Ask HN: How do systems (or people) detect when a text is written by an LLM
gemini gpt-5 perplexity
| Source: HN | Original article
A wave of curiosity has rippled through the Hacker News community after a recent “Ask HN” thread asked how developers and analysts can tell whether a passage was generated by a large language model (LLM). The post, which quickly climbed to the front page, sparked a flurry of replies that laid out the technical playbook behind today’s AI‑text detectors.
At the core of most commercial tools is the measurement of statistical “perplexity” – the degree to which a string of words follows predictable patterns typical of machine‑generated output. Low perplexity, combined with unusually uniform token distributions, flags a text as likely synthetic. OpenAI’s recent watermarking scheme, embedded directly into model logits, adds a covert signature that can be extracted with a simple classifier, while Google’s Gemini team is experimenting with similar traceable tokens.
Beyond algorithmic tricks, researchers are revisiting classic stylometry: sentence length variance, lexical richness, and the presence of idiosyncratic errors that humans tend to make but LLMs smooth over. Open‑source projects such as “guppylm” and the newly released “Modo” have incorporated these heuristics into lightweight detectors that can run on a laptop, widening access beyond big‑tech APIs.
The surge of interest matters because detection is becoming a prerequisite for content moderation, academic integrity and legal compliance. As generative models grow more capable and begin to self‑watermark, the arms race between creators and detectors is set to intensify. Regulators in the EU and Nordic countries are already drafting guidelines that could mandate transparent labeling of AI‑generated text.
What to watch next: OpenAI plans to roll out an opt‑in watermark for GPT‑5 later this year, and a consortium of universities announced a benchmark suite for detection robustness at the upcoming NeurIPS conference. The outcome of these initiatives will shape whether the industry can keep pace with ever‑more convincing synthetic prose.
Sources
Back to AIPULSEN