There are tools that claim to recognise AI images, voices and videos. Some measure something real, none of them decide a case. This guide explains what these tools actually check, where they fail, and which routine works better in practice.
An AI detector is a second AI: a classifier trained on real and generated images of a certain era. It returns a probability, not a fact. If it says 85 per cent AI, that means the image resembles the generated images it was trained on — nothing more.
Two consequences: the detector is only as good as its training data, and it ages quickly, because generators change every few months. Anything that shrinks or re-compresses the image — a screenshot, a chat app, a social network — removes exactly the fine detail the classifier relies on. And the error goes both ways: real photos get labelled AI, AI images slip through.
Keep in mind: A detector result is one piece of information. It is not a verdict, and it is definitely not something to base a public accusation on.
Many generators and some cameras write technical metadata into the file, recording which tool produced it. Some companies also embed invisible watermarks in the pixels. Where they survive, they are the strongest signal there is, because they come from the producing side.
The problem is survival. Uploading to a social network, taking a screenshot, converting to JPEG or pasting into a document usually strips metadata and often damages the watermark. So: present and intact provenance data is a strong indicator; missing provenance data means nothing at all. Most images on the internet have none.
Speech models can now imitate a voice from a few seconds of audio. That means a familiar voice in a voice message is no longer a reliable proof of identity — and, just as important, an unfamiliar or slightly odd voice is not proof that something is fake.
What helps in practice is not listening more closely, it is checking the channel: did the message arrive through the usual app, does the request make sense in context, and can you verify it through a second, known route? For anything involving money or access, confirm through a channel you chose yourself. There is no reliable everyday “voice detector”, and there does not need to be if you verify the request instead of the recording.
Not every questionable image is AI. A large part of misleading material is ordinary editing: a person cut out of a photo, a sign painted over, a screenshot with a missing line, an old photo presented as new. This kind of manipulation is often technically undetectable in a single image — the deception lives in the missing context.
So check the story before the pixels. Does the same scene appear in other independent sources? Is the image older than the event it claims to show? Is a quote attributed to someone shown together with a photo from an unrelated day? A reverse image search and a look at the original article answer more questions than any detail in the picture.
what typical AI mistakes look likeThis is the sentence that matters most on this page: no single visual feature proves that an image came from a generator. A strange hand is produced by motion blur, wide-angle lenses and aggressive phone sharpening. Blurred text is produced by compression. Extra reflections are produced by double glazing and shop windows.
The same is true in the other direction. A perfect image is not proof of a camera. A convincing voice is not proof of a person. Every clue is one line of evidence, and lines only add up.
You do not need to judge the image. You need to judge the claim built around it. Five questions, in this order: Where does the image come from, and can I see the original? What does the text around it claim? Do other independent sources show the same thing? Is anything in the image physically inconsistent, and is that inconsistency explainable? And how much would it matter if I am wrong?
Then practise. The photo, fake-news, flag, video and language modes in the game give you feedback on exactly these estimates — and the learning area shows you worked examples so you can compare your estimate with the explanation.
try the Language Detective modeNo. Detectors give probabilities, are trained on older generators and lose accuracy as soon as an image is re-compressed. They are useful as a hint, not as evidence.
No. Watermarks and provenance data disappear through screenshots, uploads and re-encoding. Only their presence is informative, never their absence.
Not by listening more closely, but by verifying the request through a channel you chose yourself — a call back to a known number, or a second known contact.
Say that you are unsure. Instead of labelling an image as AI, describe what fits together and what does not, and name your confidence level. That is more accurate and safer for the people in the picture.
Key takeaway
Detectors, metadata and watermarks can support a judgement, but none of them replaces it. The reliable part of the work is source, context and consistency — and the honest part is naming your uncertainty.
Every game has ten images: five real photos and five AI images. After each guess you see the answer, so you learn while you play. Free, no account needed.
Play the quizHow to Tell if a Photo Is AI-Generated: 6 Practical Checks
How to tell if a photo is AI-generated: six checks in the order that helps — source, first impression, details, physics and tools, plus where each clue fails.
How AI-generated images are actually created
Training data, the diffusion process and the polished look: why generators win at smooth surfaces and fail at strict rules.
Typical AI mistakes: hands, text, shadows and reflections
Why generators fail at things with strict rules, what to look for — and when real photos show the same supposed mistakes.
Wir nutzen lokalen Speicher für Betrieb, Sicherheit und Spielfortschritt. Mit „Alle akzeptieren" erlaubst du zusätzlich Werbung (Google AdSense), die Cookies setzen kann. Details in der Datenschutzerklärung.