Forget Fancy AI Detectors: A 20-Year-Old 'Classic' Algorithm Catches Chatbot Text Better

Was that blog post written by a human or a chatbot? Plenty of people have tried to build tools that answer the question. But the approach lighting up Hacker News this week isn’t a giant neural network at all. It’s classical machine learning — techniques older than the smartphone in your pocket.

The discussion, titled “Detecting LLM-Generated Texts with Classical Machine Learning,” landed on Hacker News on July 16, 2026, and pulled in 151 points and 103 comments. It’s a rare hot thread in an otherwise quiet corner, so it’s worth unpacking on its own terms.

Fighting neural nets without a neural net

When you hear “AI detector,” you probably picture another hulking AI — a neural network hired to catch a neural network. The intuitive arms race. This approach flips that script entirely.

“Classical” machine learning here means the pre-deep-learning toolkit: logistic regression, support vector machines, naive Bayes. You turn text into numbers — word frequencies, statistical features, vocabulary spread — and let a lightweight model sort the patterns. No cluster of GPUs required. It runs happily on a laptop.

The core insight is simple. Chatbot writing leaves a statistical fingerprint that human writing doesn’t. How often certain words appear. The distribution of sentence lengths. How varied the vocabulary is. These are subtle tells, and it turns out old-school classifiers are surprisingly good at spotting them.

Small is the whole point

The comment that got the most traction was about size. One participant noted the classifier “doesn’t seem that big” and floated an idea: package it as a browser extension and run it live against every paragraph on the screen.

Why does that matter? Big deep-learning detectors may be accurate, but they’re heavy. You ship your text to a server, wait for a response, and pay for the privilege. A lightweight classical model just runs on your own device. It’s private, it’s fast, and it’s close to free.

And this isn’t only a text-detection story. Another commenter described using the same approach to sort and organize thousands of datasheets, then bolting on a dead-simple image classifier to filter out pages that were mostly charts and diagrams. No grand model — just something that got the job done in production. That’s the real endorsement.

“Works today, sure”

It isn’t all rosy. One short comment landed with a sting: “today, sure.” It works now. Tomorrow is another question.

That skepticism points straight at the fundamental dilemma of AI detection. The moment a detector learns a chatbot’s tells, the next generation of chatbots evolves to erase them. Sword versus shield. A detector that’s razor-sharp today can be dead weight the day a new model ships.

Classical machine learning is tightly bound to the statistics of the data it trained on. That makes it precise at catching a specific model at a specific moment — but a genuinely new writing style means retraining from scratch. The lightness that was an advantage comes back around as a maintenance burden.

The lesson worth keeping

The takeaway is clear: you don’t have to reach for the biggest, flashiest tool first.

Even a problem this “AI-shaped” — catching AI-written text — turns out to be largely solvable with 20-year-old statistics. And the classical route is lighter, faster, more explainable, and runs on your own machine. In an era determined to hand every problem to a massive model, that’s a refreshing counterexample.

Just don’t mistake this shield for a permanent one. The race between detection and generation never stops, and today’s right answer carries no guarantee for tomorrow. So here’s the real question: do we actually want a browser that stamps every paragraph on screen as “human” or “AI” in real time? And when it does — how much should we trust the verdict?