The best AI image detectors are very accurate on clean, full-resolution images from generators they were trained on, and noticeably less accurate on everything else: compressed reposts, screenshots, lightly edited photos, and images from generators released after the detector was built. No detector is 100% accurate, and a vendor’s headline figure describes its own test set, not the image in front of you. Treat a result as a strong signal to combine with other checks, never as proof.
For comparison, people are worse. In a 2022 PNAS study, participants identified AI-generated faces 48.2% of the time, no better than chance, and training with feedback lifted that only to 59% (Nightingale and Farid).
Why a single accuracy number misleads #
Detector websites often quote one number, such as “98.9% accuracy.” That figure can be true and still tell you little about your own use, for four reasons.
1. It depends on the test set #
Accuracy is measured on a particular collection of real and AI images. If that collection is full-size images straight from popular generators, the number will be high. If it includes screenshots, compressed reposts and new generators, it will be lower. Vendors rarely publish their test sets in enough detail for you to judge.
2. New generators break old detectors #
Detectors learn the fingerprints of the generators in their training data. Researchers have shown that standard real-vs-fake classifiers are “asymmetrically tuned” to spot fake patterns they’ve seen, so images from an unseen generator tend to fall into the “real” bucket by default (Ojha et al., 2023). Their method, built on features from a large pretrained vision model, improved generalization to unseen models substantially, but the gap never closes completely. An early paper on the subject was titled, fittingly, “CNN-generated images are surprisingly easy to spot… for now” (Wang et al., 2020).
3. Real-world images are messier #
Every upload, screenshot, crop and filter changes the fine texture detectors read. An image that has been through three platforms is harder to judge than the file that left the generator.
4. Base rates change what a flag means #
This one surprises people. Suppose, purely as an illustration, that you check 1,000 images, 50 of which are AI-generated, with a detector that catches 95% of AI images and wrongly flags 5% of real ones:
| Flagged as AI | Not flagged | |
|---|---|---|
| 50 AI images | 47.5 | 2.5 |
| 950 real images | 47.5 | 902.5 |
Half of the “AI” flags would be real photos. The detector is “95% accurate” in both directions, but when most of what you check is real, false positives pile up. That’s why a flag on a photo of a real person should prompt more checking, not an accusation.
What makes detectors less accurate #
| Situation | Effect on accuracy | What to do |
|---|---|---|
| Heavy compression, many reposts | Fine texture lost, weaker signal | Find the earliest, largest copy |
| Screenshots with app interface | Clutter and resampling | Crop to the picture |
| Small images and avatars | Too little detail | Use the full-size version |
| Brand-new generators | Unfamiliar fingerprints | Weigh source and context more heavily |
| Real photos with AI edits | Mixed signals | Crop around the suspect area and check again |
| Beauty filters, heavy phone processing | False positives | Ask for an unfiltered original |
| Deliberate attacks | Detectors can be fooled | Don’t rely on one tool |
That last row is well documented. Researchers have shown that small, deliberate changes to an image can push detectors toward the wrong answer, and that some watermarks can be removed or even forged (Saberi et al., 2023). See can AI image detectors be fooled for how that works.
What independent testing has found #
Independent comparisons are rarer than vendor claims, but they exist. A 2024 study from the University of Chicago tested automated detectors against crowdworkers, thousands of professional artists and a small group of expert artists on telling human art from AI art (Ha et al.). Hive’s detector and the expert artists both did very well but failed differently: Hive was weaker against adversarial perturbations, while the experts produced more false positives. The authors concluded that a human plus an automated detector gave the best mix of accuracy and resistance to manipulation.
That’s a good rule for everyday use, too. Pair the detector with your own look at the source, the context and the details.
How to read a detector result sensibly #
- Look at the confidence, not just the label. A verdict at 98% AI is not the same as one at 62%.
- Take “Uncertain” at face value. It means the image didn’t carry enough signal. It’s not a hidden “probably AI.”
- Know which evidence decided. A signed Content Credential that says “AI-generated” is much stronger than a model score.
- Try a better copy. A second check on the original file often gives a clearer answer than the first check on a screenshot.
Expose AI is designed around these points. Every verdict (Likely real, Uncertain or Likely AI-generated) comes with a confidence meter, and the “How we decided” card says whether the metadata or the on-device model made the call. When the signal is weak it says Uncertain rather than guessing. The Pro model, a larger on-device model, returns roughly half as many Uncertain verdicts as the standard model on the developer’s evaluation set. The app itself describes results as a strong signal, not proof, and asks users to treat it as one tool among several. We explain the result screen in AI image checker results explained.
Frequently asked questions #
Which AI image detector is the most accurate? #
There’s no single answer, because accuracy depends on the kind of images tested. Independent research has found Hive’s detector among the strongest on AI art, and vendors like Sightengine cite academic comparisons in their favor. Run important images through more than one tool and compare. Our roundup of the best AI image detectors covers the options.
Can AI detectors give false positives? #
Yes. Filtered selfies, heavily processed phone photos, professional product shots and digital renders can all be flagged as AI. When most of the images you check are real, even a small false positive rate produces a lot of wrong flags. Asking for the original, unfiltered file usually settles it.
Are AI image detectors reliable enough for evidence? #
Not on their own. A detector result is probabilistic and can be wrong in either direction, so it shouldn’t be the sole basis for accusing someone, rejecting a claim or deciding a dispute. Original files, Content Credentials and corroborating sources carry more weight.
Do detectors get better over time? #
They get retrained, and good ones improve on the generators they’ve seen. But new generators keep arriving, so detection is an ongoing race rather than a solved problem. A detector that hasn’t been updated in a year will likely miss images from newer models.