Detecting LLM-Generated Texts with “Classical” Machine Learning
Efforts to detect large language model (LLM)–generated text using relatively small, “classical” machine-learning classifiers are showing surprisingly high accuracy, inspiring ideas like browser extensions that flag or filter AI-written content much like ad blockers. Commenters debate whether such detection can ever be reliable or fair enough for high‑stakes uses like grading or plagiarism enforcement, given false positives, evolving models, and easy evasion via style changes or post‑processing. Many argue that the more realistic long‑term approaches are focusing on provenance and “proof of work” (e.g., edit histories) or simply using detectors as heuristic slop filters for low‑effort AI content, rather than as definitive arbiters of authorship.
Detectability Today
- Many argue current LLM outputs are detectably different from typical human writing, especially commercial, RLHF-tuned chat models with “corporate / engagement-maximizing” tone and stylistic quirks (e.g., length, hedging, em‑dash usage).
- Others counter that text alone is too low-signal to robustly encode provenance; at best we’re spotting transient “tells” that will vanish, likening some claims to astrology.
- Several point to systems like Pangram and the blog’s classifier (synthetic “twin” texts, small transformers) as evidence that high AUC and low false-positive rates are currently achievable, at least on benchmark datasets.
Reliability and False Positives
- Strong concern about using detectors in high‑stakes settings (grades, jobs, plagiarism accusations).
- False positives, even at 1 in 10,000, are viewed as unacceptable if the consequence is e.g. expulsion; tools should be at most one piece of evidence with human review and appeal paths.
- For casual browsing, many see 80–90% accuracy as “good enough” to skip likely slop without accusing specific authors.
Arms Race and Evasion
- Recurrent theme: any public detector can be used as an adversary (GAN-style) to train generators and post-processors that evade it.
- Suggestions include high-temperature / advanced sampling, style transfer, or second-pass models that rephrase output.
- Some think large labs don’t prioritize stealth; others expect spammers/fraudsters will if platforms start blocking AI slop at scale.
Language Drift and Human Style
- Worry that pervasive LLM use will shift human writing toward “LLM voice,” degrading detectors over time and unfairly flagging humans—especially ESL writers or those with formal/corporate styles.
- Others claim humans are still more varied and worse at strict optimization than models, so detectable gaps remain for now.
Tools and Practical Uses
- Multiple references to browser extensions and small on-device classifiers used to label social posts and comments as likely AI.
- These are framed as personal slop filters, not forensic tools.
Alternatives to Detection
- Several suggest focusing on “proof of work” (edit histories, timestamped drafts, audit trails) and measuring effort/quality rather than provenance.
- Some venues reportedly combine detectors with audit trails to judge extent of AI involvement.
Attitudes toward AI-Generated Text
- Many dislike the bland, over-polite, “insufferable” tone and length of low-effort AI content.
- A minority say they genuinely enjoy LLM outputs and revisit them like reference material.