Detectors guess.
Watermarks know.

The short answer

Sometimes — statistically, on average, over many documents. On any single passage, no. AI detectors misclassify often enough, in both directions, that the most telling data point in the field is this: OpenAI, with every advantage in understanding its own models, withdrew its AI text classifier in July 2023 citing its low rate of accuracy. If the model’s maker could not do it reliably, third-party tools claiming near-certainty deserve skepticism.

How detectors work — and why that is the problem

Post-hoc detectors have no inside information. They estimate AI likelihood from learned features of the writing itself — predictability of word sequences, sentence-length variation, distributional patterns. That makes them universally applicable and universally fragile: they are guessing what the text looks like, not testing what produced it.

This is the fundamental contrast with keyed watermarks. A watermark detector re-scores every word choice against a secret key and computes calibrated statistical evidence; a style classifier pattern-matches against what model output usually looks like. One is a measurement, the other an impression. Our detection guide covers where each fits.

False positives: the expensive direction

A Stanford study found several commercial detectors flagged more than half of writing samples by non-native English speakers as AI-generated, while classifying the same authors’ ideas rewritten in more elaborate English as human. Formulaic, plain, well-organised prose — the kind produced by careful students, technical writers, and anyone writing in a second language — is exactly what detectors confuse with model output.

The cost lands on people: students accused of cheating, freelancers losing clients, writers asked to prove a negative. Any process that treats a detector score as a verdict rather than a prompt for human judgment is misusing the tool.

False negatives: the quiet direction

Errors run the other way too. Paraphrasing, light editing, and translation shift the statistical fingerprint detectors rely on — the SynthID-Text paper reports that edits weaken even purpose-built watermarks, and style classifiers have far less to hold onto. Detectors also trail the models: each new generation writes differently from the corpus the classifier trained on. A clean score does not establish that a human wrote the text.

Using detector scores responsibly

• Treat a score as a reason to look closer, never a conclusion. Corroborate with drafts, version history, and the author’s other writing.
• Never act on a single passage. Statistical evidence needs length; short texts are noise.
• Weigh the false-positive cost before the false-negative cost. A wrong accusation usually does more harm than a missed detection.
• Prefer provider detection where it exists. Keyed watermark checks, like Anthropic’s forthcoming API, produce calibrated evidence a style guess cannot match.