Do shoppers prefer real photos to AI photos?

In Getty Images research reported by CX Network in 2026, 38% preferred real images and 15% preferred AI-generated images in marketing materials. This is a stated preference from a vendor survey, with no sample given in the report, not a measure of buying.

Do AI product photos increase returns?

This is not known from independent data. In a vendor survey by Stylitics, 30% of 411 shoppers expected they might have to return an item shown on a virtual model, which is an expectation, not a measured return rate. An Alibaba test reported 7.9% lower returns, but it compared AI-designed items with human-designed items, not photos of one product.

Does an AI label stop a backlash?

A label does not prevent one. Valentino's December 2025 ad and Gucci's February 2026 images were both labelled as AI and still drew public criticism. Hiding AI carries its own risk, as J.Crew's August 2025 case shows.

Are younger shoppers more sceptical of AI images?

Gen Z is the more sceptical group in the IAB and Sonata Insights survey of January 2026: 39% of Gen Z felt negative about AI-generated ads, against 20% of Millennials. The survey covered 505 US consumers aged 16 to 43, so subgroup figures carry more uncertainty.

Does trust in AI images differ by country?

Yes, according to Getty Images research reported by CX Network in 2026, consumers in Southeast Asia, Latin America, the UAE and Saudi Arabia are much more receptive to AI imagery. The report gives this as a vendor finding, and this page has no matching independent data.

Do customers trust AI product photos? 2026 evidence

What do the studies say about trust in AI images?

Studies of AI images show a consistent split: people accept AI images they have not identified, and rate them lower once AI is disclosed or suspected. No study in this review, dated 31 October 2026, measured trust in AI product photos on a live store. The table therefore holds nearby evidence: opinion surveys, advertising tests and experiments on simulated product pages.

Read the last column first. Vendor surveys are marked: Getty Images licenses photographs, and the other vendors sell services to brands. The peer-reviewed rows are stronger on method, but several rest on small or convenience samples.

Who and when	Sample	Finding	Type
Pew Research Center, fieldwork 9 to 15 June 2025	5,023 US adults	76% say it is extremely or very important to tell whether pictures, videos and text were made by AI or people. 53% are not too or not at all confident they can tell.	Independent survey
Getty Images VisualGPS, released 30 April 2024	Over 30,000 adults in 25 countries, 7,500 per survey. The release dates the research both as 2022 to 2024 and as July 2022 to September 2023.	Almost 90% want to know whether an image was created with AI. 98% agree authentic images and videos are pivotal in establishing trust. 76% agree it is getting to the point where they cannot tell if an image is real.	Vendor survey
IAB and Sonata Insights, 15 January 2026	505 US consumers aged 16 to 43, 104 ad executives	45% of consumers feel positive about AI-generated ads, against the 82% that executives assume. Negative sentiment is 39% among Gen Z and 20% among Millennials. 73% say knowing an ad used AI would raise their likelihood of buying or make no difference.	Trade body survey
Clutch, 18 September 2025	Not stated in the release	66% felt confident spotting AI photos, 57% were wrong when tested, 84% say disclosure is important, and nearly 40% would trust a brand less without it.	Vendor survey
Stylitics and Aha Studio, undated	411 shoppers	76% call on-model photos the most useful format. 60% reacted neutrally or positively on learning an image was AI, 31% negatively. 59% want clear labels.	Vendor survey, apparel
NIQ, 12 December 2024	Over 2,000 people, about 150 with EEG	Participants identified most AI ads and found them more annoying, boring and confusing. Even ads rated high quality produced weaker memory activation than traditional ads.	Vendor study, ads
Nielsen Norman Group, 21 August 2026	77 US adults, six website hero images, participants not told about AI	AI images were not at a disadvantage against stock photos. Only authenticity differed significantly, by 0.4 points on a 7-point scale in favour of AI. Participants who suspected AI tended to rate the site lower.	Independent usability study, not product images
Xie, Muralidharan and Phua, Journal of Advertising, 2026	Over 700 US women aged 18 to 54, three studies	Disclosing AI plus-size models lowered brand attitude and purchase intention. One added sentence, saying the AI was trained only on licensed photos of real people, brought results level with no AI.	Peer-reviewed
Yang, Liu, He and Chu, Frontiers in Computer Science, 21 August 2026	276 people recruited through WeChat, Xiaohongshu and personal networks, simulated product page	For a power bank, an AI label lowered perceived authenticity, aesthetic appeal and social presence, and its total effect on brand trust was negative. That pattern did not appear for perfume. The direct effect on purchase intention was positive (b = 0.308).	Peer-reviewed
Wei, Liu, Wang and Sun, Frontiers in Psychology, 21 October 2026	410 Chinese respondents in a randomised online experiment, 81.5% aged 18 to 24	An "AI-generated" label on illustrations of Chinese ethnic-minority fashion had a negative total effect on purchase intention (-0.437).	Peer-reviewed
Ghose, Lee, Todri and Adamopoulos, 2025	Lab and field study. Sample not given on the pages read.	Fully AI-made ads raised click-through by up to 19% over expert-made ads. Disclosing AI involvement cut effectiveness by up to 31.5%.	Working paper, ads
Schilke and Reimann, Organizational Behavior and Human Decision Processes, 2025	13 experiments, over 5,000 participants	People who disclose AI use are trusted less, and trust falls furthest when hidden use is discovered by others.	Peer-reviewed, not product images

Can shoppers tell if a product photo is AI-generated?

Shoppers cannot reliably tell: in a Microsoft AI for Good paper posted in 2025, people separated real images from AI images correctly 62% of the time. Microsoft sells AI tools. The study collected about 287,000 evaluations from over 12,500 people in August 2024, and its authors state the result likely overestimates people's ability today.

Earlier work on faces reached the same conclusion. Nightingale and Farid (PNAS, 2022) found 48.2% accuracy, close to chance, among 315 participants judging StyleGAN2 faces. Training and feedback raised it to 59.0% among 219 participants.

These tests used general images, not retail pages. Two vendor surveys come closer: Clutch reports that 66% felt confident and 57% were wrong when tested, though the release does not define wrong. Stylitics reports 71% for apparel images, and words that figure in two ways on one page: could not tell the difference, and saw the same or only small differences.

Not being able to spot an AI image does not settle the question of trust. The customer compares the photo with the parcel, and that test does not depend on detection.

Should you label AI product photos?

Follow the rule of each channel first, because law and marketplace policy already decide much of this question. Beyond those rules, labelling is a brand decision, and the evidence below does not point one way.

People say they want labels: almost 90% in Getty's survey, 84% in Clutch's and 59% in Stylitics'. In experiments, plain disclosure usually costs response, as the Ghose, Xie and Wei rows show. Stated preference and measured behaviour differ here, and IAB's 73% is a stated answer.

Hiding AI carries the larger risk. Schilke and Reimann found the steepest trust loss when concealed use is discovered, and the J.Crew case below followed that pattern.

The wording of a label appears to matter. In the Xie study a plain label was not enough, while a sentence explaining how the AI was trained restored results. That is one study of adverts, not of product pages, so treat it as a lead. A practical reading: keep the machine-readable marker in every file, and where a visible note is allowed, state what is AI and what is real, such as an AI model with the garment's measurements taken from the real item.

Check a file with the AI image label checker

Which rules already decide the label?

Five rule sets decide most cases in 2026. Under EU AI Act Article 50, which applies from 2 August 2026, providers must mark output in a machine-readable format, with systems already on the market given until 2 December 2026, and deployers must visibly disclose deep fakes. Google Merchant Center asks for IPTC metadata, and Amazon asks for a keyword on photorealistic AI-generated people. Zalando does not allow visible on-image labels and says not to provide content that requires visible labelling under applicable laws. In New York, General Business Law section 396-b has required conspicuous disclosure of synthetic performers in advertising since 9 June 2026. This page is not legal advice.

Read the 2026 AI image rules

Will AI models hurt my brand?

AI models are most likely to hurt a brand when they contradict what the brand promises. The Xie study names the mechanism: women judged brands using AI plus-size models as hypocritical, because the models seemed less humanlike while the brand claimed body positivity.

Three promises have proved exposed. Diversity: Levi's was criticised in 2023 for showing diversity without paying diverse models. Craft: critics of Gucci and Valentino objected to AI from houses that sell craft at high prices. Fit: shoppers objected that Atoir's AI imagery could not show how a dress sits on a real body.

Damage is not automatic. Back Row wrote that Valentino, Suit Supply, Coperni and Prada had used AI for ads without headline-making backlash; its events place that article in January 2024, although the page now shows January 2026. A later Valentino AI ad did draw criticism, as the BBC reported in December 2025. Audience matters too: negative sentiment in the IAB survey was almost twice as high among Gen Z as among Millennials.

Rivals can also use the choice against a brand. Dove pledged in April 2024 never to use AI to represent real people in its advertising, and Aerie pledged in October 2025 not to use AI to create images of people or bodies.

The size of the damage is unknown. No brand in this review has published a measured sales effect from a backlash.

Which AI image backlash cases are documented?

Eight dated cases show what triggers a backlash: a claim the image cannot support, concealment, a missing view of fit, or an inaccurate picture. All eight were confirmed in news reports. A full list of brands that use AI models, including those with no backlash, is on the sibling page.

Date	Brand	What happened	Outcome
March 2023	Levi's	Announced a test of AI-generated models from Lalaland.ai with more skin tones and body types. Critics said it avoided paying diverse models.	Levi's said it was not scaling back live shoots, and that AI was not a means to advance diversity.
January 2024	Selkie	Used Midjourney for base images in a Valentine's Day collection. This concerned designs, not models. Artists objected.	The founder said she plans to stop using AI-generated images in future collections.
July 2025	Atoir on The Iconic	A listing disclosed AI imagery with no real model. Reddit users called it deceptive and said a $400 dress could not be judged on a real body.	Neither company had responded when Pedestrian published on 13 July 2025.
July to August 2025	Guess in US Vogue	Ads in the August 2025 issue used models made by the agency Seraphinne Vallora. Critics raised lost work and beauty standards.	Guess and Vogue had not replied to ABC News at publication. The OECD AI Incidents Monitor logged the episode as an incident.
August 2025	J.Crew	Instagram images ran without AI disclosure. Blackbird Spyplane flagged misaligned stripes and a backwards foot.	J.Crew added a caption credit to AI photographer Sam Finn and called the images digital art.
December 2025	Valentino	An Instagram ad for the DeVain bag carried an AI label. Commenters called it sloppy, cheap and lazy.	The BBC article carried no reply from Valentino.
February 2026	Gucci	Images labelled created with AI promoted its Milan show. Critics called them AI slop, while some users praised them.	No reply from Gucci or Kering was reported.
October 2026	Just Eat (food, not fashion)	A trial generated menu images for small restaurants, including sausages and eggs floating in cola.	Just Eat apologised, removed the images and paused the rollout.

When are AI product photos the wrong choice?

AI product photos are the wrong choice when the photo is the evidence the buyer relies on. Second-hand and one-of-a-kind items are the clearest case. A 2026 preprint by Pec, Proszowska and Rzewnis, not peer-reviewed, ran a field experiment on Vinted: AI images did better on attention in autumn and winter, and original photos did better in spring and summer, especially on conversion-related measures. Its abstract gives no effect sizes, and the full text was not read for this page.

Real photography is also the better choice where fit decides the sale and no real size data accompanies the image. Baymard's guidance for apparel says real human models should be used whenever possible and treats mannequins and virtual models as a last resort. The articles read do not test current photorealistic AI models, so this is a caution, not a verdict.

A brand built on real bodies or handwork should treat AI models as a change of promise, not a production decision.

What makes an AI product photo trustworthy?

An AI product photo earns trust when the item delivered matches the image. Accuracy is measurable before publication. Herofab's own check of 303 generated images flagged 20.8%; invented detail was the most common fault at 9.6%, followed by unnatural look at 5.6% and a different product at 4.0%. The check covered images made between 24 September and 12 October 2026, before Herofab's late-October image engine change.

Just Eat shows the cost of skipping that step at scale, and J.Crew shows that small errors such as stripes are what expose an image. Inspect every image against the real product at full zoom, and publish real measurements beside any on-model image.

Dress a model in the real product with Virtual try-on

Checklist

Compare every AI image with the real product at full zoom: print, logo, collar, buttons and colour.
Publish real garment measurements and fabric details next to any on-model image.
Keep the IPTC marker in the file, and check that resizing has not removed it.
Follow each channel's rule on visible labels before adding one.
Where a visible note is allowed, state what is AI and what is real instead of a bare label.
Never present AI imagery as a diversity or body-positivity effort.
Use real photos for second-hand items, unique pieces and anything where condition is the sale.

Sources

Checked 31 October 2026. Rules change; the official pages are the final reference.

Do customers trust AI product photos?