synthetic mirroring

Synthetic mirroring works by first learning the habits of AI-written text, then holding up that learned pattern like a mirror to recognize similar writing. It's a bit like a counterfeit money detector who learns to spot fakes by making counterfeit bills themselves, memorizing the distinctive texture and ink smell, so that later, the moment they see another bill, they think "this feels familiar." In other words, the tool built to catch AI has AI baked into it.

The key point is that this method gives results as probabilities. Feed in a piece of writing, and it spits out a score like "this is probably X% AI-generated" — but that score isn't solid proof, like a confession from the actual writer or a documented drafting history. It's just an estimate of how closely the patterns match. That means someone whose writing style happens to resemble AI-generated text could get an unfairly high score, while a piece that leaned heavily on AI assistance could slip through with a low one.

Recently in the publishing industry, there's been a string of cases where a single score like this led to canceled contracts or cast doubt on award-winning works, fueling growing debate over just how reliable this mirroring approach really is.

The article explains that "Pangram is a service that tells you, as a probability, how much AI was involved — but AI itself is used in making that judgment." That's synthetic mirroring in action. The easy mistake is treating that score as direct "proof that AI wrote it." In reality, it's just a probability estimate based on pattern similarity, not a definitive verdict.