An AI Label Can Change What People Believe, Even When the Content Is True

In June 2024, former White House photographer Pete Souza discovered something strange about one of his photographs on Instagram.

The image, which Souza said he had taken himself, had been labeled "Made with AI." He told TechCrunch that he suspected the label was triggered by changes in Adobe's photo-editing workflow, particularly a process involved in preparing the image for export.

Other photographers reported similar experiences, with real photographs receiving AI labels after relatively minor edits. Meta eventually changed its labeling approach, acknowledging that its earlier system sometimes failed to distinguish limited AI-assisted alterations from content generated entirely by AI.

The controversy exposed an awkward problem. A technical indicator could correctly or incorrectly suggest some connection to AI, while the label presented to viewers implied something much more consequential about the image itself.

Two years later, that distinction matters more than ever.

On October 7, 2026, Google made its SynthID Detector publicly available worldwide in English. The tool checks images, videos, and audio for invisible watermarks associated with participating AI providers, including Google, OpenAI, NVIDIA, and Kakao.

Unlike Meta's earlier automated labeling system, SynthID searches for specific watermark signals embedded by supported AI tools. There is no evidence that Google's newly expanded detector caused the labeling problems photographers experienced on Instagram.

But both developments raise a related question: What happens when people interpret a technical signal about how content was made as a judgment about whether the content deserves to be believed?

Several independent studies suggest this distinction has real consequences. People sometimes rate accurate information as less credible when it's labeled AI-generated, even when the underlying information is true.

And as Google expands access to AI-origin verification, understanding that reaction becomes almost as important as detecting the watermark itself.

SynthID tells you something about origin, not whether a claim is true

Google's SynthID technology embeds imperceptible watermarks into supported AI-generated media.

Those watermarks can subsequently be detected, helping identify whether content originated from or was modified using a participating AI system. Google says the technology is designed to withstand certain common alterations, including compression, cropping, and filtering.

The October rollout expands public access to that capability, which Google previously tested with journalists and other media professionals.

According to the company, SynthID has been applied to more than 180 billion images and videos, plus 240,000 years of audio. Those are Google-reported cumulative deployment figures, not independently verified measurements of detector accuracy or the proportion of online media carrying compatible watermarks.

The Neuron's earlier reporting on the SynthID rollout examined how the technology works, its limitations, and the implications of several major AI companies participating in Google's verification ecosystem.

One distinction is especially relevant here.

A watermark can provide evidence about the production of an image, but it cannot independently establish whether the image's message is true.

Imagine a news organization creates an AI-generated illustration of a real event and clearly identifies it as an illustration. A watermark may help verify its synthetic origin, but that origin alone doesn't make the accompanying reporting inaccurate.

Now consider an authentic photograph shared with a fabricated caption claiming it depicts an event that occurred yesterday. The image may contain no detectable AI watermark, yet the information presented to viewers could still be misleading.

These examples illustrate different verification questions. One concerns the content's origin; the other concerns the accuracy of claims made about it.

Google's detector is designed to help with the first. It doesn't independently resolve the second.

And that distinction matters because the way people interpret AI labels isn't always limited to the information those labels actually provide.

Researchers found that AI labels can make accurate information seem less credible

In October 2024, researchers Sacha Altay and Fabrizio Gilardi published a study in PNAS Nexus examining how people respond to news headlines labeled as AI-generated.

The researchers conducted two preregistered online experiments involving 4,976 participants in the United States and United Kingdom.

Participants evaluated headlines that varied in accuracy, authorship, and labeling. Some headlines were true, while others were false. Some were written by humans, while others were produced using AI.

The researchers found that attaching an AI-generated label reduced participants' ratings of a headline's accuracy and, across the combined measures, their stated willingness to share it.

The effect appeared even when the headline was true or had been written by a human but was incorrectly labeled AI-generated.

The magnitude matters. The researchers reported an average effect equivalent to approximately 2.66 percentage points across their normalized accuracy and sharing measures, compared with 9.33 percentage points for a label explicitly identifying content as false.

Those figures describe changes in experimental ratings, not the percentage of participants who stopped believing the headlines or decided against sharing them. The AI-label effect was relatively small, roughly one-third the effect of explicitly labeling a headline false.

The distinction is important because the findings don't establish a broad collapse in public trust.

In fact, the researchers also tested whether exposure to AI labels reduced participants' general trust in news and journalists. They found no statistically significant effect on those broader attitudes.

What changed was people's evaluation of particular labeled headlines.

The researchers then investigated why.

In their second experiment, participants received different explanations of what an AI-generated label meant. Some were told AI had helped improve the clarity and style of writing. Others were told AI had prepared a draft. A third group was told AI had selected the topic and written the entire article.

The negative effect was stronger when participants believed AI had performed most or all of the work. Providing explanations of more limited AI involvement reduced that effect.

The implication was surprisingly specific: people weren't necessarily treating AI-generated as synonymous with false. They were often assuming that an AI label meant the machine had done far more work than it actually had.

That matters for everything from AI-assisted journalism to marketing materials and creative work.

A writer using AI to improve sentence clarity has made a different editorial decision from someone publishing an unverified article generated entirely by a chatbot. A generic label can obscure that difference.

A larger UK experiment found the same credibility penalty—with limits

More recent research reinforces the finding while providing a useful warning against exaggerating it.

In June 2026, researchers Chuyao Wang, Patrick Sturgis, and Daniel de Kadt published a study in Telematics and Informatics examining how AI labels affect people's judgments about online information.

The researchers used a nationally representative probability sample of 3,861 UK adults, presenting participants with a mock news article about universal basic income, a policy proposal involving regular payments to citizens.

The experiment varied whether the article carried an AI-generated label, whether readers received background information about generative AI, and how the policy argument was presented.

The researchers found that AI labeling reduced the perceived accuracy of the article and participants' interest in the policy it discussed.

But the effect had important limits.

The label did not produce a statistically significant change in respondents' support for universal basic income or their general concern about online misinformation.

Providing background information about generative AI also reduced the negative effect of labeling on perceived accuracy.

This matters because it separates several outcomes that are easy to confuse.

Someone can become less confident in a particular article without changing their opinion about the policy discussed. They can also react skeptically to one AI-labeled item without becoming generally distrustful of everything online.

The researchers' findings support a narrower, more defensible conclusion: AI labels can influence judgments about the accuracy of specific content, but their broader effects depend on context and aren't necessarily large.

The study also involved written policy information, not images, videos, or Google's SynthID interface.

That distinction limits what it tells us about SynthID directly. Still, it establishes a reason to investigate whether the explanations accompanying AI-origin signals influence how people interpret the resulting information.

And the question becomes more complicated when the content being labeled is visual.

Visual AI labels create another problem: what exactly is the image supposed to prove?

A photograph and an illustration may depict the same subject, but readers don't necessarily expect the same thing from them.

A photograph presented as documentary evidence implies that a camera captured a particular scene. An illustration can communicate an idea without claiming that its depicted details occurred exactly as shown.

AI-generated imagery complicates those expectations because it can resemble photography while serving many of the same legitimate purposes as traditional illustration.

In August 2026, University of Vienna researcher Darian Harff published two preregistered experiments in Computers in Human Behavior involving 890 German-speaking participants.

Respondents encountered social-media posts from fictional nongovernmental organizations featuring either real photographs or AI-generated images, with different labeling conditions.

The research examined whether people could identify synthetic imagery and how disclosure affected their assessment of the message and its source.

One experiment involved a predominantly young, highly educated sample that demonstrated relatively strong detection skills. The second used a quota-based sample and found that ordinary respondents had more difficulty identifying AI-generated images without disclosures.

Labels improved detection, although they weren't always successful.

The study also found that labeling could reduce perceived message credibility and source trustworthiness among participants farther from the political center.

But the researcher found no corresponding effect from varying the images' apparent evidentiary value—the extent to which they seemed to document the underlying event.

These are findings about particular political communications, not every type of image or audience. They also don't establish how people would respond to SynthID.

Nevertheless, they expose an important design problem for media verification.

Consider a nonprofit organization publishing an AI-generated illustration to communicate a genuine humanitarian issue. A clear disclosure helps readers understand that the illustration isn't documentary photography.

If the accompanying message is accurate, however, the AI-generated label alone provides no basis for rejecting the underlying facts.

By contrast, a synthetic image falsely presented as photographic proof of an event that never occurred raises a genuine deception problem.

A useful verification process needs to distinguish those situations rather than treating all synthetic visuals as equivalent.

But AI labels can also help people recognize misinformation

The risk of misplaced skepticism is one side of the evidence. The benefits of labeling deserve equal consideration.

In 2025, researchers published a separate study in PNAS Nexus examining how different labels affected people's responses to misleading AI-generated images.

Across two preregistered experiments involving 7,579 US participants, researchers tested labels describing how content was produced against labels emphasizing its potential to mislead.

The images came from cases previously identified by journalists or fact-checkers as AI-generated and misleading.

Across the tested conditions, labels reduced participants' belief in the claims associated with those images.

That's useful evidence for transparency. When synthetic media is used deceptively, labeling can make audiences more skeptical of material they might otherwise accept.

But there was another distinction in the results.

Simple labels identifying content as AI-generated generally had limited effects on people's stated willingness to engage with the posts, even when they reduced belief in the misleading claims.

The findings show why the objective of a label matters.

A label intended to disclose how content was produced serves a different purpose from a warning intended to tell audiences that a particular claim is misleading.

Those two purposes sometimes overlap, but they require different evidence.

The study also concerned already misleading images. It doesn't establish what happens when an accurate photograph, legitimate illustration, or carefully fact-checked article receives an AI label.

Taken together, the research supports a more nuanced position than either labeling everything indiscriminately or avoiding labels altogether.

Transparency gives people useful information. The wording and context determine how accurately that information is understood.

Meta's labeling controversy shows what happens when those distinctions disappear

The earlier controversy involving Pete Souza and other photographers demonstrates why this matters outside controlled experiments.

In June 2024, TechCrunch documented complaints from photographers whose real images were receiving Meta's "Made with AI" label.

Meta's system relied partly on industry-standard indicators associated with AI tools, including information embedded in image files.

Some editing applications could attach those indicators even when the underlying photograph hadn't been generated by AI.

The result was a mismatch between a technical signal and the meaning an ordinary viewer might assign to the label.

A photograph created with a camera and then lightly retouched isn't equivalent to an entirely synthetic image presented as documentary evidence.

Meta acknowledged this issue. In July 2024, it began replacing "Made with AI" with "AI info" to better reflect the range of possible AI involvement.

In September, it announced another adjustment, moving certain labels for AI-edited content into the post menu while retaining more visible labels for content identified as AI-generated.

The change recognized that people needed more information about the extent of AI involvement rather than a single blanket classification.

There's a legitimate counterargument here. Even relatively minor AI edits can matter.

A news photographer removing an object from a scene can materially change what the image appears to document. A journalist has an obligation to preserve the integrity of documentary evidence regardless of whether the editing tool uses AI.

The relevant distinction therefore isn't simply minor versus major editing, or human versus machine.

It's whether the alteration changes the facts the content claims to establish.

That is where origin information becomes useful evidence rather than a substitute for editorial judgment.

And it's a distinction worth preserving as more companies adopt AI watermarking systems.

Google's expansion brings this question into a new era of transparency

Google's October SynthID rollout arrives amid growing regulatory expectations surrounding synthetic media.

Under Article 50 of the European Union's AI Act, covered AI providers and deployers face requirements involving machine-readable marking and disclosure of certain AI-generated or manipulated content.

The precise obligations depend on the system, use case, applicable exceptions, and transitional provisions.

Those requirements give companies an additional reason to develop reliable ways of identifying AI-generated content.

Google's SynthID provides one approach by embedding detectable signals within supported media. The Coalition for Content Provenance and Authenticity, or C2PA, offers another through digitally signed records describing aspects of a file's origin and editing history.

Both approaches provide evidence about how content was created or processed.

Neither independently determines whether the events or statements depicted in that content are true.

That difference becomes especially important when verification results inform consequential decisions.

A newsroom investigating a suspicious video needs to know whether its visual or audio elements were generated or manipulated. It must also establish whether the underlying event actually occurred.

A company reviewing marketing materials might need to know whether AI-generated assets were used and whether the required disclosures were made. That is separate from assessing whether the advertisement's factual claims are accurate.

And a platform evaluating a potentially deceptive post needs to distinguish evidence of synthetic creation from evidence that the content violates a specific policy.

These examples illustrate potential workflows, not documented failures of Google's newly released detector.

They also reveal why simply identifying AI involvement cannot settle every question that follows.

The most useful AI verification result may need three separate answers

The research suggests that an effective verification interface should distinguish three kinds of information that are often collapsed into one label.

The first concerns origin: How was this content made?

A detector can report whether it found a supported watermark or provenance signal, ideally identifying the relevant provider or technology where the evidence permits. A negative result must be interpreted within the detector's known coverage and limitations.

The second concerns alteration: What role did AI play?

A meaningful explanation would distinguish complete AI generation from limited editing when the available evidence supports that distinction. It could also identify whether the signal concerns an entire file or a specific portion.

Google's earlier description of SynthID Detector included the ability to highlight regions of images or segments of audio where watermarks were more likely to appear.

That provides a useful example of how detection systems can communicate more detail than a binary result, although it doesn't by itself establish the extent of every edit.

The third concerns accuracy: Is the content's claim true?

That question generally requires corroboration, context, original records, and human judgment. A watermark cannot determine whether a real photograph has been given a false caption or whether a synthetic illustration accompanies accurate reporting.

Separating these categories could help prevent two mistakes: treating every AI-generated asset as deceptive and treating every unmarked asset as authentic.

But there's an important unanswered question.

Would people actually understand these distinctions if Google's interface presented them more explicitly?

The existing experiments suggest that additional context can reduce misunderstandings about AI labels. None directly tests the current SynthID portal.

That leaves a clear research opportunity.

Independent researchers could ask users to evaluate the same files with different explanations: a generic AI label, a description of the detected watermark, and a fuller explanation of what the signal establishes and what remains unknown.

The results could measure whether people correctly distinguish AI origin, the extent of editing, and factual accuracy—and whether they make better decisions about what additional verification is needed.

That would help establish whether more detailed labels genuinely improve understanding or simply add information that users ignore.

The real challenge is making verification useful without making it misleading

Google's decision to open SynthID Detector to the public is a meaningful expansion of access to AI-origin evidence.

As more providers participate, ordinary users and professional investigators gain another way to examine suspicious media.

But the research on AI labeling shows why improved technical detection isn't the end of the problem.

People sometimes interpret AI labels as indications that information is less accurate, even when it is true. The experimental effects are generally modest and context-dependent, and they don't establish a broad collapse in trust.

Other research shows that labels can help people recognize misleading synthetic images, which provides a strong reason to make such information available.

Both findings can be true because the value of a label depends on what it identifies and what readers understand it to mean.

For a newsroom, the practical consequence is straightforward.

If a photograph or video carries evidence of AI involvement, that finding should prompt an investigation of its origin, editing history, and factual claims. The strength of the evidence—not the mere presence of an AI-related label—should determine the editorial conclusion.

A label can justify further scrutiny without justifying an accusation of fabrication.

Google's technology is designed to help identify supported AI-generated media. The next question is how accurately people interpret the evidence its detector provides.

Because knowing that AI helped create something is useful. Knowing what that fact actually proves is what makes verification trustworthy.