wyper.io logowyper.io Blog Fact-Check Add to Chrome
Home › How to check if an image is AI-generated

How to check if an image is AI-generated

There is a whole industry selling you a percentage. We measured that percentage and decided not to sell it.

Fact-Check · last reviewed 2026-08-24

The short answer

To check whether an image is AI-generated, work on provenance rather than pixels: find the earliest appearance, look for signed content credentials, and check whether the scene is consistent with independent records of the event. Detector scores are unreliable in both directions, so they may guide your attention but must never carry a conclusion.

What we measured, and what we did with it

We were asked to add AI detection to the product. Before building it we tested whether it works. The result was no, and we published the whole method with the raw per-item data.

The design is the important part: the human side was 120 Hacker News comments posted before 30 November 2022, the day ChatGPT was released, because text published before that date cannot have come from a language model. The AI side was 120 texts generated in three stances, including one explicitly instructed to sound human, across three matched length bands. That is 240 items in total. Detection was not reliable enough to put in front of a user.

So we did not ship it, and we do not ship an image classifier either. That is a feature decision made against our own commercial interest, and it is published with every number so you can check it rather than take our word for it.

A tool that hands you a 92 percent confidence score on an image has told you almost nothing, and has made you feel certain. That combination is worse than silence.

The checks that survive contact with reality

  1. Look for content credentialsSome cameras and editing tools now sign an image with its capture and edit history (the C2PA standard). Present, it is strong evidence. Absent, it proves nothing, because most platforms strip it.
  2. Reverse image search the earliest copyIf the image existed before the event it claims to show, you are done. This single check settles more cases than any classifier.
  3. Read the scene against the recordWeather, shadow direction, signage language, licence plates, uniforms, architecture. A generated image is internally plausible and externally careless.
  4. Check the text in the imageGenerators still garble small text more often than cameras do, and captions or watermarks are frequently invented outright.
  5. Test the claim it carriesMost images that mislead are not generated at all. They are real photographs with a false caption, which no detector will ever catch.

Why the honest answer is often uncertainty

A verification method that always produces a verdict is not a method, it is a slot machine. There are images where the correct output is: origin unknown, claim unsupported, treat as unverified. That is a usable result, because it tells you not to share.

This is also why our results separate what is quoted from what is not. Every statement in a wyper result is either backed by a sentence we located verbatim in the cited source, or explicitly marked as having no source behind it. Of 633 real citations we re-fetched and checked, 73.1 percent were verbatim on the page, and the failure modes are published too: 19 percent of pages block automated readers entirely, and 6.4 percent could not be found.

What wyper does with an image

The web app scans a photo with on-device OCR in twelve languages, so the image itself never leaves your device, and it takes the text it finds into a normal claim check with a sourced evidence chain. It does not score the pixels, because that part does not work.

For a claim that matters, Full Spectrum adds counter-positions, source spread and open questions, with a Triple-AI reading of one shared evidence set. Sources are never ranked by origin: across 24 measured claims, 63 percent of everything returned was a primary document and fact-checkers were 1 percent.

Questions

Can AI-generated images be detected reliably?
No, not to a standard that should carry a conclusion. Detectors misfire on compressed and re-encoded images, degrade against generators newer than their training data, and produce false accusations. We measured the equivalent problem for text across 240 items and chose not to ship detection at all.
Does wyper have an AI image detector?
No, deliberately. We published the measurement behind that decision. What the app does instead is read the text in an image on your device and check the claim it carries against real sources.
What are content credentials and are they worth checking?
They are signed provenance data, following the C2PA standard, recording how an image was captured and edited. Where present they are strong evidence. Their absence means nothing, because most platforms strip metadata on upload, so treat them as a bonus rather than a test.
What is the single most useful check on a suspicious image?
Reverse image search for the earliest appearance. If the picture predates the event it is being used to illustrate, the question is settled without any judgment about how it was made.
Are watermarks a reliable sign?
No, in either direction. Watermarks are trivially cropped or generated, and plenty of authentic images carry none. Treat a watermark as a lead to follow, never as an answer.

Check a claim and look at the evidence yourself

Free: 5 checks a day, no account needed. Every statement tells you whether a quote backs it.

Open the web app Add to Chrome, free

Read on

How to check if a video is a deepfakeThe four checks that decide, and why detectors are the wrong first step. Can AI-written text be detected? We measured it240 texts, the full method, and the reason we did not ship it.