Why do AI detectors flag human writing?

An AI detector does not look for a machine. It measures how predictable your writing is, and then reports that number as if it were an answer about who wrote the text. This is what the measurement actually is, where it comes from, and what happened when Stanford ran it on essays written years before ChatGPT existed. Chapters 00:00 The question 00:58 Your phone has been doing this for years 01:59 What a language model actually scores 05:23 What else scores low and flat 08:20 Who sells this, and what they claim 12:18 What it actually knows Sources Every figure in this video traces to one of these. • Liang, Yuksekgonul, Mao, Wu & Zou, GPT detectors are biased against non-native English writers, Patterns 4(7), 2023 — arxiv.org/abs/2304.02819 • OpenAI, New AI classifier for indicating AI-written text, 31 January 2023, and its withdrawal note of 20 July 2023 • GPTZero, What is perplexity & burstiness for AI detection?, including the autumn-2023 update note that both were retired • Chechitelli (Turnitin), Understanding the false positive rate for sentences of our AI writing detection capability, 14 June 2023 • Vanderbilt University, Guidance on AI detection and why we're disabling Turnitin's AI detector • Common Sense Media, The Dawn of the AI Era: Teens, Parents, and the Adoption of Generative AI at Home and School, 2024 • Dathathri et al., Scalable watermarking for identifying large language model outputs, Nature, 2024 • Pudasaini, Miralles-Pechuán, Lillis & Llorens Salvador, arXiv:2603.23146, 2026 — arxiv.org/abs/2603.23146 • GPT-2 model card and config.json — huggingface.co/openai-community/gpt2