Sometimes, but not reliably. When studies ask people to sort AI images vs real images, accuracy ranges from about chance to about three in four. GenImage (2023) quotes a figure of 61.3%. Microsoft’s Real or Not Quiz found 62% across roughly 287,000 judgements. Kamali and colleagues’ CHI 2025 study, whose main phase had 37,568 participants, found 76% for AI images and 74% for real photos. Automated detectors look much better on older benchmarks and much worse on current generators. The DailyBench authors (2026) found that detectors scoring 91-96% balanced accuracy on GenImage fell to 52-79% on modern fully synthetic images and 43-67% on locally edited ones.
So the useful question is not whether you, or a tool, can tell the difference. It is which checks still carry information about a given image, and how much. Visual tells, classifiers, provenance records and context searches each answer a different part of that question, and each has a known way of failing.
How accurate people are in the studies, and why the numbers differ
These figures form a range, not a set of rival headlines, because the studies measured different things. HPBench (NeurIPS 2023) reported that people misclassified 38.7% of images, which is 61.3% accuracy. GenImage quotes exactly that figure. In the Microsoft quiz, more than 12,500 participants worldwide judged a randomised set of real and AI images, and the authors describe the 62% result as only slightly above chance. Cooke et al. (2024) tested 1,276 participants across images, audio, video and audiovisual clips and found mean performance close to the 50% chance level. Frank et al. (IEEE S&P 2024) surveyed 3,002 people in the USA, Germany and China and found that most of them were simply guessing on state-of-the-art forgeries.
The Kamali study was designed to show which factors affect how accurately people judge. Its 599 main-phase images were 149 real photographs and 450 diffusion images made with Midjourney, Firefly and Stable Diffusion, each chosen through repeated attempts to look as photorealistic as possible. Accuracy on individual AI images ranged from 32% to 99%. Viewing time mattered. At one second, people were 72% accurate on AI images, rising to 80% at ten seconds and 82% at twenty. With unlimited time, only 17% of the generated images were identified below chance, but that share rose to 43% when viewing was limited to one second.
Kamali’s results tie accuracy to how carefully the fake images were chosen, how long people looked, and what the images showed. Kamali’s hand-picked images were judged more photorealistic than 84% of the matching uncurated images generated from the same prompts. People also bring biases. Participants in that study started out labelling 63% of images as fake, and the bias faded after a few images. In Frank et al., generalised trust, cognitive reflection and self-reported familiarity with deepfakes all influenced people’s answers.
The studies disagree about subject matter. In the Microsoft quiz, people did best on human portraits and worst on natural and urban landscapes. In Kamali’s study, portraits were the hardest AI category at 72.7%, against 77.2% for full-body shots. The authors explain that portraits often show one person against a blurred background, with simple poses that leave fewer places for errors to appear. Whether a portrait counts as “easy” depends on what it is compared with, so we would not treat either ranking as a general rule.
What real-or-AI photo quizzes measure, and what your score means
A real-or-AI quiz measures how well you sort one particular set of images, under that quiz’s rules, with nothing to go on but the pixels. The rules change the score. The set may have been curated for realism, the viewing time may be fixed or open, and the subject mix may lean towards portraits or landscapes.
Scores are also not as random as the headline averages suggest. In Kamali’s study, 34% of participants got at least 90% right on their first ten images. Pure guessing would put only about 1% of people at that level.
The New York Times quiz that Kamali et al. cite points out that repeated cropping and compression on social media can make a real photo look AI-generated.
The visual tells that exist, and how often they are missing
Researchers group the visible signs of a diffusion-generated image into five categories:
- Anatomical implausibilities: extra or missing fingers, overlong necks, empty or overly shiny eyes, teeth merging into the mouth, merged limbs in group shots.
- Stylistic artifacts: waxy or glossy skin, overly soft hair, a cinematic look, and parts of the image that seem cut from different scenes, with smudged edges or mismatched resolution.
- Functional implausibilities: objects that could not work, such as guitar strings hanging loose or a sandwich held sideways, straps merging into clothing, and garbled text and glyphs.
- Violations of physics: shadows, reflections, lighting and perspective that do not agree.
- Sociocultural implausibilities: scenes that break social norms, cultural context or historical accuracy.
Forensic research adds finer cues, such as mismatched reflections in the corneas and irregular pupil shapes.
How often these cues show up is a different matter from what people look for. In Kamali’s annotations, functional implausibilities appeared in 58.7% of AI images, anatomical ones in 51.4% and stylistic ones in 39.0%. Physics violations appeared in only 9.17% and sociocultural ones in 5.50%. In their comments, 61% of participants said they relied on anatomy. Yet accuracy was lower on images with functional implausibilities (64.1%) than on those with anatomical or stylistic ones. Having several artifacts does not make an image obvious either. Whether a flaw is visible also matters, not only whether it is present.
The most important caveat comes from the authors of the guide themselves: human-perceptible artifacts are not always present in AI-generated images. Without context beyond the pixels, it is often hard to draw any firm conclusion about where an image came from. A tell you find is grounds for suspicion. A tell you fail to find does not show the image is real.
Why AI-generated photos keep getting harder to spot
Generators have been removing the traces that both people and detectors relied on. GenImage’s frequency analysis found that GAN images carry a regular grid-like artifact, while diffusion images, like real ImageNet photos, carry few. Early VAEs left periodic artifacts because of their fixed-stride convolutions, and those easy-to-spot patterns have faded as architectures improved. The DailyBench authors describe the same shift in visible errors: earlier generators such as SDv1.4 often produced inconsistencies across the whole image, while recent ones produce coherent images whose synthetic traces are subtler and vary from region to region. For faces and avatars, 3D-aware rendering reduces the geometric errors, odd angles and lighting mismatches that detectors often use.
Detectors also go out of date with each new generator. The c2paviewer.com article on file provenance describes each new generation of diffusion models as producing output that defeats classifiers trained on the previous generation. AI-GenBench (2025) tests this directly by training on older generators and testing on newer ones in release order. Its results show a significant drop that coincides with the detectors’ first evaluation on Stable Diffusion generators. Human effort compounds the problem: a person who regenerates and picks the most realistic output makes the image harder to judge than a casual generation would be.
Do AI image detectors and reverse image search beat human judgement?
On still-image benchmarks, the best detectors beat people. Outside them, detector accuracy drops sharply. In HPBench, the best model failed on 13% of images under the same setting in which humans failed on 38.7%. A 2025 review reports a GPT-4o fusion detector at 93.4% and CNNSpot at 91.8%, against 86.3% for the best human annotator. For deepfake video, Groh et al. (PNAS 2022) tested 15,016 participants and found ordinary observers about as accurate as the leading detection model, though the two made different kinds of mistakes.
Where detector accuracy collapses
One major weakness is generalisation. In GenImage, detectors trained and tested on the same generator scored above 98.5%, reaching 99.9% on Stable Diffusion V1.4 and V1.5. A ResNet-50 trained on Stable Diffusion V1.4 and tested on Midjourney scored 54.9%. Averaged over all 64 train-test combinations, accuracy was 66.9%. Reproducibility work points to the cause: most methods overfit to features specific to one generator instead of learning traits common to all AI images. On DailyBench, accuracy dropped from 91-96% to 52-79% for full synthesis and 43-67% for local edits. One detector there went from 98.50% on GenImage to about 57% on both newer sets.
Detectors also lean towards calling images real. DailyBench’s authors found that existing detectors are much less accurate on fake images than on real ones. Locally edited images are often classified as real, which the authors suggest is because edits preserve most natural content. On manipulation detection, most methods stayed around 50%, and the best reached 56.62%.
Benchmarks of material collected from the internet show the same gap. Deepfake-Eval-2024 gathered in-the-wild deepfakes from social media and detection-platform users in 2024, including 1,975 images. On it, the AUC of open-source detectors fell by 45% for image models, 48% for audio and 50% for video compared with earlier benchmarks.
Compression, resizing, blur and deliberate attacks
Everyday handling of an image weakens the cues detectors depend on. GenImage found that JPEG compression has a significant impact on standard backbone detectors, unless they were trained with compression and blur as augmentations, as CNNSpot was. DailyBench found accuracy generally falling as JPEG quality falls, and most detectors degrading as blur increases, because they rely on fine-grained artifacts that smoothing removes. In an adversarial robustness study from 2024, simulated social media degradation alone (medium level: downscaling to 75% and JPEG quality 50) cut average AUROC from 87.5% to 66.6%. Corvi et al., as summarised by Kamali et al., found that recropping and compression push some classifiers close to random guessing.
Deliberate attacks are worse. The same 2024 study found that a black-box Diverse Input Attack reduced average AUROC from 87.5% to about 0.0%, without any access to the detector’s internals. Against the commercial tool HIVE, average AUROC fell from 98.9% to 76.6% under attack, and the HIVE team confirmed the results after disclosure. Robust pretrained backbones recover some of that loss, but at the cost of some accuracy on clean images.
General-purpose vision models behave like careful humans
Large vision-language models behave more like human observers than specialised detectors do. On the Mirage dataset of 5,000 AI images with subtle but visible artifacts, a zero-shot LVLM reached 94.62% accuracy, ahead of specialised detectors such as CODE at 88.65%. On the Chameleon set, which was built to look artifact-free, the same model fell to 62.02% and caught only 15.52% of the fakes, while CODE reached 71.33%. The Mirage authors frame this as a trade-off. LVLMs spot visible flaws much as a person would, while contrastively trained detectors are better at the invisible statistical patterns. Asking the LVLM to explain its reasoning step by step lowered its accuracy to 82.03%. More broadly, the 2026 Explainable Deepfake Detection Challenge found that detection and explanation are separate capabilities, and that strong classification does not guarantee explanations grounded in the right visual evidence.
What reverse image search adds
Google gives the example of its About this image feature showing that news articles had already identified a staged moon-landing picture as AI-generated. The c2paviewer.com article on file provenance groups reverse image search with classifiers as detection-based approaches that lose accuracy as generators improve.
For combining human and machine judgement, the PNAS result is the one to remember. People who saw the model’s prediction were more accurate than either people or the model alone, but wrong predictions often made them less accurate. A detector score shown to a moderator is evidence the moderator will lean on, so a confident wrong score does real damage.
Provenance: Content Credentials and watermarks prove origin, not truth
Provenance turns the problem around. Instead of hunting for signs of fakery, it records where a file came from and what was done to it. The 2026 deepfakes survey calls the C2PA family of specifications important because they provide a technical way to document the origin and transformation history of digital content, including generative edits. The C2PA explainer (v2.4) is explicit about the limit: provenance information helps establish origin, history and authenticity, but it cannot tell you whether the content is true, accurate or factual. A camera can sign a manifest for a staged scene. The signature proves which camera took the photo and when, not that the scene is real.
A missing signature is not proof of fakery, though a valid signature is strong evidence of authenticity. Records can also have gaps. Cropping in a tool that does not support Content Credentials may leave the history out of date. Metadata can be stripped, which is why C2PA defines durable Content Credentials that pair cryptographic hashing with watermarks or fingerprints, so the record can be found again in cloud storage. The 2026 deepfakes survey notes that the value depends on adoption, on metadata surviving, on key management and on platform support.
Google’s SynthID Detector checks whether an image made with Google AI carries a SynthID watermark. Google says the watermark stays detectable after sharing and a range of transformations. That is Google’s claim, and the deepfakes survey notes that, in general, watermarks can be weakened, removed or attacked through fine-tuning and other techniques.
Do the same tells work on AI faces, deepfakes and AI video?
Only partly. Faces are hard for people. Cooke et al. found people less accurate on synthetic images that feature human faces. Earlier research found GAN-generated face portraits indistinguishable from real ones. Two other studies cited by Kamali et al. found that people disproportionately judged white AI-generated faces to be human (Miller et al., 2023) and rated GAN faces as more trustworthy than real faces (Nightingale and Farid, 2022).
Video adds evidence that still images lack, which is time. Reenactment deepfakes can look plausible frame by frame yet show temporal or geometric artifacts across frames. Video-level detectors use exactly that, while frame-level detectors are cheaper but can be beaten by picking good frames. None of this makes video easy to detect. In Meta’s DFDC competition (2020), 60% of submissions scored a log loss on the private test set no better than predicting 0.5 for every video. The best models reached an average precision of 0.753 and a ROC-AUC of 0.734 on real-world deepfake videos. The authors still described deepfake detection as extremely difficult and unsolved. The 2026 deepfakes survey reports detector AUC of around 99% on controlled FaceForensics++ and Celeb-DF settings, but around 80-81% on the in-the-wild WildDeepfake set.
Newsroom practice, as the 2026 survey describes it, therefore checks whether audio, lip sync, lighting and timing agree, checks the source and how the clip spread, and cross-references independent context, instead of relying on a detector score.
AI images vs real images: what each check can tell you
| Check | What it can establish | Where it fails |
|---|---|---|
| Looking for visual tells | Strong suspicion when anatomy, function, text or physics clearly break | Many AI images have no visible artifact; compression can make real photos look fake |
| Taking more viewing time | Accuracy on AI images rose from 72% at 1 second to 82% at 20 seconds in one large study | Even with unlimited time, 17% of AI images were identified below chance |
| Specialised AI image detector | Very high accuracy on generators it was trained on | 52-79% on modern synthesis and 43-67% on local edits in DailyBench; weakened by JPEG and blur; vulnerable to adversarial attack |
| Vision-language model | 94.62% on images with subtle visible artifacts | 62.02% on artifact-free fakes, catching 15.52% of them |
| Reverse image search and context | Earlier published debunks of the same image | Detection-based, loses accuracy as generators improve |
| C2PA Content Credentials | Signed origin and edit history | Proves origin, not truth; absence proves nothing |
| Generator watermark (e.g. SynthID) | Whether content came from that vendor’s tools | Silent on other generators; watermarks can be attacked |
NIST AI 100-4, as summarised in the 2026 survey, reaches the same conclusion as this table: no single technique is sufficient on its own.
What we would check before trusting a detector in a moderation pipeline
For a team putting an AI-image classifier into a content moderation or trust-and-safety workflow, the published numbers suggest a few tests that matter more than the headline accuracy.
- Test on generators released after the training data. Use a time-based split like AI-GenBench’s “next period” metric, not a random split, because the next period is what the detector will face in production.
- Push the test set through your own platform’s transforms first. Resize, recompress and crop it the way your platform does, since degradation alone moved one set of detectors from 87.5% to 66.6% AUROC.
- Report accuracy on fakes and on real images separately. Detectors tend to prefer the real class, which hides missed fakes inside a good overall number.
- Include locally edited images. Edits are where the DailyBench detectors scored lowest.
- Measure the cost of false positives. Overly aggressive classifiers flag authentic photographs and damage the reputations of real creators and journalists.
- Decide how the score is shown to reviewers. Wrong model predictions pull human accuracy down.
A detector’s score on a benchmark does not guarantee its reliability in your pipeline. For LLM deployment decisions, see our guide to designing task-specific evaluation harnesses for LLM deployment decisions. For synthetic content feeding back into what people and models consume, see our piece on AI-generated data and internet quality, including model collapse. For diffusion models themselves, see explainable AI in generative diffusion models.
FAQ
How can I find out which AI made an image?
Content Credentials help where present, because a C2PA manifest documents the origin and editing history of an image. For Google’s tools, the SynthID Detector checks for Google’s own watermark. Fingerprinting methods in research try to identify traces specific to a generator, and detectors trained on one generator tend to struggle on others.
Why do AI images get details like hands and text wrong?
Kamali et al. trace functional errors such as garbled text and impossible objects to a lack of understanding of real-world mechanics. As models fix these weaknesses, new and subtler artifacts may appear in their place.
