Pangram AI detection tool spreads to Substack as debate grows

In short: A small startup called Pangram is becoming a widely used way to estimate whether writing was made with AI, and Substack is integrating it.

What's going on

Pangram is a 24-person company based in Brooklyn. It sells a tool that scans text and returns a percentage guess for how much of it was written using AI. Think of it like a spellcheck, but instead of looking for typos, it looks for patterns that often show up in AI writing.

The tool has been pulled into several public disputes in publishing. Wired reports that Pangram scores were shared online in cases involving novels and writing contests, sometimes with very high percentages. In late July, Substack said it would integrate Pangram so readers can quickly check possible AI use.

Pangram says its newest model, called Pangram 4, has a very low false positive rate, meaning it rarely labels human writing as AI by mistake. Still, the company also says results can change depending on how much text you scan, and it performs worse on very short passages. Outside research has also raised concerns that detectors can be fooled, for example by using “humanizer” tools that rewrite AI text to look more human.

What to watch

More platforms may add AI detection, but the big question is how these scores get used. Publishers and schools may treat them as one signal, not a final verdict, since a wrong label can harm reputations. Critics also warn about bias, including the risk that some groups of writers could be flagged more often than others.

Source: Wired