How to detect AI-generated text, and why most detectors fail

Can you tell whether text was written by AI? Detection tools say yes with a confidence percentage attached. The honest answer is that most of them can’t, and the ones that can are about to be replaced by something quite different.

This matters because people are being accused of things on the strength of those percentages.

How detectors try to detect AI generated text

The familiar tools work statistically. They look at how predictable the text is, whether word choices are the ones a model would most likely pick, and how much variation there is in sentence structure.

Machine-generated text tends to sit closer to the statistical centre, because models are trained to produce likely continuations. Human writing wanders more.

So a detector is really measuring one thing: how ordinary the writing is. That isn’t the same as measuring who wrote it, and the gap between those two things is where the damage happens.

The training process makes this worse rather than better. Models are tuned on human preference, and preference rewards fluent, conventional prose, as OpenAI’s InstructGPT work showed when a 1.3B tuned model was preferred over 175B GPT-3.

In other words, models were deliberately pushed toward the kind of writing detectors treat as suspicious, and good human writers are often pushed the same way by editors.

A detector flags text for being predictable. Plenty of people write predictably, and none of them deserve an accusation for it.

Who gets falsely flagged

The false positives are not random, which is the part that should worry anyone using these tools on people.

Writer	Why they get flagged
Non-native English speakers	Simpler constructions, smaller vocabulary range, closer to the statistical centre
Technical and academic writers	Conventional structure is required by the format
Anyone writing to a template	Reports and forms constrain word choice by design
Careful, plain writers	Clarity looks like predictability to a classifier

Read that list again. It’s close to a list of people already disadvantaged in academic and professional settings, and the tool compounds it.

There’s no threshold that fixes this. Raise it and you miss most AI text. Lower it and you accuse more people who wrote their own work.

Watermarking is a different mechanism entirely

The approach now being deployed doesn’t guess. It marks the text at the moment of generation.

Anthropic began watermarking every Claude model released after 2 August 2026, embedding a signal into the model’s token selection rather than into file metadata. TechCrunch reported it travels with the text when copied and pasted.

Files use the separate C2PA signed-metadata standard, while text carries the embedded statistical mark, as analysis of the rollout set out. The distinction matters, because metadata is trivially stripped and an embedded mark is not.

This is a real improvement over guessing. It’s also narrower than it sounds.

What a watermark proves, and what it does not

A watermark tells you the text passed through a particular model. It doesn’t tell you who wrote the ideas, how much a person edited afterwards, or whether any rule was broken.

TechTimes put the distinction precisely: the mark proves processing, not authorship.

That gap is exactly what the backlash was about. Forbes reported complaints from people who use models only to proofread their own writing, where a mark cannot distinguish editing from generation.

Coverage also noted the signal degrades under heavy editing, paraphrasing or translation, and that short passages may not carry enough text to detect reliably, as Gizmodo reported.

The coverage gap nobody solves

Watermarking only works if the model doing the generating chose to mark its output, and a large share of models will not.

Open weights are the obvious hole. Meta released a 30B model under Apache 2.0 that anyone can download and run locally, and nothing’s compelling a self-hosted deployment to mark anything.

So the systems being marked are the compliant commercial ones, and the systems most likely to be used for deception are the ones outside that perimeter. Detection improves for honest users and barely moves for dishonest ones.

Why the rules exist at all

None of this deployment happened because users asked. The EU AI Act’s Transparency Code took effect on 2 August and requires generated or edited content to be marked so other systems can identify it.

The requirement is about provenance in the information ecosystem, not about catching students. Our piece on what the Act requires covers the scope.

That mismatch is the core problem. A rule written to label machine content is being repurposed as evidence in decisions about individuals, and it wasn’t designed to carry that weight.

What to do instead of trying to detect AI generated text

If you run a classroom or a team, the useful shift is from detection to process.

Ask for drafts, outlines and version history, which aren’t easy to fabricate convincingly and reveal how work developed. Assess in conditions you control when the stakes justify it. Set an explicit policy on what assistance is permitted rather than an unstated assumption that none is.

And if a detector flags something, treat it as a prompt to have a conversation rather than as a finding. A percentage is not evidence, and the person on the other side usually can’t prove a negative.

The counter-case

There’s a reasonable argument that some detection beats none, and it deserves stating rather than dismissing.

At scale, a detector that is right most of the time still surfaces cases worth reviewing, and reviewing isn’t the same as punishing. Used as a triage signal with a human decision behind it, the tools have a defensible role.

The objection isn’t that detectors are useless. It’s that institutions rarely stop at triage, and the false positives fall on people least able to contest them.

What would change this

Universal watermarking across open and closed models would close the coverage gap, and there’s no mechanism to enforce that on weights already downloaded.

Published false-positive rates broken down by writer demographics would let institutions price the harm they are accepting. No major vendor publishes that.

Until either arrives, treat every confidence score as what it is: a statement about how ordinary the writing looks, dressed up as a statement about who wrote it.