Most advice about AI images starts with the clues. This guide starts one step earlier: what a generator learns, how an image is built out of noise, and why that process produces the very mistakes you are asked to spot. Once you know the mechanism, the clues stop looking like a random checklist.
An image generator was trained on enormous numbers of images, each paired with a written description. From those pairs it learned statistical connections: which shapes, colours and textures usually appear together with which words. It did not store the images as a searchable archive — it compressed their patterns into numbers.
That is why a generator cannot answer the question “how does a hand really look”. It can only answer “which arrangement of pixels is most typical for something described as a hand”. Those are two different questions, and the difference is exactly where the typical AI mistakes come from.
Most current systems are diffusion models. They start with pure visual noise and remove a small part of it, step by step, always nudged in the direction of your prompt. After a few dozen steps something image-like remains.
Two consequences matter for you as a player. First: the model has no scene in mind — it has no floor plan, no lighting plan, no inventory of objects. Everything appears because it is statistically plausible in that spot, not because it physically fits. Second: the result is a compromise between many similar training images, which is why faces, skin and skies often look unusually smooth and even.
Some things follow hard rules: a hand has five fingers in a fixed order, letters have a fixed number of strokes, a chain link must close around the next one, a reflection must show the same object. A statistical model learns the average of what it saw — and where the rule has many exceptions, the average is not a rule at all.
Count the fingers in a crowded hand. Trace one letter of a shop sign. Follow a fence into the distance and check whether the posts stay consistent. These are the areas where the mechanism fails most visibly.
Keep in mind: Failures are not what identifies AI. Real photos also contain merged fingers, blurred text and broken patterns — through motion, compression and perspective. Treat a failure as a reason to look further, never as proof on its own.
Because the generator aims for plausibility, it tends to produce the clean version of a scene: soft studio light, tidy backgrounds, even skin, symmetrical composition. Prompts make this stronger, because words like “perfect”, “professional” or “4K” push the result towards the polished end of the training data.
In real photography, the opposite is normal: mixed light, sensor noise, a tilted horizon, a stranger half visible at the edge. That is why an image can feel artificial without any identifiable error — and why “it looks too clean” is a feeling, not a finding.
If you understand the mechanism, three honest conclusions follow. One: a very convincing image is not proof that a camera was involved, because plausibility is precisely what these models optimise. Two: an odd detail is not proof of AI, because real cameras and real editing produce odd details too. Three: you can never verify an image with certainty — you can only weigh how much fits together.
That is uncomfortable but useful. It turns the question from “guess the answer” into “estimate how much evidence points in one direction”. Our detection guide breaks that estimation into six steps you can run in under a minute.
run the six-step checkThis is the part where reading turns into practice. Open one mode of the game — photos, fake news, flags or videos — and play five rounds slowly, without watching the timer. For each image, name one thing that fits the scene and one thing that does not. Say out loud which of the two weighs more.
Then read the explanation of the two guides about typical mistakes and about the limits of detection. Playing first and reading afterwards works better than the other way round, because your own wrong answers are the examples you remember.
read what detection can and cannot doNo. It reproduces statistical patterns from many training images, not single pictures. Occasionally a result resembles a training image very closely, but the normal case is a new arrangement of learned patterns.
Because it learned what is typical, not what is correct. Faces appear huge numbers of times in similar composition in training data, while hands are small, hidden or moving — much harder to learn precisely.
For the obvious cases, yes. Odd hands and nonsense text are becoming rarer. What remains is what has always mattered: source, context and consistency inside the image.
Key takeaway
An image generator reconstructs what is typical, not what is real. That is why it wins at smooth surfaces and loses at strict rules — and why a single detail never decides a case.
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 reliable is AI detection? Detectors, metadata and watermarks
What detectors really measure, why missing watermarks prove nothing, how to check suspicious voice messages and manipulated images.
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.