Detectors and Humanisers

An industry has grown up around detecting AI-generated text, and another around defeating the detectors. Both rest on a premise that does not hold.

Why detection does not work reliably

There is no signal to detect. Generated text is not watermarked. It is statistically ordinary text. Detectors measure properties like predictability and uniformity, and human writing varies enormously on exactly those properties.

False positives fall on specific people. Writing that is plain, structured and grammatically regular scores as machine-generated. That describes non-native English writers, people writing in a formal register, technical writers, and anyone taught to write in clear simple sentences.

This is not a hypothetical harm. Students have been accused on the basis of detector output, and the pattern of who gets accused is not random.

Published accuracy figures are measured on clean cases — fully generated text versus fully human text. Real cases are mixed: human writing edited by a model, generated text heavily revised, a draft with pasted quotations.

Light editing defeats them. A pass of rewriting drops detection scores substantially.

Nothing can be proven. A detector produces a probability. It is not evidence in the sense a disciplinary process requires, and institutions that treat it as evidence are making a mistake that will eventually cost them.

Why humanisers make things worse

Tools that "humanise" text work by introducing variation — swapping words for less common synonyms, varying sentence length, adding irregularity.

The result is usually worse writing. Odd word choices, clumsy rhythm, sentences that vary for no reason. It reads as strange rather than as human.

It defeats a probability estimate, not a reader. An editor or a teacher who suspects generated text is responding to the qualities described in the previous article — genericness, symmetry, absent specifics. A humaniser does not fix any of those. It changes vocabulary while leaving the emptiness intact.

It introduces errors. Synonym substitution changes meaning, and technical terms get replaced with near-misses that are wrong.

It defeats the purpose. If the aim is text that reads as yours, the way to get it is to write parts of it.

If you are being assessed

Keep drafts and version history. A document with an edit history showing composition over time is far better evidence than any argument about a detector.

Ask what the policy actually is before submitting. Many institutions permit assisted drafting with declaration and prohibit undeclared generation. Knowing which applies removes the guesswork.

If accused, ask for the basis. A detector score is not proof. Ask what the tool's false positive rate is on writing like yours, and whether the institution has validated it.

If you are assessing

Do not use a detector as evidence. Use it as a reason to look more carefully, if at all.

Look at the content instead. Generated work has recognisable weaknesses: no specific engagement with course material, generic examples, citations that do not exist or do not say what is claimed, no evidence of the thinking the assignment required.

Check the citations. This is the single most effective check and it takes minutes. Fabricated references are common and unambiguous.

Ask about the work. A short conversation about what someone wrote establishes authorship better than any tool.

Design assessments that are harder to generate. Specific to the course, requiring engagement with particular material, building on earlier work, or involving a spoken component.

What is actually changing

Watermarking of generated text is technically possible and has been demonstrated, and it only works if the generating provider applies it and the text is not heavily edited. It will not cover open-weight models that anyone can run.

Reliable detection of AI text is not arriving. Institutions and employers will have to adapt by changing what they assess and how, rather than by policing how text was produced.

The detector industry is selling certainty that does not exist, to people under pressure to have it. That is worth understanding before anyone's academic record depends on a percentage.

If your institution uses one

Practical steps for anyone studying or working somewhere that runs detection, whether or not the policy is reasonable.

Write in something with version history. Cloud documents, or a repository. Composition over time is the strongest evidence of authorship available and it costs nothing to have.

Keep your notes and sources. Being able to produce the material you worked from settles most questions.

Do not paste your draft into a detector to check it. Several of these services retain submissions, and some have been found to add them to their own corpora.

Ask the policy question in writing before you need the answer. A recorded reply about what is permitted protects you.

If accused, ask three things: what tool was used, what its documented false positive rate is, and what evidence exists beyond the score. Most institutions cannot answer the second, and the question is reasonable rather than obstructive.

Ask for a conversation about the work. If you wrote it, you can discuss it, and that is more convincing than any technical argument.