The honest answer
AI text detectors exist, they are widely used, and they are not reliable enough to prove that a piece of writing was generated by AI. They produce a probability based on statistical patterns, and that probability is wrong often enough — in both directions — that no single detector score should ever be treated as evidence. The honest 2024 position, echoed by researchers and even by some detector vendors in their fine print, is that detection is a weak signal to investigate further, never a verdict to act on.
How detectors work
Detectors look for the statistical fingerprints of machine-generated text. The two most common signals are perplexity — how surprising or predictable each word is to a language model — and burstiness — how much sentence length and complexity vary across a passage. Human writing tends to be bursty and less predictable; AI writing, especially default-temperature output, tends to be smoother and more uniform. A detector scores text against these patterns and outputs something like “85% likely AI.” The crucial point is that these are correlations, not proof: plenty of human writing is smooth and predictable, and plenty of AI writing can be made bursty.
The false-positive problem
The reason detectors are dangerous in practice is the false-positive rate — human writing flagged as AI. A widely reported Stanford study found that detectors flagged a large fraction of essays written by non-native English speakers as AI-generated, because simpler vocabulary and structure mimic the low-perplexity signal detectors hunt for. Formulaic but entirely human writing — technical documentation, legal boilerplate, a careful student following a rubric — can also trip the alarm. When the cost of a false positive is a failing grade or a misconduct accusation, even a small error rate is unacceptable as standalone evidence.
The tools and their limits
The best-known detectors — GPTZero, Turnitin’s AI indicator, Copyleaks, and Winston AI — all publish high accuracy figures on their own benchmarks, but independent testing consistently lands lower and inconsistent across writing styles. Worse, every detector can be defeated: paraphrasing tools, manual editing, and purpose-built “humanizer” services reliably lower scores, and the evasion side of the arms race moves faster than the detection side. So a “clean” result proves nothing (the text may be edited AI), and a “flagged” result proves nothing (it may be honest human writing).
What to do instead
Because tool-based detection cannot be trusted on its own, the workable approach is process and conversation, not software. Educators get far more from asking students to show drafts, version history, and notes, or from a short oral discussion of the work, than from any detector score. Organisations are better served by clear AI-use policies and disclosure norms than by surveillance tooling. If a detector is used at all, treat its output strictly as a prompt to look closer — never as the basis for a decision. The defensible position in 2024 is simple: AI detectors are a hint, the human conversation is the evidence.