AI Detectors Are Making AI Better at Lying. Here’s How.

You’ve probably seen the warnings: an AI detector flagged your student’s essay as machine-written. Or you ran your own email through a checker and got a ‘70% likely AI’ result. And you panicked. But here’s the thing nobody tells you: Every time a new AI detector gets good, it teaches the next generation of AI how to lie better. That’s not a bug. That’s the whole game.

Last week, researchers at a major lab released a technical report on a classifier called Pangram. It’s better than GPT-2’s detector. It catches subtle artifacts from GPT-4 and Claude. It seems like progress. But dig into the paper and you’ll see the uncomfortable truth: the classifier’s success depends entirely on the weaknesses of the current generation of text generators. The moment those generators learn to avoid the patterns Pangram relies on? Back to square one.

This isn’t a war we can win. It’s a Red Queen race, and we’re running just to stay in place. Every advance in detection forces an advance in generation. The better our classifiers get at spotting AI text, the more data we provide to train the next generation of models to sound perfectly human. We’re not building shields; we’re forging sharper swords for the other side.

Most people assume the solution is a better classifier. They think, ‘Eventually we’ll have a detector that can spot any AI writing.’ But that assumption ignores a fundamental principle: a classifier that works across all generations must be general, but a general classifier is too weak—it misses the subtle signals. A specific classifier catches current patterns but becomes obsolete in months. You can’t have both.

Let me give you a concrete example. In the Pangram paper, the classifier was trained on outputs from GPT-4, Gemini, and a few others. It performed well—above 90% accuracy in some tests. But the researchers also tested it on a newer, unreleased model. Accuracy dropped to 62%. Six months of development, and the detector was already outdated. That’s not an anomaly. That’s the nature of the arms race.

And it gets worse. The very act of deploying a classifier publicly gives adversarial actors a playground to reverse-engineer the detection methods. They can craft texts that fool the classifier while still being machine-generated. The Pangram team even noted that their classifier is ‘susceptible to simple adversarial attacks.’ In other words, the tool you’re using to separate human from machine is easily tricked by anyone who knows how it works.

So what does this mean for you? It means the next time you read a glowing product review, a heartfelt op-ed, or a breaking news story, you can’t be sure. The boundary between human and AI authorship is already a blur, and every new detector pushes it further into fog. The uncomfortable truth is that trust in online text is becoming a luxury we can no longer afford. Not because AI is evil, but because the tools we build to catch it are training it to be more convincing.

We need to stop pretending that technical solutions alone will save us. Verification will have to come from outside the text—provenance systems, cryptographic signatures, human-in-the-loop workflows. But those systems require infrastructure we don’t have and trust we’ve already lost. Until we build that, assume every paragraph you read could be written by a machine that learned its fluency from the very detectors designed to expose it.