10,320 synthetic images and 254,400 human evaluations show that powerful models can outperform humans when it comes to general marketing images. But that study never asked: Is this image the same as the product that comes in the package?
E-commerce visual strategy / August 2026
AI-generated images or real photos?
Which sells more in e-commerce?
Aesthetics alone cannot answer this question in 2026. The answer lies in the returns ledger, court rulings, marketplace metadata requirements, and the customer’s expression when they open the box. This article brings together peer-reviewed research, consumer surveys, laws now in force, and twenty years of production experience.
Editor's note, August 24, 2026.
Four things have changed since the original publication in June. Turkey’s regulation on AI advertising took effect on August 1, and the EU AI Act’s transparency provision on August 2, one day apart. The U.S. Supreme Court declined to review the Thaler case, in which AI was named as the sole author; the lower court’s ruling requiring sufficient human contribution therefore stood in that case. A new study in Journal of Advertising showed how AI models erode brand trust and two ways to counter it. Finally, returns data moved the debate from aesthetics to accounting. This version also tightens how sources are used: general AI research unrelated to product images is labeled separately, preprints are not presented as peer-reviewed studies, and return rates whose primary source could not be verified have been removed.
The first 3 minutes
Four things you need to know before joining a meeting
There is no single winner in this debate. The right choice depends on what purpose that image serves for you.
The NRF estimates that 19.3% of online sales were returned in 2025. There is no independent measurement yet showing how many percentage points AI-generated images add to that rate.
In the Harris Poll, 63% of respondents said they were less likely to buy from a brand using AI-generated ads. This measures attitudes toward advertising, not product-page conversion.
Turkey's regulation focuses on digital characters that are indistinguishable from humans. In the EU, however, the provider's machine-readable labeling and the seller's deepfake disclosure are two entirely different obligations.
How should you read this page? Not all of the following findings carry the same weight, so we’ve labeled them all:
Peer-reviewed scientific journal · Pre-print has not yet been peer-reviewed · Industry data survey or institutional report · Official legislation or platform policy.
And here’s something we’ll say once and never repeat: none of the findings presented here guarantee what will happen in your category. A survey measures self-reported attitudes, a field experiment measures the behavior of a single platform, and a peer-reviewed study measures a controlled task. Only your own test will reveal what will happen with your specific SKU. Keep this in mind as you read; we won’t remind you of it in every paragraph.
A product photo has to do more than look good. When the customer opens the box, they should say, "Yes, that’s exactly it." This is precisely where the key to the 2026 AI debate lies: the question is which image can truly stand in for the product.
Product photography serves a different purpose than advertising imagery. Advertising imagery sells an emotion. Product photography, on the other hand, makes a claim: this fabric drapes like this, this ring takes up this much space on the finger, this lipstick gives this shade on this skin tone. In a market where customers cannot touch the product, the image is not merely a decorative surface, It is proof of the item being bought. When that claim fails, the cost shows up in shipping, return forms, and review scores.
Academic studies and consumer research from 2024 to 2026 show a conditional but consistent pattern. An AI-generated image can be technically flawless; however, the moment a consumer senses or learns that it is synthetic, emotional trust drops, perceived authenticity weakens and purchase intent declines. The effect is not equally strong in every category. In fashion, cosmetics, jewelry, luxury, baby products, food, hospitality and personal care, that is, where consumers test the product on their own body or in their own experience, it is much stronger.
LUX Photo Video Production has produced e-commerce, fashion, cosmetics, jewelry, industrial product, and campaign photography in our Istanbul and Berlin studios since 2005. We do not reject AI; we use it for concept development, background variations, demos, campaign adaptations, and some image extensions. But for the main product image, the question has remained the same since 2005: Does this photo truly represent the product?
The 2026 table, in eleven points
- The issue isn't whether AI is used or not. According to the research, four factors determine the response together: output quality, product accuracy, transparency and the brand’s reason for using AI.
- There are areas where AI is clearly ahead. A large-scale peer-reviewed study shows that the most powerful generative models match or surpass human-created work in terms of quality, realism, and aesthetics in general marketing visuals. This is not proof of physical product accuracy-it is the result of a different competition.
- There is no universal rule that AI attracts attention while real photos sell. In the Vinted preprint, AI-generated images performed better on attention metrics in fall and winter, while original photos led particularly on conversion metrics in spring and summer. The same platform and year produced opposite results.
- Returns are the hidden cost of e-commerce. According to the NRF, an estimated 19.3% of online sales were returned in 2025. That is why judging visual strategy by clicks alone is misleading.
- No study was found to support the claim that AI-generated images increase returns by 1.8-4.2 percentage points. The source review found no supporting study. Track conversion and reasons for returns together for your own SKUs.
- Loss of trust in AI search nearly doubled in 12 months. In Fractl’s U.S. study, the share saying that heavy AI use reduces brand trust rose from 20% to 39%. This finding concerns AI search and content, not a product photography experiment.
- Why you use it is more important than how you use it. Experimental studies show that trust declines when the use of AI is explained as a "cost-saving measure"; however, when presented with a different rationale, it generates a level of trust similar to that of human-generated images.
- Supporting a claim of social value with synthetic imagery is backfiring. Journal of Advertising experiments found that AI models in body-positivity campaigns increased perceived brand hypocrisy.
- Transparency is now in effect-but the two countries want two different things. Turkey is regulating digital characters that could be mistaken for real people; the EU, on the other hand, is addressing provider labeling, deepfake disclosures, and certain AI interactions separately.
- Pure AI output carries a high risk of lacking copyright protection. In the U.S., protection does not arise without sufficient human creative contribution. However, other rights, such as trademark, contract, and unfair competition, should be evaluated separately.
- The origin of AI is sometimes traceable, but it can never be definitively diagnosed. C2PA, IPTC, and SynthID provide strong indicators; however, metadata can be deleted, and classifiers make mistakes.
In 2026, visual strategy cannot be reduced to whether to use AI. The question is where AI saves time and where a real shoot provides evidence.
Consumer data
Eight numbers, four different measurements
The data below was compiled from academic studies, institutional research, and industry surveys published during the 2025-2026 period. What each finding measures is indicated below the card; full citations are provided in the bibliography at the end of the page.
%78
"AI makes the ad feel less original"
Harris Poll / 4A's / Infillion, June 2026 · Perception of Advertising
%73
They trust an ad less if they suspect it was created using AI
Harris Poll / 4A's / Infillion, June 2026 · Perception of Advertising
%63
Consumers are less likely to make a purchase from a brand that uses AI-generated ads
Harris Poll / 4A's / Infillion, June 2026 · Perception of Advertising
%50
Consumers prefer brands that do not use GenAI in their consumer-facing content
Gartner · 1,539 U.S. consumers, October 2025
%39
Extensive use of AI reduces trust in the brand (up from 20% in 2025)
Fractl / Search Engine Land · 1,008 U.S. consumers · AI search/content context
%76
Consider it very important to distinguish AI-generated images, video, or text from human-created content
Pew Research Center · U.S. adults, 2025 · general media context
%53
Are not confident they can do so
Pew Research Center · U.S. adults, 2025 · general media context
%37
Say they will check the return policy when they learn that a virtual model was used
Stylitics + Aha Studio · 411 shoppers · company-sponsored survey
These findings come from different countries, samples, and methods. Gartner found that half of consumers prefer brands that avoid generative AI in consumer-facing content. Cint found that a large share expect brands to disclose AI use. In Klaviyo’s 2026 study, 61% were neutral about brands using AI-generated marketing content, while 32% said it reduced their trust. These measure general consumer-facing content and advertising, not product-page conversion.
What stands out most is speed. Fractl and Search Engine Land asked the same questions in two consecutive years. The share saying that heavy AI use would reduce brand trust almost doubled in one year. Those finding AI search more useful than traditional search fell from 82% to 54%, while the “AI skeptic” segment rose from 3% to 17%. The comparison shows a clear decline in trust, but concerns AI search and general content, not product photography.
Pew Research Center adds a layer of uncertainty: 76% of U.S. adults consider it very important to identify whether images, video, or text come from AI or people, but 53% lack confidence in their ability to do so. Most want to know while recognizing that they may not be able to tell. This is a media-literacy finding, not an e-commerce conversion test; it can reasonably be read as a risk signal for product pages. In the fashion-focused Stylitics-Aha Studio survey, 37% said learning that a virtual model was used would make them more cautious about sizing, 37% would check the return policy, and 30% thought they might need to return the product.
This is a distinction that often gets missed. DoubleVerify's 2026 study shows that consumers do not reject AI technology outright. What draws the line is quality . The negative response is driven by images looking cheap, uncanny, and careless. In the same study, 42% of EMEA consumers said they viewed even a brand associated with low-quality AI content negatively.
This finding challenges both sides, those who defend AI and those who reject it outright. What matters is not the tool used to produce the output, but the quality of the output and the honesty of the claim.
All of the images below are from the LUX Production portfolio.
100% real people, 100% real sets, 100% real light.
None of them used generative AI. Read the finding above through these images: the dividing line lies in quality, not in the tool. Real production can stay real without compromising on quality, which puts it on the stronger side of this debate.
Market timeline
What happened in 18 months?
The timeline of the upheaval in Turkey’s e-commerce visual market is now clear. The timeline below traces the changes we’ve observed from within the industry, alongside publicly available technology and market data.
- August 2025: The first wave Google is releasing Gemini 2.5 Flash Image. The model, codenamed "Nano Banana," is gaining traction in the community and is going viral for photo editing. The output quality is currently at the level of casual fun and social media; no one is canceling their studio appointments just yet.
- November 20, 2025, A milestone is reached Google DeepMind launches the Gemini 3 Pro-based Nano Banana Pro(officially Gemini 3 Pro Image) worldwide. It offers 2K and 4K output, compositions using up to 14 reference images, character consistency for up to five people, control over camera angle, focus, depth of field and color grading, and legible text generation. The same week it is integrated into Google Ads, Workspace and Vertex AI. Product imagery crosses the “good enough” threshold on this date. The first to notice are not photographers but the sales managers who plan monthly catalog budgets.
- December 2025 - February 2026: The replacement wave For a seller under cost pressure the math gets simple: the cost per image falls from the $50-500 range to a few cents. The shift happens fastest in white-background cutouts, catalog variants and generic lifestyle scenes. This is exactly when the phones start going quiet on the production side.
- March-May 2026: Saturation and backlash Feeds fill with synthetic images. Merriam-Webster chooses the 2025 word of the year "slop", while Macquarie Dictionary chooses "AI slop". These choices show that the issue has become a cultural phenomenon. “Is this real?” reaches the product page, while returns and negative reviews over mismatched sizes, colors, and textures begin to accumulate.
- June-August 2026: Regulation and the return to real production Turkey and the European Union are implementing their transparency requirements one day apart. Peer-reviewed studies and consumer surveys published during the same period confirm that the loss of trust has deepened. In categories that require a high level of trust, some brands are returning to actual production-but not with the old briefs: fewer photos, greater evidential value.
A business perspective on this timeline
What happened in November 2025 was not a leap in quality; it was category substitution. A specific subsegment of product imagery (plain-background cutouts, color and size variants, generic backgrounds) has been permanently displaced on price. Accepting this is the right starting point for production companies and brands alike, because that segment is not coming back.
Anyone who has shot a 400-SKU catalog on a white background knows that the job was never really creative. It was repetitive and measurable, and the fee was the fee for repetition. The machine took exactly that part, and that is no surprise.
The second quarter of 2026 also showed the limits of substitution. AI’s ability to generate an image and represent a product are not the same. The drape of fabric, leather’s pores, a diamond’s refraction, cosmetics on real skin, and a product’s actual size in the hand or on the body cannot be resolved by an image that merely looks realistic enough. A photograph’s evidential strength comes from its indexicality , its physical connection to the object. Anyone who has seen a gemstone’s fire change when a lamp moves by two degrees during a jewelry shoot understands why this cannot be imitated.
What was replaced was not the product photo; it was the cheapest layer of the product photo.
What’s coming back, however, is not the old volume, but a request for proof.
The practical conclusion: after 2026, the e-commerce visual budget should be split in two. On one side, high-volume, low-risk production suited to automation. On the other, main images that carry the proof of the product, shot in real production and, where possible, made verifiable with provenance data. Brands that try to cover both with a single budget line often lose on both sides: they are neither cheap enough nor trustworthy enough.
The Return Economy
The real debate is about accounting
The biggest financial impact of a visual decision isn’t reflected in the conversion rate. It appears in the returns ledger. This section sets up the table that an e-commerce manager needs to review.
The appeal of AI production is undeniable. In the controlled task in Hartmann, Exner, and Domdey’s peer-reviewed study, producing an image with DALL-E 3 cost $0.04, while an image from a freelancer given the same brief cost 100 dollars . This compares general marketing assets, not SKU photography involving physical samples, models, sets, color management, and delivery. Even so, the difference in scale is clear.
The question is where those savings may be offset. One of e-commerce operations’ largest costs is Product returns. The NRF and Happy Returns’ 2025 report estimates that 19.3% of U.S. online sales were returned. A 2023 Coresight Research study, sponsored by 3DLOOK and involving 100 U.S. apparel decision-makers, found an online apparel return rate of %24,4 according to the report. The same report cites an Optoro example in which processing the return of a $50 item can cost up to 66% of its price. This rate varies by category and operation.
Two frequently repeated figures also need scrutiny. Claims that 22% of returns result from products looking different from their images, or that misleading images cause 40% of returns, trace back to old industry infographics. Their current, independent methodologies are not publicly available, so we do not use them as 2026 benchmarks. Nor was an independent, auditable study found showing that AI-generated product images universally increase returns by 1.8-4.2 percentage points. The established point is narrower: any image misrepresenting stitching, color, texture, fit, or scale creates an expectation gap. Its financial effect on your brand can only be measured through SKU-level testing.
Standard of Proof
What can be said with confidence today is: Online return rates are high; consumers expect accuracy and clear disclosures in AI-generated visuals.
What cannot yet be said: "AI-generated images increase the return rate by this many points in every category." There is no such causal relationship, and those who claim there is have not proven it.
| Metric | Traditional production | AI-generated image | Operational impact |
|---|---|---|---|
| Cost per image | Depends on the brief and SKU scope | $0.04 in the task at hand | In the peer-reviewed example, the freelancer is $100; this is not a direct comparison with a physical product photo. |
| Time to market | Weeks (logistics + studio) | Minutes | A clear advantage in fast catalog updates. |
| Visual accuracy and texture | An exact representation of the physical product | Risk of texture deviations and anomalies | There is no universal rate; testing is required on a per-SKU basis. |
| Evidential function | High: physical contact with the object | Low: a sense of doubt and artificiality | Purchase intent may decrease with open labeling. |
| Copyright protection | Exists | May not arise in the U.S. without meaningful human contribution | Trademark, contract, and unfair competition rights are also considered. |
| Compliance Burden (Turkey / EU) | Generally, there are no additional statements specific to AI | Conditional disclosure requirements | In both methods, the rules regarding advertising, privacy rights, and product accuracy apply. |
Do the math using your own numbers. Consider a store with 5,000 orders a year, an average basket of TRY 1,200, and an 18% return rate. Assume a change in imagery raises the return rate by 2 percentage points, from 18% to 20%. That means 100 additional returns a year. If processing a return costs 40% of the item’s price, the additional annual cost is approximately 48,000 TL. The 2-percentage-point increase and 40% processing cost are illustrative inputs for sensitivity analysis, not research findings.
Repeat the calculation using your own net return-processing costs, gross margin, and actual quote for a shoot or AI production. Base the decision on total contribution margin after returns , rather than the cheapest cost per image.
Cart abandonment and pre-purchase hesitation
The impact doesn’t end with the refund. In a survey conducted by Storyblok and OnePoll in 2023 with 1,000 consumers, participants 85% said they were not interested in AI assistance with purchase decisions, while 17% said an AI recommendation would make them less likely to buy. This measures AI recommendation systems, not AI product photography. The same organization’s often-quoted 60% cart-abandonment figure comes from a different study: a 2022 survey of 6,000 people in the U.S. and Europe about overall poor web experience .
This data does not prove that AI images cause cart abandonment. A more honest conclusion: AI is accepted where it adds convenience; it creates purchase friction where it increases the need for verification. The Vinted preprint in the next section examines this possibility through real platform behavior. Its design is limited and its result changes with the season, which is why it is instructive.
So how do you measure this accurately? The only reliable way to measure the impact of a visual change is on a per-SKU basis and including the return rate to draw a comparison.
Recommended setup: pick 15-25 SKUs from the same category, keep half with their current images and reshoot the other half in real production. Track four metrics together for at least 30 days (45 days for seasonal products): conversion rate, return rate, distribution of return reasons and product review score. A test that looks only at conversion hides the cost of returns and leads you confidently to the wrong conclusion.
Why be skeptical of claims that AI increased conversion by 340%?
Most of the AI performance statistics circulating in the industry cannot be verified. They share the following characteristics: Most are generated in the blog posts of companies selling AI visual tools, and they cite each other as sources. When the same figure is repeated on thirty websites, it takes on the appearance of "industry data"; but at the end of the chain, there is no peer-reviewed publication, an audited financial statement, or an independent measurement organization at the end of the chain. There is only the original blog post.
When evaluating a proposal, ask yourself three questions. One: Who published this figure first-the party that conducted the measurement, or the party selling the vehicle? Two: What was the comparison based on-a professional photo, or the absence of any visuals, or an amateur phone photo? Three: Are return rates and customer review scores included in the metrics, or were only clicks and items added to the cart counted?
We need to turn that same criticism around, or else it wouldn’t be honest. The weakest point of the anti-AI argument is that it makes an unprovable generalization like "real photos always sell better." A defensible claim is a narrower and stronger one: A rich set of visuals that accurately represent the product reduces returns; visuals that misrepresent the product, regardless of the production method, increase returns.
Materials, textures, and surfaces, from the LUX Production portfolio.
Leather grain, metal sheen, fabric texture. Real product, real light, studio photography.
The "expectation gap" described in this section may or may not arise precisely from these details.
Academic literature
What does science say?
The 2024-2026 literature is not one-sided, and it’s important to state this from the outset. AI can create aesthetic appeal and drive clicks in general marketing visuals. What determines trust lies elsewhere: the product’s physical authenticity, whether the source is discernible, and why the brand is using this technology.
First, the strongest finding against our own thesis
Peer-reviewed study
In their study published in International Journal of Research in Marketing, Hartmann, Exner, and Domdey compared 10,320 synthetic marketing images from seven models with 2,400 human-created images and collected 254,400 human evaluations in total. The strongest models matched or surpassed human work in quality, realism, and aesthetics. In a field test with 173,022 impressions, the best AI banner also achieved a click-through rate up to 50% higher than a professional stock photo.
Any analysis that ignores this result is incomplete. The argument that "AI produces poor visuals" ends here. However, the study measured general marketing content and banner performance. The color, stitching, texture, scale, or match between the product and its packaging of an SKU were not tested at any stage. We are looking at the results of a different competition: proof of appeal, not proof of product accuracy.
1. Brand hypocrisy: the brand’s value proposition and its synthetic visuals don’t align
The strongest evidence here comes from a study by Quan Xie, Sidharth Muralidharan, and Joe Phua of Southern Methodist University’s Temerlin Advertising Institute, published in Journal of Advertising. The question arose from Levi’s announcement that it would use AI models to increase diversity, and the backlash that followed.
Across three experiments, using AI-generated human models in body-positivity and plus-size apparel campaigns led female consumers to perceive strong brand hypocrisy . Attitudes toward the brand, purchase intent, and likelihood of recommending it fell significantly.
The mechanism is called Social Presence Theory. As Xie puts it, consumers do not find AI models human, warm, or relatable. When you try to convey a message based on empathy using a synthetic figure instead of a real person, the message itself falls apart. The ad says, "We love you just the way you are," but there’s no one in the visual.
The less discussed part of the study: two antidotes. This study does not say, "Don’t use AI models." It tests two intervention mechanisms, and both are effective.
The second experiment showed that, goal concreteness moderates the effect: concrete, measurable messages reduce the impact of lower social presence on perceived brand hypocrisy. The third experiment found that in ads with an abstract purpose message, adding a well-developed transparency statement alongside a simple AI disclosure improves brand evaluations.
The researchers have three recommendations for brands: Explain your use of AI; provide that explanation alongside information about the brand’s concrete commitments in the real world; and describe how AI is being developed responsibly. In short, the problem is bare, unjustified use of AI.
Three experiments by Brüns and Meißner, published in 2024 in the Journal of Retailing and Consumer Services, point in the same direction: GenAI-produced social media content damaged perceived brand authenticity and behavioral responses. But framing AI as supporting people rather than replacing them clearly reduced the damage.
2. Signaling theory: an image describes not only the product but also the seller
Research based on signaling theory makes a simple but powerful point: a marketing image also tells consumers how much effort the seller has made , not only what the product looks like. British Food Journal found that AI-labeled images weaken consumers’ perceptions of seller effort .
When customers sense that the seller took a shortcut on the image, they also lower their expectations of product quality. The question on their mind: if they cut corners here, where else did they cut corners on the product? The effect is clearly stronger for products that command a high price premium. In other words, an AI image costs a premium brand more than it costs a budget brand.
3. Product involvement changes the picture
Israfilzade’s 2025 2×2 experiment tested both the advertising source, AI or human, and transparent labeling. The decisive variable was product involvement . Low-involvement products such as snacks encountered little resistance to AI imagery. For products with higher financial and technical risks, such as laptops or luxury goods, an explicit AI label sharply reduced trust and purchase intent.
This finding is the academic basis for the category matrix later in the article. The practical implication is clear: the same production budget yields low marginal benefit in a low-involvement category and high marginal benefit in a high-involvement one. Spreading the budget evenly across categories is therefore an expensive mistake.
4. The rationale is more decisive than the technical aspects
The most commercially useful finding concerns how brands explain their reasons for using AI. In experiments examining the effect on trust, the final experiment had n=212, when the rationale was "privacy protection" , consumers reported trust levels similar to those for human-created images. When the same technology was explained in terms of "cost-effectiveness" , trust and purchase intent declined significantly.
In this experiment, consumers reacted less to the brand’s use of AI than to the realization that it used AI only to cut its own costs .
5. Vinted preprint: results reverse with the season
Not peer-reviewed · Preprint
Published in August 2026, the study compared AI-generated product images with original photos from matched listings on Vinted across two seasonal waves. The results were not one-sided:
- During the fall-winter season AI-generated images yielded more positive results in metrics related to attention.
- During the spring-summer season The original photos stood out, particularly in metrics related to conversion.
The study’s own interpretation is that performance depends on the balance between aesthetic refinement, visual originality, perceived risk, and the need to validate the product. One limitation is clearly stated: Since AI-generated images are more polished, it is not possible to completely distinguish between "AI source" and "presentation quality," and the platform’s algorithm cannot be controlled. In other words, we have a context-sensitive signal , not a universal law of sales. This literature offers no universal rule.
6. Tolerance varies depending on the visual type
Research by Münster School of Business (FH Münster) and CECIRE, in partnership with fashion retailer Ernsting’s family, draws a practical distinction: consumers show some tolerance for AI in decorative and background images , but are highly critical of personalized product and model photography . This distinction underpins the later decision table by image type. It is probably also the most readily applicable finding in this article.
7. Six findings that are often overlooked but are important
- "Even the word "AI" itself can cause sales to drop. A study by Cicek, Gursoy, and Lu of Washington State University shows that the inclusion of the term "Artificial Intelligence" in product and service descriptions negatively affects purchase intent. The mechanism operates through emotional trust, and the effect is much more pronounced in high-risk categories such as expensive electronics, financial services, and medical devices.
- Perceptions change once the source is revealed. In Zhang and Hur’s 2025 study, when the source of the image was not disclosed, the difference between AI-generated and human-generated images remained minimal. However, when it was revealed that the image was AI-generated, trust and purchase intent dropped significantly.
- Risk is higher in hedonic categories. A study by Belanche and colleagues in the International Journal of Information Management found that AI imagery was viewed more negatively in hedonic and high-involvement decision-making.
- Emotional content can trigger moral disgust. Journal of Business Research reports that when consumers believe emotional marketing messages were written by AI rather than by people, they find them less authentic, feel moral disgust and show lower purchase intent; even though the text is exactly the same.
- An ethical stance can help soften the reaction. The finding reported by Forbes, based on four experiments led by Prof. Sean Sands of Swinburne University of Technology, suggests that a credible framework of social benefit and ethical responsibility can reduce negative reactions. Since the citation for the fully peer-reviewed article could not be found, this should be considered supporting evidence.
- There is no evidence for assuming consumers will simply get used to it. Conjointly’s 2023-2025 waves showed mixed attitude changes. Most declines were within the margin of error, but approval of AI use in marketing content fell significantly, from 55% to 36%. The loss of trust in Fractl’s separate study concerns general AI search and content, not a product photography test.
| Research / Institution | Focus | Method and Sample | Key finding |
|---|---|---|---|
| Hartmann, Exner & Domdey IJRM (2025) | Overall marketing image quality and ad clicks | 10,320 AI images; 254,400 evaluations; 173,022 banner impressions | The most powerful models matched or exceeded performance on human images; the study did not test the accuracy of physical SKU representation. |
| Xie, Muralidharan & Phua SMU · Journal of Advertising (2026) | AI models in body-positivity campaigns | Three series of experiments involving female consumers | Perceived brand hypocrisy; low social presence; decline in purchases and recommendations. A concrete purpose message and a qualified transparency statement soften this effect. |
| Brüns & Meißner (2024) Journal of Retailing & Consumer Services | AI content on social media | Three experiments | Direct damage to perceived brand authenticity; the "supportive AI" framework mitigates the damage. |
| Tainted by technology? British Food Journal (2026) | Perceived seller effort | Three consumer experiments | The "AI" label diminishes the perception of effort; resistance is growing among high-priced products. |
| Israfilzade (2025) Equilibrium | Transparent labeling × product involvement | 2×2 between-subjects design | An explicit AI label reduces trust and sales for high-involvement products. |
| Motivation framing study | The impact of the rationale for using AI | Three experiments (n=130, n=79, final experiment n=212) | A cost-saving rationale reduces trust; a privacy-protection rationale produces trust similar to that in human-created images. |
| FH Münster / CECIRE In partnership with Ernsting’s family | AI-generated images in fashion retail | Market and consumer analysis | Partial approval for decorative images; strong criticism of personal and model images. |
| Vinted field test Preprint (2026) | The role of evidence in secondhand fashion | Field experiment, S-O-R model | Seasonal reversal: AI attention metrics in fall and winter; original images and conversion metrics in spring and summer. |
| Cicek, Gursoy & Lu Washington State University | The term "AI" in the explanations | Experimental study | Emotional confidence is declining; purchase intent is decreasing. This is more pronounced in high-risk categories. |
The nature of product photography
Product photos are not advertising images
In e-commerce, visuals serve as a substitute for the product’s physical presence when consumers cannot touch it. The texture of fabric, the sheen of metal, the transparency of glass, the surface of leather, the true shade of a cosmetic pigment, the way a ring fits on a finger, the quality of the packaging: all of these are conveyed through a single image. As the customer looks at that image, they are actually trying to reach out and touch it.
That’s why an AI-generated image that makes the product appear more flawless, shinier, smoother, larger, or a different color,no matter how aesthetically successful it may be, poses a commercial risk. A single question arises in the customer’s mind, and that question determines both the conversion and the return: "Will the product really look like this?"
There is another layer in luxury, cosmetics and fashion. The problem raised by the study “When AI Doesn’t Sell Prada” goes beyond realism. In luxury consumption, value is bound up with human labor, craft, attention, materials and the story of how something is made. When an AI image thins out this sense of labor, consumers start to question how the brand creates value. A customer who wonders who stitched a bag will also wonder who made its image.
Data from 2026 brings this finding to the commercial level. In a survey conducted by Clutch in June 2026 with 408 consumers, the participants 36% said that seeing real people behind a brand was the strongest driver of their loyalty say. That answer ranks ahead of price, product quality and convenience. In the same study, 49% point to real customer stories as the most memorable campaign format.
Same image, different label, different level of trust
Research by NIM (Nuremberg Institute for Market Decisions) shows that labeling the same image as a “photo” or an “AI-generated image” changes consumer perception. An image labeled as AI can be perceived as less emotional, less credible and less memorable. In other words, transparency does not always build trust; sometimes it only makes doubt visible.
The third experiment in the SMU study rounds out the picture, and these two should be read together: While a simple "Generated by AI" label may erode trust, a well-crafted transparency statement that explains how the process works improves brand perception. For example, explaining that the AI model was trained on a dataset of consented, licensed photos of real people can reduce consumer skepticism. The difference lies in what the label says, not in whether there is one.
As of 2026, this balance is no longer a matter of choice. The dilemma NIM points to (labeling lowers trust) has so far been a strategic question. Starting August 1, 2026 in Turkey and August 2, 2026 in the EU, under certain circumstances Tagging is required.
As a result, the question is no longer simply "Should I tag it?" but has split into two: (1) In which categories should I create images that I won’t have to tag? (2) When I have to add a tag, how do I make it a qualified tag?
Psychology of Buying
The nine mechanisms underlying the reaction
1. Loss of evidential value
In e-commerce, a photo is physical evidence of a product. The moment a consumer suspects the image might be synthetic, that evidence loses its validity.
2. Transfer of human labor
A genuine photograph conveys the sense that someone standing in front of the product made decisions about the lighting, the angle, and the materials. That sense of effort quietly transfers to the product.
3. The uncanny valley
Shadows, symmetry, texture, fingers, faces, fabric, or minor anomalies in perspective leave the consumer with an indescribable sense of artificiality.
4. Persuasion alert
Sensing that the image has been overly optimized to be convincing, the customer goes on the defensive: "If they cut corners on the image, did they cut corners on the product, too?"
5. Lack of touch
When shopping online, no one can touch the product. Real lighting, real surfaces, and the correct scale are the only things that can partially make up for this shortcoming.
6. Ethical response
The fear that AI will replace creative labor is triggering a reflexive rejection based on values among certain consumer groups.
7. A signal of cheapness
Consumers interpret an AI-generated image as "less effort was put into it," and this perception extends to the entire brand. If the image looks cheap, the product is also considered cheap.
8. Collective threshold of doubt
The effect is environmental. Consumers also penalize a brand for appearing next to low-quality AI content, on top of its own advertising.
9. The perception gap is not on the consumer’s side, it is on yours
The most striking finding in the IAB’s 2026 report, covering 505 U.S. Gen Z and millennial consumers and 104 advertising executives, concerns the executives: 82% believed young consumers viewed AI ads positively. The actual consumer figure was 45%. The gap widened from 32 percentage points in 2024 to 37 in 2026.
LUX Production portfolio-all real photography.
Real models, real sets, real products, real lighting.
Generative AI was not used in any of these images. All nine of the mechanisms listed above
are either activated or deactivated depending on whether an image was generated in this way.
Framing thesis
Same technology, opposite results
When the two brand cases are placed side by side, a pattern emerges that is consistent with academic findings. Let us say upfront that this is not proof of causality; still, it is instructive.
Crisis: March 2023
Levi Strauss & Co. × Lalaland.ai
Levi’s announced plans to test an AI modeling system with Amsterdam-based Lalaland.ai, representing different ethnicities, ages, and body types on its e-commerce site. It positioned this as a method for increasing diversity, inclusivity, and sustainability .
The announcement sparked a wave of harsh reactions in the global media (The Independent, The Guardian, The New York Times) and on social media. The gist of the criticism was this: producing synthetic minority models instead of hiring and paying real human models strips diversity of its value and reduces it to a mere effect. New York Magazine called this "artificial diversity.".
Conclusion: After the backlash, Levi's said its AI models would not replace real models or its diversity commitments and would only support shoots with human models. It was a narrowing and clarification of a controversial positioning.
A more subdued reaction, July 2024
Mango, "Mango Teen / Sunset Dream"
Mango launched an ad campaign for its youth collection created entirely using artificial intelligence. Technically, it was a step beyond what Levi's had done: not a synthetic model, but synthetic campaign.
Mango framed the process in terms of operational speed, design flexibility, and production efficiency , rather than social value, diversity, or inclusion. AI was presented as a production tool, not a claim to virtue.
Conclusion: The campaign also faced criticism over job losses, representation, and product accuracy. However, the sources reviewed revealed no measured backlash on Levi’s scale. That does not establish commercial success. The cautious inference is that an efficiency rationale produced a different kind of debate from a claim about social value.
The difference between the cases lay in the claim , rather than the technology. Levi’s presented AI as a commitment to social value , which the public read as hypocrisy. Mango presented it as a production method ; no backlash of comparable scale was found.
These two cases provide a field example consistent with the SMU study’s findings on brand hypocrisy and its motivation-framing experiments. A case comparison alone does not constitute causal evidence.The one-sentence rule
AI should never be presented as a substitute for a value commitment. When claims of diversity, body positivity, sustainability, local production, handcrafted products, or social responsibility are supported by synthetic images, in the consumer’s mind brand hypocrisy is triggered. The technology should be clearly positioned as a tool that supports operational speed and the creative process. Other uses often cost more than they bring.
Case analysis
The form of transparency: Guess and H&M
The second comparison is even clearer: two brands, similar technology, different transparency architecture, different responses.
Guess, Vogue, August 2025
In the August 2025 issue of *Vogue*, *Guess* featured a two-page ad starring two "models" created entirely by AI. The characters, created by Seraphinne Vallora, who produced the images, were named Vivienne and Anastasia. There was an explanation-but it appeared in the corner of the page as a small-print production note.
A TikTok video that highlighted this phrase garnered over two million views. The backlash was directed at both Guess and Vogue; subscription cancellations were discussed. Condé Nast was forced to clarify that the AI model had never been used on Vogue’s own editorial pages. Professional organizations such as the Model Alliance drew attention to the lack of consent and compensation standards.
H&M, digital twins, 2025
H&M created digital twins of 30 real models with their consent and royalty payments . Participation was closed: models were selected directly, the twins were used only in H&M campaigns, and the models retained their rights. Published content carried prominent watermarks and transparent labels.
H&M also faced criticism about job losses, consent standards, and the industry’s future. But the debate centered less on deception than on labor force policy . Some participating models described the process as professional, collaborative, and transparent.
The difference was not the technology, but the disclosure’s size and placement, and what happened to the people behind it. Transparency is not a legal checkbox but a trust mechanism. Where the label sits and how large it is can sometimes shape the reaction more than the AI itself.
A practical takeaway for e-commerce brands: every piece of information you try to hide will cost you exponentially more once it’s discovered.| Brand | Year | Method and positioning | Reaction | Reputation Effect |
|---|---|---|---|---|
| Levi's | 2023 | Diversity-focused AI models with Lalaland.ai | Severe; Accusations of "artificial diversity" | After the backlash, the company clarified that AI was intended to support human models. |
| Mango | 2024 | A fully AI-generated campaign for Mango Teen; efficiency framework | No reaction on the scale of Levi's was observed in the sources reviewed. | There is no publicly available, causal data regarding the commercial impact. |
| Zalando | Q4 2024 | AI in approximately 70% of editorial images, framed as an enrichment tool | There is no brand-level sentiment analysis in publicly available sources. | Company statement: turnaround fell from weeks to days, and costs fell by up to 90%. |
| Guess | 2025 | AI model in a Vogue ad; small-print caption | Viral: Allegations of a lack of transparency | Loss of reputation due to the perception of label concealment. |
| H&M | 2025 | Consent-based digital twins + royalty payments + visible watermark | Mixed; criticism of job losses and an emphasis on model rights go hand in hand. | The first set of images was released in July 2025; the commercial impact was not disclosed. |
| Coca-Cola | 2024 & 2025 | AI-generated New Year's ads | A broadly negative characterization: "soulless" | The brand continued; a debate over the measurement method arose. |
2026 market reality
"No AI" has become a new sign of trust
In 2026, some brands began marketing their decision not to use AI not as an ethical stance, but directly as a signal of brand trust and product authenticity.
In categories such as beauty, baby care, food, luxury goods, ready-to-wear, and analog photography, the message "We don’t use AI" is becoming increasingly visible. The reason is clear: in these categories, purchasing decisions aren’t based solely on price and function. Physical reality, tactility, human representation, the physical feel of the product, and the brand’s perspective on labor are all part of the decision.
Dove’s commitment not to create or distort images of women using AI, Aerie’s policy of not using AI-generated bodies or people, and Coterie’s decision not to use AI-generated social media images , and Polaroid’s campaigns highlighting the physical experience all point to the same commercial reality: In the age of AI, physical reality is becoming a brand asset again.
In 2026, this stance moved from campaigns to brand commitments. Aerie expanded its “100% Aerie Real” pledge through a Pamela Anderson campaign. The ad addresses the issue directly: Anderson asks an AI tool to generate models, tries to make them “more natural, less lifeless,” then gives up, steps onto a real set, and tells the camera "You can't prompt this" . In January 2026, Almond Breeze released a Jonas Brothers film explicitly titled “No AI Needed.” Brands such as Le Creuset began explaining their production processes in pinned Instagram comments.
In 2026, “human-made” became a new organic certification: what was once assumed is now a value stated explicitly.
“No AI” is not the right strategy for every brand. In categories where the product touches the body, skin, babies, food, fabric, jewelry, luxury or perceptions of health and personal care, a real production process is a direct source of trust. In other categories it is often just noise.
The second warning is stronger: “No AI” should be an auditable commitment , not a campaign slogan. New York State’s disclosure law for ads featuring synthetic performers took effect in June 2026; it is not a general “No AI” certification. An unverifiable claim that a brand does not use AI can still pose a separate risk under general deceptive-advertising rules.
Not using AI may not be enough: Appearing to use AI is also risky
The Quip Case
Quip’s ad, filmed entirely without AI (a physical set, human models, miniatures and practical effects), drew an “Is this AI?” reaction on social media. The brand had to highlight its production process and post a statement: “No AI, just us.”
Meaning for e-commerce
Images that are overly sterile, overly smooth, physically too flawless, and devoid of texture can raise suspicions of AI, even if they’re real photos. Flawlessness is now a clue.
That’s why, in 2026, production quality shouldn’t mean 'perfection." The true surface texture, the product’s scale, its physical response to light, the character of the materials, and a controlled yet believable balance of retouching must be preserved. An overly plastic aesthetic that makes a real photograph look like it was created by AI is squandering the brand’s trust capital.
This pressure is also changing the aesthetics of commercial photography. In 2025-2026 there is a clear trend in fashion and lifestyle photography toward less polished, more dynamic, less flawless images: direct flash, slight grain, unplanned details in the frame, the clutter of real spaces. The reason is not nostalgia: imperfection is now an authentication signal.
Still, it would be a mistake to apply this trend to product photography as is. In e-commerce, the product’s color, texture, and scale must be accurately conveyed. The right balance is this: The product information should be flawless; the presentation should feel natural. The product itself should be depicted technically accurately; but the setting, lighting, and composition should make the viewer feel as if they are actually there in a physical space. This is precisely where AI struggles the most-because the models converge toward an average aesthetic, that is, toward perfection.
Control surface
A product image should answer three questions
1. What is for sale?
Do the details regarding color, texture, stitching, surface, packaging, and variants match the delivered SKU?
2. What are the dimensions?
Is the product shown in a way that allows us to gauge its actual size, whether held in the hand, on the body, or next to a reference object?
3. Where is the evidence?
Are the original source plate, the shoot file, the editing history, or a verifiable provenance record stored?
Platforms and law
AI imagery is now a compliance issue
As of August 2026, regulators, platforms, and technical standards don’t just influence your visual production decisions-they dictate them.
Two regulations, two different scopes. The regulations in Turkey focus on digital characters that are indistinguishable from humans in advertisements and on AI replicas of real people. Article 50 of the EU AI Act, meanwhile, separately regulates the system provider’s machine-readable labeling, the chatbot’s disclosure, and the deployer’s deepfake disclosure. Circulating "If you used AI, tag every product image" is not found in either text.
Amendment to the Regulation on Commercial Advertising and Unfair Commercial Practices
Official Gazette: July 1, 2026, Issue No. 33297 · Effective Date: August 1, 2026The amendment drafted by the Ministry of Trade regulates a wide range of areas, from targeted advertising to influencer posts, and from discount campaigns to environmental claims. There are two provisions that directly concern visual production:
- Disclosure requirement: Ads featuring AI-created digital characters that are difficult to distinguish from real people must disclose this to consumers in a clear, understandable, and distinguishable way.
- Direct ban: Advertisements that give the impression that digital replicas of real people, created using artificial intelligence, are experiencing or recommending a product or service are prohibited. Labeling is not a solution here-the act itself is prohibited.
The test is the impression created in the average consumer , rather than the tool used. AI-assisted backgrounds or color corrections may fall outside this scope, but disclosure requirements apply to characters that appear human and claims about their experiences.
AI Act Article 50: transparency obligations
Effective Date: August 2, 2026 · Labeling (legacy systems): December 2, 2026Reports that “the EU postponed the AI Act” are misleading. The Digital Omnibus regulation took effect on July 27, 2026, and postponed obligations for high-risk systems. However, Article 50: Transparency obligations have not been deferred.
- The output of systems that generate synthetic audio, images, video, or text must be marked in a machine-readable format and it must be detectable.
- When users interact with an AI system, they must be informed: the chatbot on your site should introduce itself.
- The deployer must clearly label deepfakes upon their first display; visible disclosures are also required for certain AI-generated texts related to the public interest. Standard editing functions and certain narrow use cases may be exempt.
- Generative systems placed on the market before August 2, 2026, have a transition period until December 2, 2026, for the marking requirement.
Depending on the violation, the maximum administrative fine can reach EUR 15 million or 3% of worldwide annual revenue . This is a ceiling, not a penalty automatically imposed in every case.
For a brand that sells to the EU, the right question is: What role do I play? An AI system provider based outside the EU may fall within the scope of the regulation if the system’s output is used in the EU. A retailer’s liability, however, does not stem from the shipping address; rather, it arises from whether the retailer is a provider or a deployer and how it uses chatbots, deepfakes, or synthetic content.
Shipping products to Germany does not by itself trigger every Article 50 obligation. Directly offering an AI chatbot to EU customers, using a deepfake that appears to depict an actual person, product, or place, or putting an AI system on the market under your own brand each requires separate analysis. E-commerce exporters should document the system, output, target market, and contractual allocation of roles together.
Copyright
Copyright may not arise automatically from pure AI output
U.S.: On March 2, 2026, the Supreme Court declined to review Thaler v. Perlmutter (No. 25-449). The D.C. Court of Appeals’ ruling therefore stood: the Copyright Act requires a protectable work to originate with a human . Works produced entirely by autonomous systems without meaningful human creative contribution are not eligible for copyright registration. The U.S. Copyright Office’s official 2025 report explains that prompts alone do not establish sufficient human authorship, while human creative selection, arrangement, or modification may receive protection on a case-by-case basis.
Germany: In its decision No. 142 C 9786/25 dated February 13, 2026, the Munich Local Court ruled that three logo designs generated by AI did not qualify for copyright protection under German law in this specific case. The court examined the various prompt scenarios one by one and concluded that even a detailed instruction of 1,700 characters was insufficient to determine the expressive elements of the output. In other words, writing a longer prompt does not make you the author.
If the hero image on your product page was generated solely using a prompt, your ability to take copyright-based
action against an infringer may be weakened. This does not mean you "have no rights at all": provisions regarding trademarks, design, contracts,
personal rights, and unfair competition may apply depending on the specific circumstances.
For e-commerce, visual assets are a brand’s real assets, even when they do not appear on the balance sheet. Catalog images are licensed, resold, transferred to dealers, and protected on marketplaces. Synthetic images that do not demonstrate sufficient human creative input may be excluded from this protection. Photographs, on the other hand-despite being produced by a machine-have been considered human creations for over a century, because there is someone who selects the light, the framing, the moment, and the composition.
The boundary is unclear, but the direction is clear. Neither decision specified exactly where the line lies between "AI-assisted" and "AI-generated." The question of how much meaningful human input is required is still being resolved on a case-by-case basis.
In this uncertainty, the safest position is clear: Make sure the source image is a real photo. In AI-assisted editing based on real footage, human creative input can be documented; in visuals generated from scratch using a prompt, this is debatable. In practice, this means that archiving shoot briefs, set photos, and editing histories is now driven by commercial necessity rather than aesthetic considerations.
Marketplaces
What is required, and where?
Policies are changing rapidly. The table below is based on publicly available policy documents as of August 24, 2026; please verify through the seller dashboard before making any major changes to your listing.
| Channel | Is AI-generated imagery banned? | What is expected of the seller |
|---|---|---|
| Amazon | No | Photorealistic media featuring people generated entirely by AI must include the XMP tag contains-synthetic-performer . AI edits of real people, non-photorealistic characters, and images without people fall outside this tag’s scope; general product-accuracy rules still apply. |
| Etsy | No | Products created using AI based on the seller’s prompt can be sold under the "Designed by" category; the use of AI must be specified in the listing description. Prompt packages cannot be sold separately. This policy should not be interpreted as a rule that is automatically applied to every ordinary product photo. |
| Google Merchant Center | No | IPTC in all product images generated by generative AI DigitalSourceType: TrainedAlgorithmicMedia A metadata tag must be present; the existing source tag must not be removed. |
| TikTok Ads | No | The misleading content policy requires that AI-generated content be disclosed; there is an option in Ads Manager that says "This ad contains AI-generated content." Content that is not disclosed may be rejected. Do not confuse the Shop listing policy with the ad policy. |
| Meta (Facebook / Instagram) | No | The platform applies "AI info" labels based on industry signals and user declarations. The general content labeling approach is not the same policy as the product image requirements in Commerce Manager; please review the current ad feed in your account before launching a campaign. |
| Walmart | No (No specific rules for AI) | The public product image requirements focus on accuracy and technical quality. This review found no separate general AI disclosure rule. That does not mean misleading synthetic images are allowed. |
| eBay / Shopify | No | No separate general AI image description rule applicable to all sellers was identified in this review. Liability for misleading representations, intellectual property, and product accuracy remains subject to the general rules; please verify the regional text in the seller dashboard. |
| Turkey / EU Legislation | No, but with conditions | Disclosure is required for digital characters that cannot be distinguished from real people. It is prohibited in Turkey for a digital replica of a real person to recommend products. In the EU, the obligation to provide machine-readable labeling rests with the provider. |
Three conclusions drawn from the table
First: no major channel bans AI imagery. The issue penalized is misrepresentation. A listing is being suppressed because it shows something different from the product that was shipped-not because of who produced the pixels.
Second: disclosure requirements are limited but growing and fragmented. A brand using one image across five channels faces five sets of rules. This is as much an operational challenge as a compliance issue.
Third: In most cases, real-world footage does not trigger the origin and tagging requirements specific to AI; however, the compliance cost is not zero. Product accuracy, advertising claims, model consent, right of publicity, copyright, and platform technical requirements apply just as they do to real-world footage.
“No one will notice” is no longer a strategy-but detection is not infallible
The unspoken assumption brands relied on when using AI-generated imagery was this: if it’s good enough, no one will notice. As of 2026, this assumption has collapsed on two fronts.
On the production side, origin signals have become more widespread. Google uses an invisible watermark called SynthID in its supported generative image products; some outputs also carry visible markers and metadata . This does not mean that all models use the same system or that every recompressed file can be unambiguously identified.
An ecosystem has been established on the detection side. Google Merchant Center requires an IPTC source tag, while Amazon requires specific XMP metadata for photorealistic synthetic people. C2PA, IPTC, XMP, and model watermarks provide technical indicators; however, metadata can be removed, and AI classifiers can produce false positives and false negatives.
Conclusion: A process based on concealing the use of AI is vulnerable both legally and in terms of reputation. Technical finding: Since technical detection alone does not constitute conclusive evidence, the correct approach is to preserve the source record and provide the necessary explanation to the channel in an appropriate manner. In the “Guess” case, it was not the technology itself that amplified the backlash, but rather the visibility of the explanation and the context of the synthetic model.
Origin verification
C2PA: Making real photography verifiable
C2PA (Coalition for Content Provenance and Authenticity) is an open technical standard that can cryptographically sign the declared provenance and editing history of digital content. In practice it is also known as "Content Credentials." A common misunderstanding needs correcting here: C2PA does not indicate that the content is accurate, but rather shows who created the signed record and how it was created. Synthetic content can also carry a C2PA manifest. Adobe, Microsoft, Google, Sony, Nikon, Leica, the BBC, and the Associated Press support the standard.
Its strongest form is signing inside the camera at capture . The camera becomes a root of trust: the file receives a cryptographic birth certificate before it leaves the device. Sony’s Camera Authenticity Solution uses this approach and records not only the signature but also 3D depth information that only a camera manufacturer can provide . This allows verification that the image was captured from a real, three-dimensional object. The verification site also confirms server time, which the photographer cannot alter, and indicates whether generative AI editing was used.
As of May 2026, the following Sony camera bodies support the photo license: α1 II, α1, α9 III, α7R VI, α7R V, α7S III, α7 V, α7 IV, FX3, and FX30. The license is fee-based and operates on an annual subscription basis.
Two things to know before purchasing this technology:
A signature is not conclusive evidence. Nikon suspended its own identity verification service and invalidated the certificates it had issued during that period; as of mid-2026, the service had not resumed. The problem wasn’t with the cryptography,the signatures were valid. The issue stemmed from the signer being fed manipulated content. The ecosystem is real, but it’s not yet mature.
The workflow is fragile. To preserve the signature after editing, every tool in the chain must be C2PA-compliant, and the relevant function must be enabled. A single non-compliant tool breaks the manifest. Additionally, in some commonly used editing applications, origin information can only be applied in certain export formats.
Commercial implication: Real production is now a provenance asset
Camera-source files, on-set footage, behind-the-scenes footage, crew information, color management, the retouching process, and physical proof of the product’s shoot: together, all of these form a verifiable visual presence for the brand. The strategic value lies here: AI models will continue to improve and will largely bridge the aesthetic gap. The only thing they cannot replicate is what is captured through a reliable chain of devices and processes, the history of a physical shoot. An AI output can also be cryptographically signed; the difference is that the manifest indicates its synthetic origin. In other words, origin verification is not a certificate of authenticity, but a verifiable record of the process.
FTC's reasoning: AI is no exception when it comes to deceptive conduct
The U.S. Federal Trade Commission’s approach keeps the basic principle: using AI does not create an exemption for misleading or unsubstantiated advertising claims. “AI-assisted”, “AI-free”, “real shoot”, “exact image of the product”: every one of these claims must be provable. The principle matches the Turkish regulation, which also requires academic titles, awards and environmental claims to be documented.
August 2026 compliance checklist
In order of priority. This list is not a substitute for legal advice.
- Inventory digital copies of individuals right away. In Turkey, ads that give the impression that an AI copy of a person is trying or recommending a product are banned outright. Labeling does not make them allowed.
- Label AI characters that could be mistaken for real people. A small-print footnote isn't a solution; the Guess case demonstrated the consequences of this. Make the disclosure visible and meaningful.
- Present the disclosure together with the rationale. The academic finding is clear: a bare “Generated by AI” label can lower trust; a statement that explains how the process works and describes the brand’s concrete commitment improves brand evaluations.
- Never present AI as a substitute for a value commitment. Supporting claims such as diversity, sustainability, and craftsmanship with synthetic imagery triggers a perception of brand hypocrisy.
- Check the first message from the chatbot on your site. In an AI system with direct interaction within the scope of the EU, the user must be aware at the start of the first interaction that they are speaking with an AI; the exception applies only in the limited circumstance where this is self-evident to the average user.
- Evaluate the copyright status of the Hero images. If sufficient human creative input cannot be documented in your primary product images, which were generated entirely via a prompt, copyright protection may not apply. Review the source file and records of creative choices and edits.
- Generate a channel-based declaration matrix. The same image is subject to different requirements on Amazon, Etsy, Trendyol, Meta, and your own website.
- Keep a production log. Shoot briefs, set photos, crew lists, capture dates, and editing histories: this archive supports both your copyright claim and your claim that the images come from real shoots.
- Mark December 2026 on your calendar. The transition period for the EU's labeling requirement for legacy systems ends on December 2, 2026.
Implementation model
Positioning AI based on its task
The reason we’re sharing this data isn’t because we’re against artificial intelligence. As LUX Photo & Video Production, we also carried out AI-supported visual and video processes for our clients during the 2025-2026 period: background variations, concept tests, campaign adaptations, quick demo visuals, explorations of creative directions, and some expansion tasks.
AI is powerful when it acts as an assistant. It becomes risky when it tries to serve as proof of the product.
When it comes to the main product image, the product on a model, the result of a cosmetic application, the display of fabric/texture/color, the scale of jewelry, the authenticity of food, claims related to baby and personal care products, or the value of luxury goods, real production is still the safest ground. This is a risk management decision.
Four principles that define our approach in 2026
- The source plate principle. Every project includes at least one real photograph representing the product, and derivative images are based on that source plate. We do not generate product images from scratch using prompts. The reasons are copyright and accurate representation, rather than aesthetics.
- Disclosure architecture. The delivered image set documents how each layer was produced. We provide the brand with ready-to-use guidance on how to disclose each image’s production method on each channel.
- Framing consulting. In projects that use AI, we also work together to determine how to explain this to the client. As the Levi's and Mango cases demonstrate, the same technology can create entirely different reputational risks when presented for different reasons.
- Origin verification option. For projects where evidence value is critical, the option of delivering with an in-camera signature compliant with the C2PA standard is being evaluated. The limits of the ecosystem are clearly explained to the customer; no promises such as "unbreakable" are made.
Real production shoot
Set design, lighting engineering, model management, color calibration, proper lens selection, material analysis, and controlled post-production. Real sets, real people, real products, real light.
AI-assisted visual production
Mood boards, conceptual direction, background variations, quick campaign drafts, social media adaptations, and supplementary visual content. The one non-negotiable rule: the product’s physical reality must not be altered.
Mixed strategy
A separate risk analysis for each product category. The hero image, product page, campaign visuals, social media variations, and ad adaptations are evaluated individually.
Category matrix
What should I use in which category?
High risk
Luxury fashion, cosmetics, jewelry, baby products, food, lodging, personal care. The main product image must be a real-life photo; AI should play only a supporting role.
Moderate risk
Furniture, home decor, general fashion, lifestyle products, and accessories. The hero image should be a real photograph; AI can be used for backgrounds, variations, and campaign adaptations.
Low risk
Office supplies, industrial parts, cables, and simple technological accessories. AI can be used more widely; product dimensions, color, and technical accuracy must still be maintained.
Decision table by visual type
Category alone isn't enough. Even within the same brand, the decision depends on the purpose of the image.
| Image type | Recommended production | Rationale |
|---|---|---|
| Main product image (hero) | Real photography | Evidence surface. It directly affects the return rate, copyright protection, and platform compatibility. |
| Detail / macro (texture, stitching, stone, surface) | Real photography | The place where the reality of materials is revealed; the realm where even minor production errors have commercial implications. |
| On the model / on the body | Real photography | Scale and fit information. In apparel, size mismatch is one of the top reasons for returns. |
| Scale reference (in hand, on a table, in a space) | Real photography | The image that makes it easiest for the consumer to judge the product’s size. Any error turns directly into a return. |
| Cosmetics / Results on the skin | Real photography | Health and personal care claims; verifiability is required. |
| Background variation / seasonal scene | AI derivative from the source plate | The product doesn't change; the environment does. Efficiency is justified here. |
| Format adaptation (square, portrait, banner) | AI derivative from the source plate | Reframing and canvas extension. Low risk as long as product information is preserved. |
| Mood board / Concept research | AI | In-house production not intended for consumers. It offers a significant speed advantage; however, the risks associated with licensing and confidential data must also be managed. |
| Campaign draft / presentation slide | AI | Quick visualization prior to approval. This should not replace the final production. |
| A digital copy of a real person | Do not use | In Turkey, its use is prohibited if it gives the impression of a product recommendation. Labeling is not a solution. |
A free bonus: include concrete data alongside the image. Research shows that concrete, verifiable product data placed next to an image can soften the negative effect of perceived artificiality.
When the product page displays the fabric composition, full size range, measurements in centimeters, weight, material certifications, and color code, the consumer’s focus shifts from how the image was created to the product’s actual qualities. An intervention that can reduce both the barrier to purchase and the risk of returns, with a cost of nearly zero.
Studio / AI comparison
The photo we took, the photo we edited
The comparisons below show how AI-assisted editing can support real photography while preserving its evidential value. Each image on the left is a source plate-a real photograph.
Frequently asked questions
AI images in e-commerce: 42 questions
A reference section covering everything from fundamental questions to regulations, marketplace rules, and technical verification.
A · Basic questions
Do AI images reduce sales in e-commerce?
Yes, with conditions. When the AI origin of an image is disclosed or sensed by the consumer, emotional trust can fall and purchase intent drops. The effect is strongest in high-trust categories such as fashion, cosmetics, jewelry, luxury, baby products, food, hospitality and personal care. In a large US study published in June 2026, 63% of consumers said they were less likely to buy from a brand that uses AI-generated ads.
If AI imagery is cheap and fast, why is it risky?
Because in e-commerce, an image is digital proof of the product. If an image misrepresents the product’s color, texture, size, surface, or feel, it may lead to returns, negative reviews, and a loss of brand trust, even if conversion rates increase in the short term. The cost advantage of AI is real; however, product accuracy and the impact on returns must be measured separately.
Which sells better: real photos or AI-generated images?
There is no single winner. A large peer-reviewed study found that AI could outperform human-created general marketing images in quality and click-through rates. In the Vinted preprint, results reversed by season: AI led on attention metrics in fall and winter, while original photos led particularly on conversion metrics in spring and summer. Where physical SKU accuracy matters, a real source image has greater evidential value.
Can a consumer distinguish an AI-generated image from a real photo?
Not always. In the Pew Research Center’s 2025 U.S. survey, 76 percent of adults considered it very important to be able to distinguish whether images, videos, and text were generated by AI or by humans; 53 percent did not feel confident they could do so. This is a general media finding, not an e-commerce conversion test. In a 411-person Stylitics-Aha Studio survey focused on fashion, however, 71 percent reported that they could not distinguish between real and AI-generated clothing images at first glance.
Are consumers completely opposed to artificial intelligence?
No; the findings are mixed depending on the context. While the Hartmann-Exner-Domdey study shows that the best AI-generated marketing visuals can outperform human-created ones in terms of quality and click-through rates, the Harris and Gartner surveys point to skepticism regarding trust. Quality, product accuracy, category, description style, and the brand’s rationale for using AI are all determining factors.
Won't consumers eventually get used to AI-generated images?
Automatic adaptation has not been proven. In Conjointly’s 2025 wave survey of 301 U.S. adults, attitude changes from 2023 to 2025 were mixed; while most declines fell within the margin of error, approval of AI’s use in marketing content dropped significantly from 55% to 36%. Fractl’s finding of a loss of trust rising from 20% to 39% comes from a separate general AI search and content study.
B · Sales, conversions and returns
Do AI-generated images increase return rates?
Any image that misrepresents the product creates a risk of return. However, a review of the literature did not find any independent study showing that AI-generated product images universally increase the return rate by 1.8-4.2 points. According to the NRF, an estimated 19.3 percent of online sales were returned in 2025. The causal effect of AI on your category should be measured by tracking conversion rates, return rates, and reasons for returns collectively for the same SKUs.
What is the true cost of a return?
There is no single universal rate; shipping, repackaging, depreciation, and customer service costs vary by category. In an example calculation from Optoro shared by Coresight, the cost of processing a return for a $50 product can reach up to 66 percent of the price. Make your decision based on your own net reverse logistics costs and contribution margin.
In which categories is the return rate the highest?
In a 2023 study conducted by Coresight and sponsored by 3DLOOK, which surveyed 100 U.S. ready-to-wear decision-makers, the online ready-to-wear return rate was found to be 24.4 percent. Since it could not be confirmed that the rates of 31.4 percent for footwear and 27.8 percent for women’s apparel came from the same study, they should not be used as current benchmarks. When making category decisions, base your assessment on your own reasons for returns, such as size, fit, color, and appearance.
Do AI-generated images affect cart abandonment?
There is no direct causal relationship yet. The 85% and 17% findings from Storyblok’s 2023 survey measure AI-driven purchase recommendations; the 60% cart abandonment rate, meanwhile, measures overall poor web experience in a separate 2022 study. These figures do not prove that AI-generated product images alone cause shopping cart abandonment, but they do show that trust and usability must be managed together.
What does a consumer do when they realize an image was created by AI?
In a company-sponsored fashion survey of 411 people conducted by Stylitics-Aha Studio, 37 percent of those who learned that a virtual model was used said they would be more careful about sizing, 37 percent said they would check the return policy, and 30 percent said they thought they might need to return the product. The survey measures purchase intent, not the actual return rate.
How can I accurately measure the impact of a visual change?
A test that looks only at the conversion rate hides the cost of returns and leads you to the wrong conclusion. Recommended setup: pick 15-25 SKUs from the same category, keep half with their current images and reshoot the other half in real production. Track conversion rate, return rate, distribution of return reasons and product review score together for at least 30 days (45 days for seasonal products).
C · Category and image type
In which product categories can AI images be used more safely?
Tolerance is higher for utilitarian products with low emotional involvement, such as industrial parts, office supplies, cables, and simple technology accessories. Research supports this: AI images meet less resistance for low-involvement products. Color, dimensions, technical form, and material accuracy must still be preserved.
Why is AI riskier in fashion, cosmetics and jewelry?
These products sell on more than technical specifications. Effects on skin, fabric drape, jewelry scale, perceptions of luxury, material feel, and human representation all influence the purchase. Apparel data also indicates sensitivity to fit- and appearance-related returns; the same rates cannot be assumed for cosmetics or jewelry. Research reports more negative perceptions of AI imagery in hedonic and high-involvement categories.
For which types of images is AI considered lower-risk?
Background variations derived from the source image, seasonal scene changes, format and aspect ratio adaptations, mood boards, concept research, and pre-approval campaign drafts. What they all have in common is that the product’s physical specifications remain unchanged. However, licensing, personality rights, and platform rule compliance must still be verified.
For which type of image should AI not be used?
Main product image, close-up of fabric and stitching, product on a model, scale reference, and the result of cosmetic application. These are the images through which consumers assess the product’s true quality; any inaccuracy leads directly to a return. Additionally, having a digital replica of a real person endorse a product is explicitly prohibited in Turkey.
Are consumers tolerant of background AI?
Yes. A study conducted by the Münster School of Business and CECIRE in partnership with a fashion retailer confirms that consumers show a certain degree of tolerance for AI in decorative and background images, but are extremely critical of the use of AI in personalized product and model photos.
D · Turkish legislation
Is it legal to use AI-generated advertising images in Turkey?
Its use is not prohibited; however, the amendment to the Regulation on Commercial Advertising and Unfair Commercial Practices, which takes effect on August 1, 2026, sets two limits. First, advertisements featuring digital characters created using artificial intelligence that are difficult to distinguish from real people must clearly, understandably, and distinctly state this fact. Second, advertisements that give the impression that a real person’s AI-generated digital replica is experiencing or recommending a product are directly prohibited.
On what date was the regulation in Turkey published, and when did it take effect?
The amendment regulation prepared by the Ministry of Trade was published in the Official Gazette No. 33297 dated July 1, 2026; the provisions took effect as of August 1, 2026.
Do I need to tag every image I generate using AI?
No. The test is the impression created in the average consumer, rather than the tool used. AI-assisted backgrounds, color corrections, or abstract images may fall outside the same scope. But characters, faces, and claims about experiences that appear to involve real people trigger disclosure requirements. In borderline cases, assess the impression the ad creates in the average consumer.
What other areas does the regulation in Turkey cover?
The regulation is far-reaching: profiling and targeted advertising based on children’s personal data have been banned; the terms "advertisement" or "promotion" are now required in influencer posts; a requirement has been introduced to disclose the criteria for displaying targeted ads; the unsubstantiated use of general terms like "environmentally friendly" in environmental claims has been prohibited; the misleading use of academic titles has been banned; and new rules regarding consumer reviews have been introduced.
What happens in Turkey if AI image guidelines are violated?
The Ministry of Trade continues to monitor the compliance of advertising and promotional content with applicable regulations; in the event of non-compliance, administrative sanctions may be imposed through the Advertising Board. Expert advice should be sought for a legal assessment specific to the campaign; this article does not constitute legal advice.
E · European Union and International Legislation
I sell to the EU. Does the AI Act apply to me?
It depends on the situation. AI system providers based outside the EU may fall within the scope if the system’s output is used in the EU. The retailer’s liability, however, depends on the provider/deployer role and the manner of use; simply shipping goods to the EU does not automatically trigger all obligations under Article 50. Providing an AI chatbot directly to an EU customer or using deepfakes requires a separate analysis.
Has the EU AI Act been postponed?
Partially. The Digital Omnibus Regulation entered into force on July 27, 2026, and deferred the obligations regarding high-risk systems to later dates. However, the transparency obligations under Article 50 were not deferred; they are enforceable and subject to penalties as of August 2, 2026. The general perception that 'the AI Act has been deferred" is therefore misleading.
What exactly does Article 50 of the AI Act require?
Obligations depend on the role: providers must design directly interactive AI systems to inform users, and must mark synthetic output in a machine-readable format. Deployers must visibly label deepfakes and disclose certain public-interest AI texts and some biometric or emotion systems. Standard editing functions and narrow exceptions apply, so the same visible-label requirement does not apply to every AI product image.
What is the penalty for violating the AI Act in the EU?
Depending on the type of violation, the maximum administrative fine can be up to 15 million euros or 3 percent of global annual turnover. This is a maximum amount and is not automatically applied in every case. Authority may lie with national market surveillance authorities, the AI Office, or other competent bodies, depending on the relevant system.
F · Copyright and ownership
Do I own the copyright to the product image I created using AI?
In the U.S., only the portion of a work produced solely by AI, without sufficient human creative input, is not protected by copyright. The Supreme Court’s dismissal of the Thaler v. Perlmutter petition upheld the lower court’s ruling; the dismissal alone does not constitute a new rule of law. Elements selected, edited, or creatively modified by a human may be protected depending on the specific circumstances of the case.
Is the copyright status of AI-generated images different in Europe?
Along the same lines. In its decision dated February 13, 2026, the Munich Local Court ruled that three logo designs generated by AI did not qualify for copyright protection under German law. The court examined various prompt scenarios and concluded that even a detailed instruction consisting of 1,700 characters was insufficient to determine the expressive elements of the output.
Can my competitor use the product image I created with AI?
If the image is entirely synthetic and no meaningful human contribution can be documented, it may be difficult to raise an objection based on copyright. However, the conclusion that "a competitor may use it freely" is not automatic; provisions regarding trademarks, designs, contracts, personality rights, and unfair competition may also apply. Documenting source files and human creative choices strengthens your claim to rights.
Where is the line between “AI-assisted” and “AI-generated”?
Neither decision has clearly defined this line; the extent to which meaningful human contribution is required is being determined on a case-by-case basis. Amid this uncertainty, the safest approach is for the source image to be a genuine photograph. Human creative contribution can be documented in AI-assisted edits made to a genuine photograph. In practice, archiving shoot briefs, set photos, and editing histories therefore holds commercial value.
G · Marketplace and platform rules
Are online marketplaces banning AI-generated images?
There is no single, universal rule. No general, absolute ban was found on most of the major channels examined in this audit; instead, there are specific conditions regarding category, region, synthetic humans, ad format, and metadata. Misrepresentation of products, however, can violate general rules regardless of whether AI is used.
What do I need to declare on Amazon regarding AI-generated images?
Amazon’s public guidance requires photorealistic images featuring people generated entirely by AI to include, in the XMP
dc:subject field, the keyword contains-synthetic-performer . AI edits of real people, non-photorealistic characters, and images without people fall outside this specific tag’s scope; general product-accuracy rules still apply.What are the AI image guidelines on Etsy?
Products created with AI from the seller’s prompts may qualify as “Designed by” on Etsy, with AI use disclosed in the listing description. Prompt bundles cannot be sold on their own. The public policy does not substantiate claims that every product photo is automatically scanned for metadata. Distinguish the product itself from the image used to list it.
Can I use the same AI-generated image on more than one marketplace?
Technically yes, but channel rules differ. Google Merchant Center requires the IPTC
TrainedAlgorithmicMedia tag for generative AI imagery; Amazon uses a different XMP keyword for synthetic people; TikTok Ads offers a disclosure tool. Meta’s general content label is separate from store product-image requirements. Maintain a disclosure matrix by SKU and channel.H · Technical: Detection, watermark, and authentication
Can AI-generated images really be detected?
Occasionally and probabilistically. Supported Google outputs may carry SynthID; Google Merchant Center requires IPTC, and Amazon requires XMP tags for certain synthetic human images. C2PA and metadata are strong indicators of origin, but they can be removed; image classifiers can produce false positives and false negatives. A single detection score should not be considered definitive proof.
What is C2PA, and what is its purpose in product photos?
C2PA is an open standard that carries a cryptographic manifest for the declared origin and editing history of content; "Content Credentials" applications use this standard. A valid signature does not indicate that the content is absolutely accurate, but rather that the chain of records can be verified. Both real camera footage and synthetic content can carry a C2PA manifest; the source and actions in the manifest must be interpreted together.
Can't the C2PA signature be cracked?
No, it is not absolute proof. Nikon suspended its own authentication service and invalidated the certificates it had issued during that period; as of mid-2026, the service has not resumed. The problem wasn’t with the cryptography itself-the signatures were valid; rather, it stemmed from the fact that the signer was fed manipulated content. The ecosystem is real, but it’s not yet mature. Nevertheless, it is currently the strongest proof-of-origin available.
Is the C2PA signature kept after retouching?
Only if every tool in the chain is C2PA-compatible and the relevant function is enabled. One incompatible tool can break the manifest. Some widely used editing applications also support provenance information only in certain export formats. Test the entire retouching workflow before offering provenance verification.
I · Strategy, framing and communication
How should I explain my use of AI to the customer?
The rationale can be more decisive than the technique itself. In Zhang and Hur’s latest experiment (n=212), the "privacy protection" rationale made AI-generated images comparable to human-created ones, while the "cost-effectiveness" rationale reduced trust and purchase intent. View this not as a universal psychological rule, but as an experimental finding that highlights the importance of the communication frame.
Why did Levi's face backlash but Mango didn't?
The difference lay in the claim and context as much as in the technology. In March 2023, Levi's linked its partnership with Lalaland.ai to diversity and inclusion; following the backlash, it clarified that the plan would not reduce the use of real models. In July 2024, Mango framed its campaign around operational speed and design flexibility; no wave of backlash on the scale of Levi’s was found in the sources reviewed. 'There was no backlash' cannot be proven; a more accurate conclusion is that a social value claim is more sensitive to synthetic representation.
Doesn't the phrase "generated by AI" undermine trust?
A bare label can reduce trust. NIM research finds that AI-labeled images are perceived as less emotional, credible, and memorable. The third experiment in the Journal of Advertising study adds an important distinction: in ads with abstract purpose messages, a well-developed transparency statement alongside a simple AI disclosure improves brand evaluations. So the problem is a label left on its own.
How do you write a qualified transparency statement?
The researchers’ recommendation has three components: explain the use of AI; present this explanation alongside information about the brand’s concrete, real-world commitments; and describe how the AI was developed responsibly. For example, explaining that the AI model was trained on a dataset of real human photos obtained with consent and under license can reduce consumer skepticism.
What was the difference between Guess and H&M?
Both used AI; the difference was created by the architecture of transparency. The disclosure in Guess’s August 2025 Vogue ad was in small print in the corner of the page; a TikTok video highlighting this garnered over two million views, and the backlash was directed at both the brand and the magazine. H&M, on the other hand, proceeded by obtaining consent from its 30 models, paying royalties, and using a prominent watermark; it faced criticism, but the debate centered on labor practices rather than deception.
Is “No AI” the right message for every brand?
No. For some brands it can be a strong trust signal; for others, using AI openly in a limited, supporting role is more appropriate. “No AI” should be an auditable commitment, not a slogan. Unverifiable claims can pose risks under general deceptive-advertising rules. New York’s law, effective June 2026, is narrower: it regulates disclosures in ads with synthetic performers.
How else can I break the perception that the image is artificial?
With concrete product data. Research shows that verifiable technical details added alongside an image can soften the negative impact of the perception of artificiality. When fabric composition, full size range, measurements in centimeters, weight, material certifications, and color codes are provided, the consumer’s focus shifts from how the image was produced to the product’s actual qualities.
J · Cost, process, and implementation
How cheap is AI-generated imagery?
AI-generated content can be much cheaper for most tasks. In the peer-reviewed Hartmann-Exner-Domdey study’s controlled general marketing brief, a DALL-E 3 image cost $0.04, while an image commissioned from a freelancer cost $100. This 2,500-fold difference in task cost does not represent the universal price for a physical SKU shoot; samples, models, sets, color management, and usage licenses are separate line items.
What are the compliance costs of real production?
Most real shoots do not trigger AI-specific disclosure requirements, but their total compliance cost is not zero. Model releases, personality rights, copyright, product accuracy, advertising claims, and platform technical rules still apply. Real photography makes the physical source and human creative contribution easier to document.
Does LUX Photo Video Production also produce AI images?
Yes. LUX produces AI-assisted photo and video imagery according to clients’ needs. We also offer professional production shoots in Istanbul and Berlin with real sets, models, products, and lighting. We follow one principle: every image representing a product starts with a real shoot; AI works on the derivative layer.
How does LUX Production decide whether or not to use AI in a project?
For each project, the category, target audience, budget, distribution channel, and trust risk are evaluated together. The approach is based on four principles: the source plate principle (product visuals are always derived from actual photography), declaration architecture (how each layer is produced is documented), framing consultation (how the use of AI will be explained to the client is established collaboratively), and the option for provenance verification upon request.
Why real photography is still strong in e-commerce
Real photos are powerful not because they are always more beautiful or generate more clicks, but because they can document the physical SKU. Real lighting, surface, scale, and accurate color management provide evidence about the product. To this can be added the documentation of human creative contributions and a reliable record of origin; however, general advertising and rights management obligations still apply.
Conclusion: the winning strategy is evidence-based visual management
In e-commerce, visuals are the product’s physical form in the digital world. The credibility of that form takes precedence over its beauty. Artificial intelligence can add speed, variety, and production flexibility to this representation; but the moment it alters the product’s physical characteristics, the consumer’s need for verification and the brand’s reputational risk both increase.
The academic literature does not offer a single universal conclusion. A practical synthesis of the findings on this page points to three risk areas . Incorrect positioning, Presenting AI as a substitute for a value commitment. Uncontrolled production: visual anomalies that distort the physical reality of the product. Lack of transparency: a disclosure that is hidden or left bare. Once these three are managed, AI delivers a real operational gain.
Main images representing the physical product, including model shots, usage scenes, fabric, texture and color images, cosmetic results, and health claims, should use a real photographic source plate, accurate color management, and a transparent editing process. The more measured lesson from the Vinted preprint is also that: The performance of AI-generated and real photos varies depending on the season, category, aesthetic quality, and the buyer’s need to verify the product.
For images with products, surfaces, and lighting created on a physical set, see our Product photography with props on a physical set service.
The change in 2026 is that this is now more than a quality preference. Turkey and Europe have different transparency obligations; copyright protection is being defined through human contribution; and some major channels require metadata. The value of real production comes from its evidential value, documented human creative contribution, and record of physical origin , rather than aesthetic superiority.
At LUX Photo Video Production, we work with both approaches. We use AI’s operational capabilities while preserving the trust that real production and a photograph composed by a human eye bring to the product.
Let’s work together to develop the right visual strategy for your products.
LUX Photo Video Production | Istanbul & Berlin
Legal Notice: The legislation summaries in this article are for general information only and are not legal advice. Regulations may change; base campaign and listing decisions on the current official texts and, where needed, on your legal counsel’s opinion. Platform policies change often. Before a major catalog change, check the current policy page in the relevant seller dashboard.
References: academic studies, consumer research, legislation, platforms and media sources
Source reading note: Peer-reviewed publications, preprints, institutional surveys, and official policies do not carry the same weight of evidence. In the text, these types have been distinguished; definitive figures for which a primary source could not be found have been excluded as benchmarks. Access and policy review date: August 24, 2026.
Academic publications and compilations
Only Peer-reviewed Articles bearing this label were verified during this review using journal and DOI records; institutional notes, encyclopedia entries, and press coverage are listed separately.
- Peer-reviewed Hartmann, J., Exner, Y., & Domdey, S. (2025). The Power of Generative Marketing: Can Generative AI Create Superhuman Visual Marketing Content? International Journal of Research in Marketing, 42(1), 13-31. DOI · TUM research summary.
- Peer-reviewed Xie, Q., Muralidharan, S., & Phua, J. (2026). How Women Respond to Computer-Generated Inclusive Advertising: Advocating for Body Positivity in the Age of AI. Journal of Advertising. DOI · SMU Summary.
- Peer-reviewed Brüns, J. D., & Meißner, M. (2024). Do you create your content yourself? Using generative artificial intelligence for social media content creation diminishes perceived brand authenticity. Journal of Retailing and Consumer Services, 79, 103790. DOI.
- Peer-reviewed Belanche, D., Ibáñez-Sánchez, S., Jordán, P., & Matas, S. (2025). Customer reactions to generative AI vs. real images in high-involvement and hedonic services. International Journal of Information Management, 85, 102954. DOI.
- Peer-reviewed Cicek, M., Gursoy, D., & Lu, L. (2025). Adverse impacts of revealing the presence of Artificial Intelligence technology in product and service descriptions on purchase intentions. Journal of Hospitality Marketing & Management, 34(1), 1-23. DOI.
- Zhang, L., & Hur, C. (2025). The Impact of Generative AI Images on Consumer Attitudes in Advertising. Administrative Sciences, 15(10), 395. DOI.
- Peer-reviewed To, R. N., Wu, Y. C., Kianian, P., & Zhang, Z. (2025). When AI Doesn't Sell Prada: Why Using AI-Generated Advertisements Backfires for Luxury Brands. Journal of Advertising Research, 65(2), 202-236. DOI.
- Institutional Note Buder, F., & Unfried, M. (2025). Transparency without trust. NIM INSIGHTS, 7, 36-41.
- Peer-reviewed Bui, H. T., Filimonau, V., & Sezerel, H. (2024). Ai-thenticity: Exploring the effect of perceived authenticity of AI-generated visual content on tourist patronage intentions. Journal of Destination Marketing & Management, 34, 100956. DOI.
- Encyclopedia entry Desveaud, K., & Pavone, G. (2026). Consumer Perceptions of AI-Generated Marketing Content. In R. Ladhari (Ed.), Encyclopedia of Artificial Intelligence in Marketing. Springer, Cham. DOI.
- Literature review Baryshkov, K. (2026). Consumer Trust in AI-Generated Marketing Content: A Systematic Literature Review and Research Agenda. American Impact Review. DOI. The journal’s peer-reviewed status could not be independently verified during this review; it was not used as a peer-reviewed primary study.
- Press release Sands, S. et al. Swinburne University of Technology, Four experimental studies on AI ad aversion. Presented by: Forbes, May 28, 2026; no citation for a fully peer-reviewed article was found.
- Peer-reviewed Mori, M. (2012). The Uncanny Valley [from the field]. K. MacDorman & N. Kageki, Trans. IEEE Robotics & Automation Magazine, 19(2), 98-100. DOI (Originally published in 1970).
Experimental and field studies
- Peer-reviewed Tainted by technology? The effect of AI-generated image labels on consumers' food purchase intentions. (2026). British Food Journal, 128(6), 2214-2230. DOI. Three scenario experiments examining perceived authenticity, seller effort, and price premiums.
- Peer-reviewed Israfilzade, K. (2025). AI-generated versus human-created advertising: Effects on consumer trust and purchase intent. Equilibrium: Quarterly Journal of Economics and Economic Policy, 20(4), 1301-1337. DOI. 2×2 layout; snacks and a laptop.
- The Zhang & Hur experiments on the regulatory role of motivation for AI use (n=130, n=79, final experiment n=212), a comparative analysis of the effects of the "privacy protection" and "cost-effectiveness" justifications.
- A fashion retail study conducted by the Münster School of Business (FH Münster) and CECIRE in partnership with Ernsting's family, partial acceptance of decorative images, but strong criticism of personal and model images.
- Pre-print Pec, M., Proszowska, A., & Rzewnis, G. (2026). When Do AI-Generated Product Images Work? Visual Authenticity, Trust, and Seasonal Engagement in C2C E-Commerce. DOI / Preprints.org. It is not peer-reviewed; it reports seasonal reversals across two Vinted waves.
- Kishnani, D. (2025). The Uncanny Valley: An Empirical Study on Human Perceptions of AI-Generated Text and Images (thesis).
Consumer research and industry data
- Organizational Survey The Harris Poll, 4A's & Infillion (June 2026). Study on AI-generated advertising and consumer confidence. Primary summary. Since the publicly available abstract did not specify the sample size, the sample size was not included in the text.
- Organizational Survey IAB & Sonata Insights (2026). The AI Ad Gap Is Widening. 505 U.S. Gen Z and Gen Y consumers and 104 advertising executives; October 2025-January 2026.
- Fractl & Search Engine Land (June 2026). AI search adoption is on the rise as consumer trust declines (1,008 U.S. consumers, 150 marketers). This is an AI search and general content research study; it is not a product photo experiment.
- DoubleVerify (2026). Global Insights: Media Quality in the Age of AI, EMEA edition. Source: Advanced Television, https://www.advanced-television.com/2026/08/06/study-ai-slop-poses-growing-risk-to-brand-trust
- Klaviyo (2026). Consumer Trust in AI: What Brands Need to Know. This article uses its findings that 61% were neutral and 32% reported lower trust.
- Clutch (June 2026). “AI in Branding” study (408 consumers). https://clutch.co/resources/ai-in-branding
- Organizational Survey Gartner (March 16, 2026). 50% of Consumers Prefer Brands That Avoid Using GenAI. 1,539 U.S. consumers in October 2025.
- Cint. (2026). 63% of U.S. Consumers Believe that Brands have a Moral Duty to Disclose AI-generated Content. https://www.cint.com/newsroom/63-of-u-s-consumers-believe-that-brands-have-a-moral-duty-to-disclose-ai-generated-content/
- EMARKETER. (2026). Shoppers aren't impressed by AI-generated marketing. https://www.emarketer.com/content/shoppers-aren-t-impressed-by-ai-generated-marketing
- Emplifi. (2026). AI and authentic reviews: what consumers really trust. https://emplifi.io/resources/infographic-ai-and-authentic-reviews-what-consumers-really-trust/
- Research Pew Research Center (September 17, 2025). How Americans View AI and Its Impact. The 76% and 53% findings concern the general U.S. media context.
- Storyblok & OnePoll (2023). Survey on AI-based purchasing recommendations among 1,000 consumers. Separate study: Storyblok & OnePoll (2022), General web experience and cart abandonment among 6,000 people.
- Organizational research Lee, C. H. (2025). Will consumers be able to tell the difference between real and AI-generated images in 2025? 301 U.S. adults; comparison with the 2023 and 2024 waves.
- Industry data NRF & Happy Returns (2025). The 2025 Retail Returns Landscape: an estimated 19.3% of online sales returned.
- Coresight Research / 3DLOOK (2023). The True Cost of Apparel Returns. Sponsored by 3DLOOK; 100 U.S. decision-makers; a 24.4% online apparel return rate.
- Stylitics & Aha Studio (2026). On-Model AI Imagery in Fashion. 411 shoppers; company-sponsored survey, self-reported attitudes.
Legislation and case law
- Official Ministry of Trade of the Republic of Turkey (July 1, 2026). Summary of the Amendment to the Regulation on Commercial Advertising and Unfair Commercial Practices. Official Gazette No. 33297; effective August 1, 2026.
- Official European Commission. Transparency obligations under Article 50 of the AI Act.
- European Commission. Guidelines on transparency obligations for providers and deployers of certain AI systems. https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations
- Official Regulation (EU) 2026/1744, Digital Omnibus on AI. Effective Date: July 27, 2026 · European Commission summary.
- Goodwin (August 2026). Not Delayed, Not Deferred: EU AI Act Transparency Obligations Are Now in Force. https://www.goodwinlaw.com/en/insights/publications/2026/08/alerts-technology-dpc-eu-ai-act-transparency-obligations-now-in-force
- Morgan Lewis (August 2026). The EU AI Act's Transparency Rules: What Took Effect on August 2. https://www.morganlewis.com/blogs/sourcingatmorganlewis/2026/08/eu-ai-acts-transparency-rules-what-went-into-effect-on-2-august
- U.S. Copyright Office (January 29, 2025). Copyright and Artificial Intelligence, Part 2: Copyrightability.
- Official Thaler v. Perlmutter, No. 25-449 (U.S., March 2, 2026), cert. denied. The Supreme Court did not rule on the merits of the appeal; the lower court’s decision recognizing human authorship remained in effect in this case.
- Court ruling Munich Local Court, February 13, 2026, 142 C 9786/25. Bayern.Recht ruling text.
- Official New York State (June 9, 2026). Disclosure law for advertisements featuring synthetic performers · General Business Law § 396-b.
- Federal Trade Commission. (2024). FTC Announces Crackdown on Deceptive AI Claims and Schemes. https://www.ftc.gov/news-events/news/press-releases/2024/09/ftc-announces-crackdown-deceptive-ai-claims-schemes
Platform policies and technical standards
- Platform Etsy Seller Handbook. What is Etsy's stance on AI-generated creations?
- Platform Amazon Seller Central. Product image guide · How to tag media that features an AI-generated person.
- Platform Google Merchant Center. AI-generated content and IPTC metadata.
- Platform TikTok Ads. Policy on misleading and false content · AI-generated content disclaimer.
- Meta. (2024). Our Approach to Labeling AI-Generated Content and Manipulated Media. https://about.fb.com/news/2024/04/metas-approach-to-labeling-ai-generated-content-and-manipulated-media/
- YouTube (2026). Improving AI labels for viewers and creators. https://blog.youtube/news-and-events/improving-ai-labels-viewers-creators/
- C2PA. Verifying Media Content Sources. https://c2pa.org/
- Sony. Camera Authenticity Solution, supported camera bodies and verification service. https://authenticity.sony.net/camera/en-us/
- Google DeepMind (November 20, 2025). Introducing Nano Banana Pro (Gemini 3 Pro Image). https://blog.google/innovation-and-ai/products/nano-banana-pro/
- TechCrunch (November 20, 2025). Google releases Nano Banana Pro, its latest image-generation model. https://techcrunch.com/2025/11/20/google-releases-nano-banana-pro-its-latest-image-generation-model/
Brand stories and the press
- Primary Merriam-Webster (2025). "Slop", 2025 Word of the Year · Macquarie Dictionary (2025). "AI slop", Word of the Year.
- Brand source Mango Fashion Group (July 10, 2024). Mango Teen / Sunset Dream AI Campaign.
- Brand source Zalando (May 7, 2025). AI content and digital twin pilot. Approximately 70% of Q4 2024 editorial campaign assets.
- Brand source H&M Group (July 2, 2025). The first set of images featuring digital twins.
- The Guardian, The Independent, The New York Times (March 2023), Reports on the backlash against the Levi Strauss & Co. and Lalaland.ai partnership.
- TechCrunch (August 3, 2025). The uproar over Vogue’s AI-generated ad isn’t just about fashion. https://techcrunch.com/2025/08/03/the-uproar-over-vogues-ai-generated-ad-isnt-just-about-fashion
- Forbes (July 29, 2025). Vogue Erupts: AI-Generated Models Spark Reader Fury and Industry Panic. https://www.forbes.com/sites/moinroberts-islam/2025/07/29/vogue-erupts-ai-generated-models-spark-reader-fury-and-industry-panic/
- FashionNetwork (2025). Vogue US faces backlash over a Guess ad featuring an AI-generated model. https://us.fashionnetwork.com/news/Vogue-us-faces-backlash-over-guess-ad-featuring-ai-generated-model,1754907.html
- Retail Dive. Aerie’s “100% Aerie Real” campaign and its anti-AI pledge. https://retaildive.com/news/aerie-real-pamela-anderson-campaign-against-ai-advertising/815942/
- DesignRush (2026). 7 Brands Using 'No AI' Disclaimers to Win Consumer Trust. https://news.designrush.com/no-ai-disclaimers-brands-consumer-trust-2026
- EMARKETER. AI in retail: Balancing automation with the human touch (The Coca-Cola case). https://www.emarketer.com/content/ai-retail-balancing-automation-with-human-touch
- Moneywise (2026). 'Cheap' and 'lazy': How AI-generated images signal lower quality. https://moneywise.com/life/lifestyle/ai-generated-images-brand-quality-human-made-luxury