Why AI image detectors disagree, and what a percentage score means
Each pixel-based detector is its own model, so two of them can score one picture differently. A percentage is one model's output, and the vendor page we read gives no cutoff or accuracy figure for it. A different kind of tool looks for signals a maker put in the file. We have not measured any detector's accuracy.
Why do AI image detectors give different answers?
A pixel-based detector is a trained model. It looks at the picture itself and returns a score. Sightengine, one vendor, says in its documentation for AI-generated image detection that its model analyzes the visual (pixel) content, and that it was trained on millions of artificially-created and human-created images.
Nothing on that page says another vendor trained on the same pictures. Two models with different training data have no reason to agree. Sightengine also says its model is updated on an ongoing basis as new generators become available, so a score for one file may change over time.
The same page says detection is robust to re-encoding, compression and resizing. It adds that confidence may drop somewhat on heavily degraded images or screenshots with significant UI chrome. We did not test that, and we cannot say how any other detector handles a shared or screenshotted picture.
What does a percentage score mean?
It means what the vendor's model says it means, and no more. On the Sightengine page we read, the ai_generated value is defined as the model's confidence, from 0 to 1, that the image was produced or edited by a generative model. Higher means more likely. As of October 30, 2026 we found no cutoff on that page for deciding what counts as a yes, and no accuracy figure.
That matters when you compare tools. A score of 0.9 from one vendor and 60% from another are not on a shared scale, and nothing we read says they line up. The same page calls the value a probability score in one place, but it says nothing about how the number relates to real-world odds. Read a percentage as a model's output, not as a measured chance.
Google draws a similar line on its SynthID Detector page. In its words, general AI detectors analyze passive artifacts and patterns left behind by generative models to estimate the likelihood of AI generation, and these tools can sometimes be less reliable than watermarking. Google makes SynthID, one of the watermarks.
Is a detector the same as a maker's own checker?
No. Some makers run checkers that look for the signals their own products add. Those tools answer a narrower question.
- Google. Its SynthID Detector page says "This is not a general AI detector." It says it can only detect media from the companies that have adopted SynthID. For a result of "watermark not detected", it says this does not rule out AI generation or manipulation.
- Gemini. Google's Gemini help page on verifying AI-generated content says Gemini can currently only recognize content created by Google AI tools. It also says that when no SynthID watermark is detected, another AI system could have created the file.
- OpenAI. Its guide to content provenance says the tool is not a general-purpose AI detector and checks for supported OpenAI signals. OpenAI says to treat "not detected" as an absence of detected evidence, not proof that a human made the content.
These tools answer "did the signal we look for show up?" and nothing broader.
What does each kind of tool read, and what does a blank result say?
| Kind of tool | What it reads | What a low or empty result says |
|---|---|---|
| Pixel-based detector (Sightengine's page is the example) | The picture's pixels. Sightengine says metadata, EXIF tags, C2PA provenance and invisible watermarks are ignored. | A low score is the model's low confidence. The page gives no cutoff and no accuracy figure. |
| A maker's watermark or record checker (SynthID Detector, Gemini, OpenAI) | Signals tied to that maker's products, such as a watermark or a signed record. | Nothing found says only that the signal was not found. Google's detector page says this does not rule out AI. Gemini's page says another AI system could have made the file. OpenAI's page says it is not proof of human creation. |
| Our checker | Signed Content Credentials, source labels in metadata, and one invisible pattern, the one Stable Diffusion's reference software adds. | Nothing found is reported as an absence, never as a score. |
What did we measure about files that carry no signals?
We ran 32 real files through our checker: 29 from Wikimedia Commons categories named "Images generated by" a product, 2 vendor samples and the C2PA project's test image. Third parties uploaded the Commons files, so we cannot know how each was handled before upload. We also cannot say how each image was really made, and we draw no conclusion about any product.
Of the 31 files other than the test image, our signal reader found a signed record on 8, an unsigned source-type declaration on 1 and a tool name in metadata on 1. The other 21 carried nothing that says how the image was made: 16 read "No supported provenance signals found" and 5 read "Metadata found; source not established". We measured this on October 30, 2026 in Chrome. The per-file results are on our accuracy page.
A pixel-based detector is built to return a score for each image it reads, so it would normally have an answer for these 31 too. We did not run one, so we cannot say whether those numbers would be right. A score fills the space where a signal reader has an absence. That is why our checker reports an absence and never a probability.
What this means when you check an image
Open the checker, choose the file, and look at the headline. For 16 of the 21 files above that carried nothing, it was "No supported provenance signals found". The line under it says: "This does not mean the image is real or human-made." The limits line says: "These checks cannot determine whether an image is AI-generated."
If the file does carry a signed record, the headline says so instead, for example "Trusted Content Credentials declare AI use". Even then it reports what a signer declared, not what the picture shows.
If you use a detector, treat its score as one model's output and keep the file's source in mind. Our main page on this topic has more, and What is SynthID? covers Google's watermark. We do not check SynthID.
Guides on this topic
- How to tell if an image is AI is the main page on this topic.
- Check whether a photo has been edited covers what editing leaves in a file.
- How to verify a photo someone sent you covers a picture with an unknown source.
- Is this profile picture AI? covers a common case.
Frequently asked questions
- Are AI image detectors accurate?
- We do not know, and we have not measured any. The Sightengine page we read states no accuracy figure, and we did not run any detector. Ask a vendor for results on files like yours before relying on a score.
- Does a 90 percent score mean there is a 90 percent chance an image was made by AI?
- Not necessarily. The Sightengine page defines its value as the model's confidence from 0 to 1. We found no statement there about how the number relates to real-world odds.
- Why did one tool say AI and another say not?
- They may be different kinds of tool. A pixel model scores the picture. A maker's checker looks for that maker's own signals, and the vendor pages we read say that finding nothing does not show who made the file.
- Does stripping metadata change a pixel score?
- Sightengine says metadata, EXIF tags, C2PA provenance and invisible watermarks are ignored, so stripping them has no effect on its result. A signal reader, by contrast, reads those very signals.