Can people spot AI-generated images?

The direct answer: not reliably in this controlled 2024 test

Most participants could not reliably tell the AI-generated images from authentic ones. Their probability judgements were slightly better than a simulation that produced random answers between 0 and 100, but they were worse than a deliberately uninformative rule: giving every image a 50% probability of being AI-generated. That second comparison is the practical headline because it shows that confidence in visual clues often failed to add useful information.

The finding does not mean every person was wrong half the time, nor that all synthetic images are undetectable. The researchers used Brier scores, which reward both correct discrimination and well-calibrated confidence. A confident wrong answer is penalised more than a cautious one. The study therefore tests whether people's probability judgements tracked the true image source, not only whether a forced real-or-fake label happened to be correct.

It also does not measure whether anyone believed the depicted event, shared the image or changed a political view. Every authentic and synthetic image referred to a real event, and participants were told that some images had been generated. The evidence supports caution about unaided visual inspection; it does not quantify misinformation exposure or prove a particular policy intervention works.[1]

How 2,091 adults judged a 52-image pool

The polling company Epinion ran the survey from 17 to 24 June 2024 among Danish adults aged 18 and over. The final sample contained 2,091 respondents, with an 8.4% dropout rate, and was constructed to represent the adult population on age, gender and region. The experiment was preregistered and approved by the University of Copenhagen's institutional review board.

Each respondent saw seven images, one at a time, randomly drawn from a pool of 52. The pool covered 26 events; each event had one authentic image and one AI-generated version. A person could not see the same event twice. The images spanned four categories: everyday scenes, people, crises and political subjects. After each image, respondents rated from 0 to 100 how likely it was to have been generated with AI.

The authentic images came from Wikimedia Commons. The synthetic images were made with OpenArt.ai, taking roughly five to 15 minutes per image, in an attempt to represent the resources of a small online actor rather than a state-backed production team. Participants received a short description of the event but no account identity, source reputation or engagement metrics. The median viewing time was seven seconds, reflecting rapid social-media browsing more closely than forensic examination.

That design isolates visual discernment and avoids one major confound: recognising a famous false image from memory. But it also creates an artificial task. People were warned that some images were synthetic and all examples depicted true events. On a real feed, users must also judge whether the event itself is real, whether a caption is misleading, and whether the account deserves trust.[1]

What the scores show—and what 'chance' means here

The researchers calculated a Brier score for each participant from the seven probability estimates. A score of zero would be perfect; lower is better. Participants' average was 0.05 lower than the distribution produced by random probability estimates, a statistically significant difference with a 95% confidence interval from -0.057 to -0.042. In that limited sense, the group did extract some source information from the images.

Yet a constant 50% answer produces a Brier score of 0.25. Participants' mean score was 0.034 higher, and therefore worse, than that benchmark; the 95% confidence interval was 0.028 to 0.039. The distinction matters. Randomly moving a probability slider is a weak comparator. A constant 50% response represents honest uncertainty and, on average, outperformed attempts to infer authenticity from the available visual cues.

No image category was statistically distinguishable from the others in discernment performance. Images of people were not reliably easier than ordinary scenes. Participants who scored highly on a three-question cognitive-reflection test did better than those who scored poorly, with an average Brier-score improvement of 0.04. Their average of 0.26 still did not beat the 0.25 constant-uncertainty benchmark. The only subgroup reported below 0.25 was people aged 18 to 24.

Women and respondents who reported experience with generative AI also had better Brier scores in the main models, while older respondents did worse. These are subgroup associations inside one experiment, not universal demographic traits. The paper notes that the gender difference was no longer statistically significant when using raw probability estimates rather than Brier scores, suggesting that calibration, not necessarily discrimination, contributed to the result.[1]

Why the result should change verification practice

For readers, the useful lesson is not to become suspicious of every polished or imperfect image. It is to treat visual intuition as one weak signal. Reverse-image search, provenance information, corroboration from independent sources and examination of the original account can provide evidence that a seven-second glance cannot. A visible oddity may prompt a check, but the absence of an oddity is not proof of authenticity.

For publishers and platforms, the result argues against making users the final detection layer. Content credentials, accessible provenance, clear labelling, rapid correction routes and professional verification can reduce reliance on individual eyesight. None is a complete solution: metadata can be removed, labels can be misread and automated detectors can fail after generators change. Effective systems need multiple signals and a way to express uncertainty rather than a single real-or-fake badge.

There is a symmetric risk. Teaching people that synthetic media are common without supplying reliable verification tools can make them dismiss genuine evidence—the so-called liar's dividend. A useful intervention should therefore improve discrimination and calibration while measuring false suspicion of authentic images. Raising doubt alone is not success.

The study's policy discussion points to watermarks, transparency standards, platform monitoring and stronger independent media and fact-checking. Those proposals were not evaluated in the experiment. They remain hypotheses whose effects should be tested against evasion, accessibility, user comprehension and the risk that bad actors simply move to unlabelled tools.[1]

Important limits: one country, one older generator and a primed task

The data are Danish and were collected more than two years before publication. Denmark has high internet use and digital literacy, but its media system, institutional trust and population do not represent the world. The authors argue that these characteristics may make Denmark a demanding test for the pessimistic claim, yet that inference still needs direct replication across languages, cultures, ages and levels of connectivity.

All synthetic images came from one freely available 2024 tool and were manually selected after multiple prompts. Current models, editing workflows and provenance systems differ. Better generation could make detection harder; new visible artefacts or labels could make it easier. A result about one stimulus set should not be converted into a timeless percentage for all AI imagery.

The experiment also told participants they would see authentic and AI-generated material. That may encourage unusually close attention to pixels while removing the source cues people normally use. Conversely, participants spent little time and saw relatively low-quality images, which may resemble casual feeds. These design choices pull external validity in different directions, so the study is most persuasive as evidence of a problem, not a precise forecast of real-world performance.

Data collection was conducted as part of a grant from Denmark's Agency for Culture and Palaces and Agency for Digital Government. The authors separately reported no financial support for the research, authorship or publication of the article and declared no potential conflicts of interest. Data and code are available through a linked Dataverse repository, supporting independent reanalysis.[1]

What evidence would change the assessment

The next decisive studies would repeat the preregistered design with current generators, multiple tools and images produced at several effort levels. They should include countries with different languages, media environments and levels of digital access; report accuracy and calibration separately; and predefine how subgroup comparisons will be handled. Reusing some common stimuli across countries would help distinguish cultural effects from image effects.

Realistic feed experiments should then add account identity, captions, engagement signals and competing demands on attention. They should measure not only source classification but also belief, sharing, reporting and memory. A promising intervention would improve correct identification of synthetic images without increasing rejection of authentic ones, and its benefit would persist when labels are incomplete or adversaries adapt.

Until those tests exist, the careful conclusion is narrow but consequential: in this large Danish experiment, unaided visual judgement was not a dependable authentication system. People, newsrooms and platforms need evidence that travels with or can be independently connected to an image, plus verification processes that remain useful when certainty is impossible.[1]