What is an AI detector, and can it really spot writing by no one

The writing platform Substack has unveiled a new tool designed to help readers determine whether the content they are reading may have been written by AI. Developed by an AI detection company called Pangram, the tool scans posts, notes, replies and comments to provide an estimate of how much of the text may have been AI-generated or written with AI assistance.

The emergence of such tools responds to a problem created by large language models becoming increasingly capable of producing written content: readers can no longer reliably tell whether the text in front of them was produced by a human or an algorithm. Companies like Pangram are building statistical models aimed at resolving that uncertainty.

So how do these detection tools actually work, technically? Most approaches rest on the idea that AI-generated text exhibits certain patterns that are statistically distinct from human writing. AI models, for instance, may tend to use certain word combinations with more predictable frequency than humans do, or show a certain consistency in sentence structure. Detection models are trained to recognize these kinds of subtle statistical signatures.

But this approach has an inherent limitation: as AI models improve, they move ever closer to human writing, narrowing the statistical gap. That creates an ongoing race between detection tools and generative models — not unlike the relationship between malware and antivirus software, where each side keeps adapting to the other.

The accuracy of these tools is also a contested issue. Independent tests have shown AI detection tools sometimes mistakenly flag entirely human-written text as AI-generated — a problem that occurs more frequently in text written by non-native English speakers, or by people writing in a highly formal, repetitive style. Such false positives can carry serious consequences; in academic settings, for example, students can be unfairly accused of plagiarism or of using AI.

Errors in the opposite direction are also possible: some AI-generated text, particularly if edited by a human or “humanized,” can evade detection tools entirely. That means publishers and platforms need to treat the results of such tools not as definitive fact, but as a probability estimate.

Substack's tool appears to acknowledge this nuance: rather than presenting a definitive yes-or-no verdict, the platform offers an estimate of how much of the text may be AI-derived. That approach gives readers room to form a more informed judgment, while implicitly acknowledging that the tool itself can be wrong.

For publishers, the appeal of such tools is understandable: preserving reader trust is critical, particularly at a time when the independent journalism and writing economy is growing. Assurance that a writer is genuinely writing in their own voice is part of the value underpinning the subscription model.

Critics, however, note that such tools can also create a false sense of security. A post being labeled “likely human-written” does not mean that content is accurate or reliable — it merely offers a statistical estimate about the text's origin.

For now, the tool is rolling out on the web and iOS, with an Android version coming soon. Substack's move is seen as part of a broader debate over how platforms can preserve reader trust at a time when verifying the origin of written content is becoming increasingly complex.