What an AI Image Detector's Percentage Actually Means

A detector's percentage is how confident that model is, not how likely it is to be right. Those are different numbers, and only one of them is on the screen. A tool that says 94% is saying "this file looks like the class I was trained to recognise" — it is not saying "I am correct 94% of the time". Open the same image in two detectors and you will usually get two different percentages, because they were trained on different image sets and none of them publishes the set it was scored against.

Confidence is not accuracy

The distinction is standard in machine learning and almost never appears on a detector's landing page. A confidence score is a property of one prediction on one file. An accuracy figure is a property of a model across many files with known answers. The first can be computed without knowing the truth; the second cannot exist without a test set.

DeepAI, whose detector is one of the most-used free ones, says this plainly on its own page: "Detection is probabilistic" — and adds that digital art, airbrush work and AI-based editing can sway the result. That sentence is doing a lot of work. It means the output is a probability, and the probability is sensitive to things that have nothing to do with whether a human made the file.

Why the same image scores differently in two tools

Three reasons, in order of how often they cause the gap:

Different training sets. A model trained mostly on one generator's output learns that generator's fingerprints. Feed it a different generator and it is guessing.

Different thresholds. Some tools report the raw probability; others report a verdict after a cut-off. A 0.6 raw score can surface as "likely AI" in one product and "inconclusive" in another with no change to the underlying signal.

Different preprocessing. Resizing, re-compression and metadata stripping happen before the model sees the file. Two tools that prepare the image differently are not answering the same question.

None of these is a bug. They are the reason a percentage is a local reading, not a measurement with units.

What vendors' accuracy numbers are and are not

Vendor pages compete on figures. TruthScan advertises "99%+ AI Image Detection Accuracy". isthisai displays 98.7% ACCURACY next to 500+ MODELS. Sightengine titles a section "Super-human accuracy". Every one of those is a claim about a test the vendor ran, and none of them, in the pages reviewed for this article, names the dataset, the mix of generators, or the threshold behind the number.

That matters because the figure is not portable. When a study tested people rather than tools, the result was worse than most vendors imply: "Analysis of approximately 287,000 image evaluations by over 12,500 global participants revealed an overall success rate of only 62%" (Thomas Roca et al., 2025, "How good are humans at detecting AI-generated images? Learnings from an experiment", arXiv:2507.18640, arxiv.org/abs/2507.18640). Note what that 62% describes: human performance on that study's own image set. The same abstract adds where that performance was strongest and weakest: participants were most accurate on human portraits, and clearly worse on natural and urban landscapes. That is the empirical version of the rule this page keeps repeating — the images that carry no obvious artifact are the ones people miss, because a portrait hands you a face to look at and a landscape hands you nothing. It cannot be used to pass or fail a detector measured somewhere else, and a vendor's 99% cannot be compared to it either.

The one kind of answer that is not a percentage

Not every check returns a probability. Provenance checks return a yes or no: does this file carry a watermark or a signed credential, or not. Google's SynthID portal works this way, and its own page states that it "is not a general AI detector" — it only speaks for content from companies that adopted the technology.

The trade is clean and worth stating explicitly:

How to read a score without being misled

Four habits, none of which require trusting the tool:

  1. Read the label, not the number. If the page says "AI-generated: 87%", check whether that is the model's probability or a confidence interval — most tools never say.
  2. Run the same file twice. A probabilistic model can return different scores for identical input. If it swings, the number is noise at that margin.
  3. Test it on files you know. Five images you generated plus five photos you took, originals not screenshots. The count of wrong verdicts on your own photos is worth more than any published accuracy claim.
  4. Never quote a percentage as proof. "94% AI" is not a finding you can put in a report. Write down which tool, which file, and what it returned.

When a percentage means nothing at all

FAQ

Is a higher percentage more reliable? No. A higher number means the model is more certain, and models can be confidently wrong — especially on generators they never saw in training.

Why do two detectors disagree on the same photo? Different training sets, different decision thresholds, different preprocessing. Disagreement is the expected result, not a malfunction.

Can I compare a detector's 99% to a study's 62% for humans? No. They were measured on different images with different methods. A number only bears comparison when it comes from the same test set and the same mix of generators.

What should I write in a report instead of a percentage? What you ran, on which file, and what it returned — plus the fact that the result is an assessment, not a determination. Only a provenance hit supports a statement of fact.

Sources

Links on this site may be affiliate links.