July 16, 2026
Bot or not? Comment section meltdown
Detecting LLM-Generated Texts with "Classical" Machine Learning
Old-school software may spot AI writing, but the comments turned it into a culture war
TLDR: A developer claims old-school software can often spot AI-written text surprisingly well, without using giant fancy models. Commenters turned that into a bigger fight over whether this could save the web from bot sludge or whether AI detection is basically digital astrology.
A developer says the big twist in the AI cheating panic may be hilariously low-tech: instead of fancy futuristic systems, plain old machine-learning tools can often tell whether text was written by a person or a chatbot. The demo reportedly catches single AI-written sentences about 85% of the time, and the whole thing grew out of one very relatable rage spiral: the author got fed up seeing fanfiction tags flooded with suspiciously robotic posts and decided, essentially, “fine, I’ll build my own detector.”
But the real fireworks were in the community reaction, where people instantly split into camps. One side was basically cheering for an AI ad blocker for words. Commenter Krssst imagined a browser extension scanning every paragraph online so people don’t waste time reading machine-made sludge. That idea got big “please save the internet” energy. Meanwhile, the skeptics came in swinging. Akersten dismissed the whole dream of detection as “tarot card reading,” arguing text just doesn’t carry enough secret fingerprints to prove where it came from.
And then came the existential doomposting. Cyanydeez warned that today’s detector might become tomorrow’s useless relic because humans themselves are starting to sound like the bots they read all day — a bleakly funny “the kids will be trained by autocomplete” take. Elsewhere, people traded war stories about painfully slow tools, hallucinating systems, and OCR setups that crawl along while making stuff up. So yes, the article is about catching AI text — but the comments are really about a bigger fear: what happens when the internet becomes unreadable, and nobody agrees whether the detector is a shield or a superstition?
Key Points
- •The article claims that mainstream LLM-generated text in early 2026 shows statistical patterns distinguishable from human writing using classical machine learning.
- •An online demo and GitHub repository are provided for a preliminary detector called AITextDetector.
- •The demo model is described as non-general-purpose and not rigorously optimized, with about 85% single-sentence accuracy on a test set.
- •The author says a perplexity-based detection method performed poorly because of false positives, false negatives, thresholding issues, and deployment costs.
- •The article identifies scikit-learn classifiers such as LinearSVC and Naive Bayes as a more successful approach, supported by labeled human and AI-generated text data.