Photo by Zulfugar Karimov on Unsplash
Film people have long known that an audience will happily accept an invented image as long as it knows the image is invented. Nobody feels deceived by a matte painting or a digital creature. The trouble starts when a picture quietly asks viewers to believe something it cannot support. Generative AI makes that line easier to cross.
This piece walks through what an overview of research on trust in AI-generated images, published by Everypixel in September 2026, says about labeling and disclosure, and what that means for brand teams, publishers and film marketers.
The headline finding: disclosure shifts trust, but not equally everywhere
The central study in the overview is a 2026 CHI paper involving 43 participants and six websites. According to the overview, participants initially showed a preference for some AI-generated images. After they learned which images were AI-generated, trust shifted toward websites that used photographs. The important detail is that the shift was not uniform: it was more pronounced in government and health contexts than in entertainment, where participants showed greater tolerance for AI imagery.
So the honest summary is not “disclosure destroys trust.” It is closer to this: once people know an image is synthetic, they re-evaluate it against what they expected it to represent. Where they expected documentation, trust moved toward photographs; where they expected fiction, the shift was smaller.
Why the label changes the reading
The overview explains this through a distinction worth borrowing for any creative brief:
Realism describes how an image looks.
Authenticity concerns its relationship to what it appears to represent.
Provenance tells us where an image came from and how it was made or changed.
Evidence helps us assess the claim associated with it.
A label changes no pixels; it changes the viewer’s understanding of the claim the image makes. A testimonial portrait implies that a real customer exists; a product photo implies the product looks as shown; a before-and-after image implies the result actually occurred. If the viewer later learns that the image was generated, the implied claim collapses, and trust in the source goes with it.
The overview also notes that looks alone are becoming a poor guide. It cites a 2025 study in which 104 participants classified 50 images, half photographs and half AI-generated from five text-to-image models. Participants correctly identified AI images in 63.7% of cases overall, and for images from FLUX.1-dev accuracy fell to 29%. It also mentions a 2025 study by Aqsa Farooq and Claes de Vreese with 292 UK participants, which found that realistic-looking AI images were more likely to be judged authentic, although participants were less confident in those judgments. Everypixel’s own blind test of AI images against real photographs, with 13 production professionals and 48 scenes, found GPT Image 2 received 46.2% of preferences against photography, with a confidence interval that included 50%; the authors describe this as directional, not as proof that AI beats photography.
If audiences increasingly cannot tell by eye, they learn an image’s origin through disclosure, or later, when someone exposes it.
What it means for brands, publishers and film marketing teams
Brands. The overview’s table of risky contexts reads like a brand checklist: product pages, testimonials, before-and-after imagery. It cites an ABC News investigation that found online sellers using AI-generated images and videos to make businesses appear to be small, family-run operations. Imaginative visuals, such as an impossible landscape, ask nobody to believe they were photographed.
Publishers. News imagery carries the strongest implied claim: that an event happened. The overview points to AI-generated images of aircraft landing near a burning Beirut airport, which Reuters confirmed were made with Midjourney and circulated in a context where they could be read as documentary photographs.
Film marketing teams. Entertainment is where the research found the most tolerance, which is good news for key art, concept art and stylized teasers. But a campaign also makes factual claims: behind-the-scenes stills, real locations, cast at a premiere, audience reactions. There, a synthetic image, once discovered, reads as a misleading claim rather than a creative choice.
Practical disclosure practices
Sort assets by the claim they make. Fictional artwork needs clear context; a concept image needs an accurate description of its status; a product image needs correspondence to the real product; a documentary image needs source, provenance and evidence.
Disclose most where the audience expects documentation. The research suggests the trust shift is stronger in health and government contexts, so treat factual and informational material as high priority.
Distinguish AI-generated from AI-assisted. An image made from a prompt and a photograph with an extended background imply different things. Say which one it is.
Never use synthetic people as real customers, cast or witnesses. This is the testimonial case the overview flags directly.
Attach provenance where you can. The overview describes C2PA Content Credentials, which attach cryptographically signed information about an asset’s creation and editing history.
Do not treat a label as proof. Provenance can show where content came from, not whether the caption or claim is true.
The overview’s own rule of thumb is a good one to pin above the edit suite: the more an image’s AI origin changes what the audience is likely to believe about its subject, origin or meaning, the more important disclosure becomes.
What this research does not tell you
The CHI study is small, with 43 participants and six websites, so it shows a direction of change rather than a reliable size of effect for your audience.
The overview does not report how much trust shifted, or how disclosure compares with being caught without it.
None of the cited studies tests specific label wording, placement or design.
The blind test covered 48 scenes and 13 evaluators, so it does not show that AI images are universally indistinguishable from photographs.
The lesson, as any good film teaches, is about the contract with the audience. Invention is welcome when it is declared. It becomes a betrayal only when the picture pretends to be a record.
