AI Images Have Become Good Enough. So Why Don’t We Trust Them Yet?

What Are Your Ideas Today?

BRING THEM TO LIFE WITH EVERYPIXEL

AI-generated images have become remarkably realistic. Faces are more consistent, lighting is more convincing, and synthetic visuals can now compete with photography in some visual comparisons.

Yet looking real and being trustworthy are different things.

An AI-generated image can show a person who never existed, a place that was never photographed, or an event that never happened. And once people learn that an image was generated by AI, they may interpret it differently.

A 2026 CHI study involving 43 participants and six websites found that disclosure of AI-generated imagery changed trust judgments. Participants initially showed a preference for some AI-generated images, but after learning which images were AI-generated, trust shifted toward websites using photographs. The shift toward photographs was more pronounced in government and health contexts than in entertainment.

The distinction is simple:

• 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.

So if an AI image can look real without documenting something that actually happened, what exactly are we trusting when we trust it?

AI Images Are Getting Harder to Distinguish From Photography

Earlier generations of image generators were often associated with obvious artifacts: distorted hands, inconsistent faces, unreadable text, impossible reflections, or lighting that did not quite make physical sense.

Those clues are becoming less reliable.

A 2025 study involving 104 participants asked people to classify 50 images — half photographs and half AI-generated images from five text-to-image models. Participants correctly identified AI-generated images in 63.7% of cases overall. For images generated by FLUX.1-dev, accuracy fell to 29%. Participants relied on familiar clues such as geometric inconsistencies and unrealistic lighting, while some newer images were described as “too perfect” or subtly uncanny.

The finding does not mean that AI images are impossible to detect. It suggests that visual inspection alone is becoming a less reliable way to establish whether an image is synthetic.

A 2026 blind evaluation by Everypixel illustrates how narrow the perceptual gap can become in some conditions. Thirteen production professionals compared 48 scenes generated by GPT Image 2, Qwen Image 2 and FLUX.2 Klein 9B with real photographs. In direct comparisons, GPT Image 2 received 46.2% of preferences against photography, with a confidence interval that included 50%. The authors describe the findings as directional rather than evidence that AI images outperform photographs.

The benchmark was limited to 48 scenes and 13 professional evaluators, so it does not show that AI images are universally indistinguishable from photographs.

It does show something more limited but increasingly important: under some conditions, visual appearance alone may not provide a reliable answer about how an image was made.

The question is therefore no longer simply:

Does this image look real?

It is:

What does this image give me reason to believe?

Realism Is Not the Same as Authenticity 

Three concepts are easy to conflate:

.trust-table-wrap { overflow-x: auto; margin: 28px 0; -webkit-overflow-scrolling: touch; } .trust-table { width: 100%; min-width: 620px; border-collapse: separate; border-spacing: 0; border: 1px solid color-mix(in srgb, currentColor 18%, transparent); border-radius: 10px; overflow: hidden; font-size: 16px; line-height: 1.55; color: inherit; background: color-mix(in srgb, currentColor 2%, transparent); } .trust-table th { padding: 15px 18px; text-align: left; font-weight: 700; border-bottom: 1px solid color-mix(in srgb, currentColor 18%, transparent); background: color-mix(in srgb, currentColor 6%, transparent); } .trust-table td { padding: 16px 18px; vertical-align: top; border-bottom: 1px solid color-mix(in srgb, currentColor 12%, transparent); } .trust-table tbody tr:last-child td { border-bottom: 0; } .trust-table td:first-child { width: 26%; font-weight: 700; white-space: nowrap; } .trust-table tbody tr:nth-child(even) { background: color-mix(in srgb, currentColor 3%, transparent); }

An image can be realistic without being authentic. It can also be authentic without proving every claim made about it.

A real photograph of a protest, for example, can establish that a photographer captured a particular scene. It does not automatically prove when the photograph was taken, who the people are, why they were there, or whether the accompanying caption is accurate.

A 2025 study by Aqsa Farooq and Claes de Vreese examined how people judge the authenticity of AI-generated disinformation images. In an experiment involving 292 UK participants, the researchers found that realistic-looking AI images were more likely to be judged as authentic, although participants were less confident in those judgments.

That distinction matters.

A convincing appearance can influence an authenticity judgment, but it is not evidence that the depicted event actually happened.

An AI-generated image of a street protest could contain realistic crowds, signs, motion blur and imperfect framing. It might look like documentary photography.

The protest could still be entirely fictional.

Realism describes appearance. Authenticity concerns the image’s relationship to reality.

What an Image Asks Us to Believe Matters

An AI-generated image can be perfectly appropriate in one context and misleading in another.

A fictional landscape for a video game does not need to document a real location. A concept image can show a possible product direction without pretending to depict a finished product. A surreal advertising image can deliberately create a scene that could never exist in the physical world.

In these cases, the audience understands that the image is a creative representation.

The situation changes when an image carries an implied factual claim.

A news photograph suggests that an event happened.

A testimonial portrait suggests that a real person exists and had a particular experience.

A product photograph suggests that the product looks as shown.

A before-and-after image suggests that the visible result actually occurred.

A medical image may suggest that it represents a real condition, procedure or patient.

The issue is therefore not simply whether AI was used. It is whether viewers are given a misleading impression about what the image represents.

That makes context more important than the simple question of whether an image is AI-generated.

The more consequential the factual claim, the more important it becomes to understand the image’s origin and supporting evidence.

Why Disclosure Changes How People Trust AI Images

Disclosure matters because information about an image’s origin can change how viewers interpret it.

The 2026 CHI study found that participants’ trust judgments changed after they learned which website images had been generated by AI. The response varied by context: participants showed greater tolerance for AI imagery in entertainment, while the shift toward photographs was stronger in government and health settings.

This suggests that people do not evaluate an image only on visual appearance. They also consider what they expected the image to represent.

That is why “AI-generated” does not carry the same meaning in every setting.

The distinction between AI-generated and AI-assisted also matters.

One image may be created almost entirely from a prompt. Another may begin as a photograph and use generative tools to extend a background, remove an object or alter part of a scene.

Those changes can have very different implications for what the audience understands the image to represent.

A useful principle is:

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.

Disclosure, then, is not just about attaching an “AI” label. It is about giving viewers enough information to interpret the image appropriately.

Photography Is Evidence, But Not Proof

It would be too simple to divide images into two categories:

photography = trustworthy
AI = untrustworthy

Real photographs can be staged, edited, cropped, retouched, miscaptioned or taken out of context.

What traditional photography can provide is a physical connection to a captured scene. Light from an environment reaches a camera sensor or film, creating a record of something that was physically present at the time of capture.

Generative AI does not require the same kind of underlying scene.

It can produce a convincing representation of a person, place or event without that person, place or event ever existing in the depicted form.

That does not make every photograph truthful or every AI image misleading.

It means the two forms of image-making can provide different kinds of evidence.

A photograph can be one piece of evidence that something was physically captured. Additional information — such as metadata, the photographer’s account, other images, video, witnesses or provenance records — can help establish context.

An AI-generated image may require different supporting information because its visual appearance does not itself establish that the depicted scene existed.

The useful question is therefore:

What connects this image to the reality it claims to represent?

Sometimes the answer is a photograph.

Sometimes it is provenance data.

Sometimes it is the source publishing the image.

Sometimes it is independent documentation.

And sometimes there is no factual claim at all because the image is simply a creative representation.

When AI Images Do Not Need to Prove Anything

Not every AI-generated image creates an evidence problem.

AI imagery can work naturally in:

• concept art;
• illustration;
• fictional characters and worlds;
• moodboards;
• early visual concepts;
• entertainment;
• decorative social media content;
• advertising that clearly uses fictional or imaginative scenes.

Imagine a campaign showing a sneaker floating above an impossible landscape or a fictional city inside a glass bottle.

The audience is not being asked to believe that these scenes were photographed in the real world. Their value comes from the concept and execution.

The same applies during creative development. Designers can use AI to explore compositions, directors can create moodboards, and brands can test visual directions before producing a final campaign.

In these contexts, AI is simply another image-making tool.

The problem begins when the visual is presented as evidence of something factual.

When AI Images Can Mislead

The stakes change when viewers reasonably expect an image to correspond to reality.

.ai-context-table-wrap { overflow-x: auto; margin: 32px 0; -webkit-overflow-scrolling: touch; } .ai-context-table { width: 100%; min-width: 900px; border-collapse: separate; border-spacing: 0; border: 1px solid color-mix(in srgb, currentColor 18%, transparent); border-radius: 10px; overflow: hidden; font-size: 15px; line-height: 1.6; color: inherit; background: color-mix(in srgb, currentColor 2%, transparent); } .ai-context-table th { padding: 16px 18px; text-align: left; vertical-align: top; font-weight: 700; line-height: 1.4; border-bottom: 1px solid color-mix(in srgb, currentColor 18%, transparent); background: color-mix(in srgb, currentColor 7%, transparent); } .ai-context-table td { padding: 17px 18px; vertical-align: top; border-bottom: 1px solid color-mix(in srgb, currentColor 12%, transparent); } .ai-context-table tbody tr:nth-child(even) { background: color-mix(in srgb, currentColor 3%, transparent); } .ai-context-table tbody tr:last-child td { border-bottom: 0; } .ai-context-table td:first-child { width: 18%; font-weight: 700; white-space: nowrap; } .ai-context-table td:nth-child(2) { width: 27%; } .ai-context-table td:nth-child(3) { width: 55%; } .ai-context-table a { color: inherit; font-weight: 600; text-decoration: underline; text-decoration-thickness: 1px; text-underline-offset: 3px; } .ai-context-table a:hover { opacity: 0.7; }

The dividing line is not whether AI was involved.

It is whether the image creates a misleading impression about reality.

An AI-generated product concept can be useful when it is presented as a concept. The same image could be misleading if presented as a photograph of a product that customers can actually buy.

Can Provenance Make AI Images More Trustworthy?

If visual inspection is becoming less reliable, another option is to provide information about where an image came from.

This is the idea behind content provenance.

Provenance systems can record information about an asset’s origin and editing history. Depending on the implementation, this can include the creation tool, editing actions, timestamps and assertions made by organizations or devices.

One of the most prominent efforts is C2PA, the Coalition for Content Provenance and Authenticity. Its Content Credentials framework is designed to attach cryptographically signed provenance information to digital assets so users can inspect aspects of their creation and editing history.

That information could become increasingly useful as AI images become harder to identify from appearance alone.

But provenance has a clear limit:

Provenance is not the same as truth.

C2PA distinguishes information about an asset’s provenance from factual accuracy. Provenance can help establish where content came from and what happened to it, but it cannot by itself establish that the content is factually accurate or that the claims surrounding it are true.

For example, provenance might tell you:

• Where did this image come from?
• Who created or published it?
• How was it made?
• Was generative AI or another editing process used?

But it cannot automatically tell you:

• Did the event depicted actually happen?
• Is the caption accurate?
• Is the claim surrounding the image true?

Technology can establish an image’s history without replacing independent evidence.

The Real Problem Is Unsupported Claims

This leads to a more useful way of thinking about visual trust.

Instead of asking whether AI images are trustworthy in general, ask what claim the image is being used to support.

Consider four different cases.

.image-context-table-wrap { overflow-x: auto; margin: 30px 0; -webkit-overflow-scrolling: touch; } .image-context-table { width: 100%; min-width: 700px; border-collapse: separate; border-spacing: 0; border: 1px solid color-mix(in srgb, currentColor 18%, transparent); border-radius: 10px; overflow: hidden; font-size: 15px; line-height: 1.55; color: inherit; background: color-mix(in srgb, currentColor 2%, transparent); } .image-context-table th { padding: 16px 18px; text-align: left; vertical-align: top; font-weight: 700; line-height: 1.4; border-bottom: 1px solid color-mix(in srgb, currentColor 18%, transparent); background: color-mix(in srgb, currentColor 7%, transparent); } .image-context-table td { padding: 17px 18px; vertical-align: top; border-bottom: 1px solid color-mix(in srgb, currentColor 12%, transparent); } .image-context-table tbody tr:nth-child(even) { background: color-mix(in srgb, currentColor 3%, transparent); } .image-context-table tbody tr:last-child td { border-bottom: 0; } .image-context-table td:first-child { width: 24%; font-weight: 700; white-space: nowrap; } .image-context-table td:nth-child(2) { width: 40%; } .image-context-table td:nth-child(3) { width: 36%; }

The same technology can therefore produce images with very different evidentiary implications.

An AI-generated fantasy landscape does not need to prove that its landscape exists.

An AI-generated image presented as a photograph of a real event does.

The difference lies not in the pixels alone, but in the claim attached to them.

What Will Visual Trust Look Like Next?

AI images may become even more convincing.

That will not necessarily make them more trustworthy. It may simply make visual appearance a weaker signal of how an image was created and what it represents.

Earlier generations of synthetic imagery often revealed themselves through obvious technical mistakes. Those artifacts gave viewers a simple detection strategy: look for the mistake.

As those mistakes become less common, that strategy becomes less useful.

A perfectly formed hand does not prove that an image is a photograph.

Correct lighting does not prove that a scene existed.

Natural skin texture does not prove that the person depicted exists.

There is also a reverse problem: genuine photographs of unusual events may become easier to dismiss as synthetic when viewers know how convincing generated imagery can be.

That makes context increasingly important.

Viewers may need to ask:

• Who created or published this image?
• How was it created or changed?
• What is it supposed to represent?
• What evidence supports the claim surrounding it?

The answers will matter differently depending on context.

A fictional AI image does not need to prove that its subject exists.

A synthetic advertising visual can be useful when its fictional nature is clear.

An AI-assisted photograph can remain valuable when relevant alterations are understood.

A documentary image can carry stronger evidentiary weight when its source, provenance and surrounding evidence can be established.

The better AI becomes at imitating photographic realism, the less useful realism itself becomes as evidence.

Visual trust will increasingly depend not on whether an image looks real, but on whether its source, context, provenance and purpose support what the image asks us to believe.