An AI assistant can help people make decisions, but a study of fake faces suggests that some users may become less discerning when they believe the advice came from AI. Researchers asked participants to judge whether photographs showed real people or computer-generated faces. The research examined AI advice on fake faces, but the troubling finding was not that such advice made everyone worse. It was that participants with more positive attitudes towards AI became poorer at distinguishing the two when the guidance was presented as coming from an AI system.
The peer-reviewed research, published in Scientific Reports and indexed by PubMed, involved 295 people judging 80 face images. Half the faces were real and half were synthetic. Alongside each image, participants saw a prediction presented as either human or AI guidance. Crucially, the guidance was right only half the time. The design let the researchers explore how attitudes and perceived sources affected a person’s own judgement.
Why AI advice on fake faces can sway judgement
Imagine viewing an unfamiliar portrait online and wondering whether it was made by a camera or an image generator. You have an instinct, but a tool beside the image announces that it believes the face is authentic. You might trust the tool, question your first impression or simply feel more confident that a decision has already been made for you. None of those reactions changes the pixels in the picture. They change the way you weigh the evidence.
That is the subtle problem the researchers examined. The human participants were not merely asked to identify synthetic images in isolation. They received outside advice and had to decide how much of it to use. The prediction appeared to come from an AI system for some people and from a human source for others. Because it was correct just half the time, treating the guidance as a reliable shortcut was not a sound strategy.
The experiment was therefore not a demonstration that a sophisticated detection tool had failed in the wild. Nor did it prove that an actual commercial AI service gave misleading advice. The source label and the advice were part of an experimental setup designed to measure reliance. The practical issue is how people respond when they believe a machine has assessed something they are uncertain about.
The result was about trust, not a universal collapse in accuracy
Participants who saw guidance attributed to AI and already had more favourable attitudes towards AI showed poorer ability to separate genuine from synthetic faces. Among those receiving advice attributed to a human source, the corresponding level of trust in people did not show the same relationship. The researchers describe this as a reason to take the possibility of AI-specific reliance bias seriously.
That result needs careful phrasing. It would be wrong to claim that every person who received AI guidance became less accurate, or that simply labelling advice as AI necessarily lowered the average score. The relationship concerned differences between people in the AI-guidance condition, particularly their existing attitude towards the technology. It is a more interesting finding than a sweeping slogan because it suggests that confidence in an assistant can shape whether somebody scrutinises its output.
The experiment also relied on a particular task and a relatively limited participant group. Distinguishing human from synthetic faces is challenging, and the balance of real and generated pictures was deliberately controlled. Everyday settings are different. A person might be judging a profile photograph, a news image, a video call or a suspicious message alongside other context. The results cannot be assumed to reproduce the same magnitude of effect in every situation.
Even so, the central mechanism is familiar. When an apparently knowledgeable source offers a conclusion, people may stop checking the underlying details as carefully. An assistant that is fluent, friendly and usually helpful could encourage that response. Its confidence may feel like evidence even when no reliable evidence has been supplied. That is precisely why the source of an assertion and the quality of the assertion need to be considered separately.
Fake faces are a useful test case
Synthetic portrait images have become convincing enough that an ordinary viewer can struggle to rely on visual clues alone. Skin texture, lighting and facial proportions are not dependable universal giveaways, particularly after a picture has been compressed or edited. A detection tool might provide useful information, but its answer should be treated according to what the tool was designed and tested to do. A statement delivered with certainty is not the same as a validated probability.
In the experiment, the supposedly helpful prediction was deliberately unreliable. That made it possible to observe who was swayed rather than merely reward people for following good advice. If the guidance had always been correct, stronger reliance could have improved outcomes, at least within the test. The researchers were interested in the relationship between source perceptions and decision-making, not in proving that asking AI for help is inherently irrational.
This distinction also matters when discussing responsibility. People use maps, calculators and medical instruments because those tools can outperform unaided judgement in appropriate circumstances. The goal should not be to distrust every machine. It should be to understand the conditions under which a particular tool is reliable and to avoid giving greater weight to an answer solely because it was produced by AI.
Checking the origin is different from guessing by eye
There are technical ways to investigate synthetic media that do not depend entirely on a person’s visual confidence. The US National Institute of Standards and Technology’s work on synthetic-content transparency examines approaches including provenance information, labelling and other methods for assessing digital material. Such methods have limitations, but they shift the question from whether a face looks suspicious to what evidence exists about its creation or handling.
LiveAIWire has examined hidden labels that can reveal the origin of AI images. Those labels may help in some workflows, but absence of a label is not proof that an image is real. Files can be stripped of information, and not every creation or editing service uses the same practices. Content provenance is an additional clue, not a magical detector that settles every case.
Google’s SynthID detection technology offers another angle on the same problem. Where compatible content includes an identifiable signal, a tool may provide evidence that AI was involved. That is different from a generic assistant looking at a face and expressing an opinion. Understanding what a tool actually checks prevents both unjustified trust and unjustified dismissal.
What a sensible second opinion looks like
If a portrait seems suspicious, one sensible approach is to ask where it came from before relying on its appearance. Was it published by an accountable organisation? Is the original source available? Does other evidence support the identity or event being claimed? When an AI assistant offers a verdict, ask whether the answer is grounded in a specific detection method or is simply a judgement generated from patterns in the image.
It is also worth resisting the temptation to turn a low-confidence guess into an accusation. Declaring a real photograph fake can unfairly harm the person depicted. Declaring a fabricated image genuine can help a scam or misleading claim spread. The appropriate response may be to say the available evidence is insufficient, especially when the picture is being used to support a consequential allegation.
The challenge extends beyond photographs. LiveAIWire has previously explored fake journalists presented as real news sources. In such cases, polished presentation can create a sense of credibility even before anyone verifies the source. The new face experiment highlights an additional risk: a reassuring assessment by an AI system might itself become part of the persuasive presentation.
The hardest part is knowing when to doubt a helpful tool
Most people will not face a carefully balanced test containing forty real and forty synthetic faces. They will encounter images intermittently, while scrolling or replying to messages. Some will be obvious, others ambiguous, and many will not matter enough to investigate. The important skill is recognising when the stakes justify a deliberate check rather than accepting the first plausible answer, whether it came from a person or a program.
Researchers still need to test different tasks, advice quality and participant groups to establish how broadly the observed relationship holds. They also need to understand when an AI label improves decision-making by encouraging useful scrutiny, and when it weakens it through overconfidence. The present study offers an experimental warning, not a prediction about every future AI-assisted judgement.
An assistant can be genuinely useful without being automatically right. The lesson from the fake-face experiment is to judge the evidence behind its advice, particularly when the verdict feels reassuring. The very people most comfortable with AI may need to make an extra effort to check whether the system is providing reliable information or merely a confident-sounding nudge.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.