Why AI Detectors Fail

1. Compression eats the evidence

Most forensic signals live in fine detail: the noise floor, high-frequency content, subtle compression artefacts. Every platform re-encodes uploads, and each pass removes some of it.

An image that went camera → phone gallery → WhatsApp → Instagram → screenshot has been re-encoded four times. What reaches you is a shadow of the original signal, and any detector working on it is guessing more than measuring.

What to do: get the original file. Ask for it to be sent as a document rather than as a photo — messengers do not re-compress attachments.

2. Stylised content has less to measure

Photographs carry sensor noise, optical falloff and texture. Flat illustration carries none of that by design, so several methods lose their footing at once.

This is why anime, vector art and clean digital painting produce lower confidence — see our anime page for how far that goes. A tool reporting equal confidence on a photo and a cartoon is not measuring the difference.

What to do: for artwork, ask for process evidence — layered files, timelapse, sketch history. It settles the question better than any score.

3. Partial generation confuses everything

Generative fill, background replacement and object insertion produce a hybrid: mostly real photo, one synthetic region.

Generation detection looks for global properties and finds a mostly-real image. Manipulation detection looks for regional inconsistency and finds it. Run only the first and you get a clean result on an edited photo.

What to do: run both. Our manipulation analysis covers the case that generation detection misses.

4. New generators arrive faster than training data

It is an arms race and detection is structurally behind. New models are tuned, sometimes explicitly, to remove the traces detectors look for.

Methods based on general statistical properties of generation generalise better than classifiers trained to recognise specific products — which is why we run nine forensic methods alongside three neural networks, as explained in how detection works.

What to do: treat confidence as provisional, especially on very recent output.

5. False positives — the failure that hurts people

Everything above is about missing a fake. This one is the reverse, and it does more damage.

A clean digital illustration flagged as generated can end an artist's commissions. A retouched photograph flagged as manipulated can sink an insurance claim. In AI-text detection this pattern already produced documented harm to students accused on the basis of a score.

The asymmetry matters: missing a fake is usually recoverable, wrongly accusing a person often is not.

What to do: never act on a score alone. Look at which methods fired, get the source files, and put a human in the loop before any consequence lands on someone.

What a good result actually looks like

Not a high percentage. A *coherent* one.

Nine methods agreeing at 85% is far stronger evidence than one method shouting 99% while the others shrug. That is why every result shows the full breakdown, and why an honest tool returns “inconclusive” instead of inventing certainty — roughly one check in eleven lands there for us.

If you want to test any of this yourself, our comparison page explains how to run a fair test on images where you already know the truth.

Key points

• Compression destroys forensic evidence — always work from the original file.
• Stylised art gives less to measure, so confidence is genuinely lower.
• Partially generated images need manipulation analysis, not just generation detection.
• New generators outrun training data; treat recent output as provisional.
• False positives harm real people and are the more serious failure mode.
• Coherence across methods beats a single high percentage.

Frequently asked questions

Check an image yourself

Reading about it only goes so far. Our AI image detector runs twelve independent methods and shows you every score, not just a verdict.

More from the blog