AI Detector

Programs like ChatGPT can write texts at the push of a button that sound as if they were written by a human. This raises a new question: who actually wrote this? An AI detector is a tool that tries to answer exactly that. You feed in a text, an image, or a video, and the tool estimates whether a human or a machine is behind it. The word estimates is important here: the result is a probability, such as “87 percent machine-generated,” and not proof. These tools are used above all in schools, universities, and newsrooms.

What’s at stake with a false alarm

The need is obvious. Teachers want to know whether a piece of homework was written by the student themselves. HR departments receive job applications that are suspiciously polished. And news desks have to decide whether a photo from a crisis region is genuine or was created on a computer. Without any tools, all these cases come down to nothing more than gut feeling.

The problem is that mistakes here are costly. If a detector flags a self-written essay as machine-generated, a student ends up under suspicion that is hard to disprove. Such false alarms are called false positives. They affect certain groups more often: people writing in a foreign language, and authors with a very plain, terse style. Studies have shown that texts by non-native speakers are wrongly accused significantly more often.

That’s why it’s now considered a rule of thumb: a detector result is a hint, not a verdict. Several major providers have even shut down their own detection tools again because the accuracy rate was too low. Anyone using such tools should always cross-check the result against conversations, intermediate drafts, or earlier versions.

What the programs use to spot machine patterns

Language models choose their words based on probabilities. They tend to reach for the most obvious phrasing. This produces texts that are noticeably even and predictable. It is exactly this evenness that many detectors measure. Experts speak of perplexity, meaning how surprising the word choice is for a model. Human texts jump around more: they contain unusual turns of phrase, clunky sentences, and breaks in rhythm.

You can picture it like a handwriting sample, albeit a very crude one. A printer writes every letter identically, a human never does. The detector looks for this telltale regularity. With images it works similarly, just with different features: unnatural noise, odd details on hands, teeth, or lettering in the background.

This approach has a fundamental weakness. If you slightly rewrite a machine-generated text or run it through a translation step, you break up the pattern. That’s why providers are increasingly turning to a different approach: watermarking. Here, the program marks its own outputs already at the moment of generation, for instance through invisible signals in the image data. This works more reliably, but it only helps with models that voluntarily take part.

AI detectors in everyday school life and in image forensics

This technology is most commonly found in schools and universities. Learning platforms and plagiarism checkers like Turnitin have built in AI detection. Many universities have responded by changing their forms of assessment: more oral defenses, more work done in traceable steps. The detector is only one building block among several here.

The second major area of use is images and videos. After major news events, fake photos regularly circulate on social networks. Fact-checkers use detection tools to sort out suspicious cases, but then additionally check the image source, location, and metadata. One should not confuse an AI detector with a plagiarism checker: the latter compares a text against existing sources, while the former only judges style.

In business news, AI detectors come up in the context of fraud. Banks check video calls during identity verification, and insurers check photos submitted with damage claims. Regulation plays a role here too: the European AI Act requires that artificially generated content be labeled. This is gradually shifting the task from after-the-fact guessing toward proof that is supplied right from the start.