6 min read · Updated August 2026
What AI-detection tools actually measure — and where they fail
Detection tools are often presented as a single verdict with a percentage attached, which invites people to treat them as a lie detector. They are closer to a set of independent tests, each with its own blind spots. Here is what those tests actually look at, and what a result can and cannot support.
Metadata and provenance
The cheapest check reads the data written alongside the pixels: camera make and model, lens, exposure settings, timestamps, editing software, and increasingly C2PA content credentials that record how a file was produced and modified.
Its strength is that contradictions are hard to fake carelessly — a claimed camera that does not produce that resolution, or an edit history that names an image generator. Its weakness is that absence proves nothing: every major platform strips metadata on upload, so most honest images arrive bare.
Compression and resave history
JPEG compression leaves a periodic grid, and every resave stacks another grid on top. Analysing those layers can reveal that one region has been compressed a different number of times than the rest — a classic sign that something was pasted in.
This is powerful on lightly handled files and nearly useless on heavily reposted ones, where dozens of recompressions have flattened the history into noise.
Texture, noise and frequency statistics
Camera sensors produce noise with a characteristic, uneven distribution. Generators produce something statistically different: often too little high-frequency detail, sometimes periodic patterns introduced by the upscaling step, and telltale smoothness in dark regions.
These measurements are strong on original files and fragile in the face of denoising, beautification filters, screenshots and social media resizing — all of which alter exactly the statistics being measured.
Learned classifiers
The best-known approach trains a model on large sets of real and generated media and asks it to judge new files. On material resembling its training data, accuracy can be very high.
The catch is generalisation. A classifier trained before a new generator existed frequently misjudges that generator's output, and its confidence stays high while doing so. This is why a bare percentage with no reasoning behind it deserves scepticism.
Why we show reasons instead of a single score
Because each method fails differently, the useful output is not one number but a list of findings you can check yourself: what was examined, what looked normal, what looked wrong, and how much weight that deserves.
That framing also makes the honest cases legible. A heavily compressed screenshot with no metadata genuinely does not contain enough evidence to judge, and saying so is more useful than manufacturing a confident verdict.
How to use a result responsibly
Treat a verdict as one input alongside provenance, corroboration and plain plausibility. Where the stakes are high — legal, journalistic, safety-related — seek the original file and a second opinion rather than acting on an automated opinion.
And be as sceptical of a clean result as of a damning one. 'No evidence of generation found' means the tests did not fire, not that the content is true.
The short version
- Detection combines several independent tests, each with different blind spots.
- Reposting and filters destroy the very signals most methods rely on.
- Classifier confidence does not fall when the tool meets an unfamiliar generator.
- Reasons you can verify are worth more than a percentage you cannot.