Do AI Image Detectors Work? We Tested Them
In short: Sometimes. In Debunk.org's test of four free detectors, two caught all 10 AI images, one caught only 4, and two called a real photo AI-made. Independent studies of real-world images found the best commercial detector right about 8 times in 10. Treat a detector's score as one clue, never as a verdict.
Key facts
On 1,975 images that circulated online in 2024, the best commercial detector reached an accuracy of 0.82, and "No commercial models that we evaluated had an accuracy of 90% or above." (Deepfake-Eval-2024)
In Bellingcat's 2023 test, the detector AI or Not caught all 100 original Midjourney images, but after compression it "falsely identified seven of ten images as real".
People are not much better: in a 2025 Microsoft study of about 287,000 judgements by more than 12,500 people, the success rate was 62%, "slightly above chance".
In Debunk.org's test, Hive Detect and DeepAI caught all 10 AI images, Illuminarty caught 4, and Hive and Illuminarty each called one real photo AI-made.
At least one detector maker advertises "99%+" accuracy on its own product page, without a linked study.
Do AI image detectors work?
They can help as one clue, but independent tests show they miss many real-world fakes and sometimes flag real photos asAI-made.
An AI image detector is a tool that looks at a picture and gives a score for how likely it is to be AI-generated. Most free ones work in a browser: you upload a picture and get a label or a percentage.
The catch is where accuracy figures come from. Makers usually report results on their own test images. Researchers who collected images that really circulated online in 2024 found much weaker results: open-source image detectors lost 45% of their performance compared with older academic test sets. Commercial tools did better, but none reached 90%, and the researchers concluded that they "still need improvement to reach the accuracy of human deepfake forensic analysts".
The free AI image detectors we tested: results
Hive Detect and DeepAI caught every AI image in our set, Illuminarty missed most of them, and Hive and Illuminarty each called one real photo AI-made.
Detector | AI images caught (original) | AI images caught (compressed) | Real photos wrongly flagged as AI | Account needed? |
Hive Detect | 10/10 | 10/10 | 1/10 | No |
DeepAI | 10/10 | 10/10 | 0/10 | No |
Illuminarty | 4/10 | 4/10 | 1/4 (6 larger photos were not analysed) | No |
Sightengine | 7/7 (stopped by its free limit) | not tested | not tested | No, until the free limit; then it asks you to sign up |
AI or Not | not tested | not tested | not tested | — |
Source: Debunk.org test. Twenty images and ten compressed copies, tested on one day: a guide, not a benchmark. Detectors are updated often, so results can change.
How we tested. Debunk.org tested free detectors that work in a browser, without payment and without an account. We chose tools from Google's results and from the studies on this page. Each got the same files, in the same order:
10 AI-generated images whose creators had published them on Wikimedia Commons, labelled with the generator they used: Midjourney, Flux, Gemini, ChatGPT, Adobe Firefly, Grok, Stable Diffusion, Imagen and Microsoft's Bing Image Creator. One is a 2023 Midjourney image of Donald Trump being "arrested", made by Eliot Higgins;
10 real photos from Wikimedia Commons, with a named photographer, camera data and a date between 2006 and 2014, before AI image tools were common: landscapes, city views, food, a train and a football match;
10 compressed copies of the AI images, resized to 1,080 pixels and saved as lower-quality JPEGs, because that is how pictures usually reach people on social media.
We counted an AI image as "caught" when the tool labelled it AI or scored it 50% or more; a real photo treated the same way counted as a "false alarm". We saved every score with the time it was shown.
What we couldn't test. Sightengine stopped after seven files with the message "Free limit reached". We didn't sign up to continue. Illuminarty didn't analyse six of our real photos, all files of 3.5 MB or more, and gave no error message. AI or Not's site blocked us from loading test files, so they aren't in the results.
Limits. Our images are clean files, mostly straight from the generator; fakes that spread online are often edited or screenshotted. All 20 have also been public for months or years, so detectors may have seen them in training.
Which AI image checker worked best?
In our test, DeepAI labelled all 30 files correctly, and Hive Detect caught every AI image but called one real photo AI-made.
That makes them the best performers on our set, not proof that they are reliable. Our files were clean and few, and the studies above found much weaker results on images that really circulated online. Illuminarty's scores varied widely: 87.7% for the Midjourney image, but 3% for the Flux image.
Compression made little difference in our test: no tool's count changed between originals and compressed copies, although Hive's score for the Firefly portrait fell from 97.9% to 87.3%. Bellingcat saw a much bigger effect in 2023, so don't assume this holds for every tool or image. Whichever tool you use, check the picture's source as well.
What the makers say, and what independent tests found. Detector makers publish their own accuracy figures. Here, we quote them as the makers' own claims, alongside independent evidence.
Tool | What the maker says on its own page | Independent evidence |
TruthScan | "99%+ AI Image Detection Accuracy" (TruthScan) | We found no study linked on that page. We didn't test TruthScan's own site, but DeepAI's detector page says it is "Powered by TruthScan", and it labelled all 30 of our files correctly. |
Sightengine | "In an independent benchmark by researchers at the University of Rochester and the University of Kansas using 80,000 images across multiple generators, our detector achieved the highest accuracy among the tools tested." (Sightengine) | The study, published in April 2024, found Sightengine the most accurate of five commercial detectors in seven of its nine tests, including 98.72% on real images. On two sets of AI images created from real photos, it scored 25.12% and 43.76%. In our test it caught all 7 AI images it checked before its free limit. |
AI or Not | "Detect AI-generated images, text, videos, and deepfakes via API with 98.9% accuracy." (AI or Not) | Bellingcat's 2023 test is described above. We couldn't test it. |
The same study found that "most of the detectors also have a strong tendency to label most of the images as human-generated". Sightengine's own FAQ also says that "no detector is 100% reliable".
Can an AI photo detector be wrong about a real photo?
Yes. In Bellingcat's test and in ours, detectors called genuine photos AI-made.
In Bellingcat's 2023 test, AI or Not marked six of 20 real photos by photography competition entrants as AI-generated. In our test, Hive Detect gave a 2012 photo of match officials at an international football match a 78.2% AI score. Illuminarty gave a 2006 photo of prayer flags on a temple roof in Kathmandu a 51.9% AI score. Both are ordinary photos with camera data from the photographers.
A false alarm can do real harm: it can make a genuine news photo look fake and help people dismiss real evidence. If a detector says "AI" about a photo from a credible source, check the source before you trust the score.
Why AI image detectors fail
Detectors learn the traces left by the generators they were trained on, so compression, screenshots, added text and new generators all weaken them.
Compression and re-sharing. Social networks shrink and re-compress images. In Bellingcat's test, AI or Not identified ten Midjourney images correctly when they were uncompressed, but called seven of the ten real once they were compressed.
Text on the image. Memes and posters often add captions. Deepfake-Eval-2024 saw the accuracy of its fine-tuned models fall by 9% on images with text overlays.
New generators. A detector trained before a new image model was released may not recognise its output. Deepfake-Eval-2024 says many older academic test sets "use outdated manipulation techniques".
AI edits of real photos. In the Rochester and Kansas study, most commercial detectors did worse on AI images made from a real photo than on images made from a text prompt alone.
This is also why detectors can be fooled on purpose, and why a detector should never be your only check.
How to check an image without a detector
Start with the source, not the pixels: find where the image first appeared, look for provenance labels, and only then look for visual errors.
Run a reverse image search in at least two engines, such as Google Lens and Bing, and look for the oldest credible copy. Our free OSINT tools page lists the ones we have tested.
Look for a label or provenance data. Some images carry Content Credentials, which record which tool made a file and when. OpenAI says images made with ChatGPT also carry an invisible SynthID watermark, which "may survive some transformations that remove metadata". Our deepfake checklist explains how to check.
Look for an existing fact-check of the image: search for its key words with "fact check".
Then look for visual errors: hands, text, reflections, shadows, and details that don't match. Use the 7-point checklist.
Use a detector last, as one more clue. If two detectors disagree, treat the result as "unknown".
People are not much better than detectors on their own. Microsoft's 2025 study found a 62% success rate, and participants struggled most with natural and urban landscapes. For a wider method, see our guide to OSINT investigation techniques.
Is it safe to upload photos to an AI image checker?
It depends on each tool's terms, and some keep broad rights over what you upload, so never upload private photos, pictures of children or anything confidential.
This is what the terms or privacy policies of the detectors in our test say:
Hive Detect: its terms, now published under the company name Deploy, give the company an "irrevocable, perpetual, nonexclusive, royalty-free and fully paid, worldwide license" to use what users submit to its demo services.
Sightengine: its terms give it a licence "to store, copy, transmit, and display Subscriber Data" in order to provide the service. They mention training its models only on media sent through its "Feedback API or similar reporting mechanisms". We found no retention period in the terms.
DeepAI: its privacy policy says it uses information "to train our artificial intelligence models", with an opt-out, and that "We also sell Identifiers and User Content to Other Third Parties".
Illuminarty: its privacy policy says that uploaded images are "kept only for the duration necessary to provide our service to you".
Terms change, so read them again before you upload anything. Reverse image search engines differ too: TinEye says it keeps your image "only … briefly in order to show you your results".
FAQ
Are there any accurate AI image detectors?
Some detectors are accurate on some kinds of images, but none is reliable on its own. In Deepfake-Eval-2024, no commercial detector reached 90% accuracy on real-world images. In our small test, DeepAI labelled all 30 files correctly, but Illuminarty caught only 4 of 10 AI images.
How accurate are AI image detectors?
It depends on the image. The best commercial detector in Deepfake-Eval-2024 reached 0.82 accuracy, but results drop on compressed copies, screenshots, and images with added text (Bellingcat, 2023). In our test, the three free detectors that checked all our files caught between 4 and 10 of 10 AI images.
Can ChatGPT identify a picture?
ChatGPT can describe a picture, but that is not the same as proving where it came from. For images made with OpenAI's own tools, OpenAI offers a checker at openai.com/verify. It looks for OpenAI's own signals, such as a SynthID watermark or Content Credentials, and if it finds one, the content "was likely generated by an OpenAI model". Because it looks only for OpenAI's signals, it says nothing about images made with other companies' tools. Debunk.org has not tested it.
Can AI image detectors be wrong?
Yes, in both directions: they miss AI images, and they flag real photos as AI. In Bellingcat's test, one tool called seven of ten compressed AI images real and six of 20 real photos AI-made.
Can I run a screenshot through an AI detector?
You can, but expect worse results: screenshots and compressed copies lose some of the traces that detectors look for (Bellingcat, 2023). Try to find the original file first.
Course: Learn to recognise, analyse and counter AI-driven disinformation and FIMI in Debunk.org's 10-hour, self-paced online course → Start FIMI 101
Note: We tested the detectors on 26 October 2026, and read their terms and checked every link on 27 October 2026.
