AI Detectors Are Answering a Question No Platform Asked

There is a whole workflow that has grown up around a false premise: write with AI, run it through a detector, rewrite until the score drops, publish. Hours per week, sometimes a subscription, spent optimising against a number that no platform in your distribution chain ever looks at.

Two separate problems here. The detectors are unreliable, and even a perfect detector would be measuring the wrong thing.

Problem one: the accuracy claims do not survive contact with independent testing

The research picture, with sources:

• A 2026 paper in the academic-integrity literature argues that detector findings are "prone to generate more false accusations than correct identifications," reporting a positive detection rate as low as 26% in its analysis.
• An academic library guide summarising multiple studies states detectors were "neither accurate nor reliable," producing high numbers of both false positives and false negatives.
• Vendor and independent figures diverge widely. One major vendor publishes accuracy claims above 98% with under 1% false positives; independent write-ups report accuracy falling to roughly 60-80% once text has been manually edited or reworded. Another vendor claims a false positive rate of 0.004%.

Any tool whose published accuracy ranges from "1 in 25,000 errors" to "more wrong than right" depending on who measured it is not a measurement instrument. It is a claim.

The specific failure that matters is the false positive: human writing flagged as machine-written. If the tool is meant to reassure you, a tool that flags real writing as fake cannot do that job.

The limitation nobody states: this is all about essays

Be precise, because this cuts both ways. Nearly all the research above studies academic writing — student essays, several hundred to several thousand words, formal register, a corpus detectors were specifically built and tuned for.

Social copy is nothing like that. A 40-word caption, a hook line, a thread post. Shorter text carries less signal; genre conventions dominate; the "voice" a detector keys on barely has room to appear.

So the correct statement is not "detectors are 26% accurate on your captions." It is: the evidence base for detector reliability on short social copy essentially does not exist, and what evidence exists in the adjacent domain is unflattering. Using a tool with no validation in your domain to make decisions in your domain is the actual problem.

Problem two: nobody in your distribution chain is running one

This is the part that should end the workflow. Look at what platforms actually published in mid-2026:

Three of the four describe form. LinkedIn's is the most explicit: the human annotators labelled posts generic or original — not human or AI. That single design decision tells you everything about what is being optimised for.

Which means an AI detector, even a perfect one, is measuring an attribute nobody is grading. You can pass a detector with an entirely generic post. You can fail one with the most original thing you have ever written. Neither result predicts your distribution.

What to optimise instead

Take the platforms' own language and turn it into a checklist. Before publishing, ask whether the piece contains at least one of:

1. A judgement. Not a summary — a position, with the reason attached. This is the direct opposite of "little commentary."
2. First-hand material. Something you tested, measured, filmed, were told, or got wrong. Unshareable input is the only durable defence against "generic."
3. A named situation. Written for a specific person in a specific circumstance, not for "people interested in the topic."
4. A structure that is not the current genre template. If your format is the format everyone uses this quarter, similarity is doing the damage, not authorship.

None of these require you to stop using AI. They require you to stop using AI on the input side. The full argument is in why AI content sounds generic, and the review workflow that makes this checkable at volume is in reviewing AI content at scale.

The cost of the workflow nobody counts

Suppose the detectors were perfectly accurate. The detect-and-rewrite loop would still be a bad use of time, and it is worth being concrete about why.

Every rewrite pass aimed at lowering a score is a pass that optimises for statistical unremarkability. You break up rhythmic sentences, swap precise words for looser ones, add hedges and asides. The output reads more human by the detector's definition and less useful by the reader's — because the properties that make text score as machine-written overlap heavily with the properties that make it clear.

Meanwhile the edit that would actually change your distribution is a completely different operation: adding a judgement, adding a fact only you have, cutting the paragraph that says nothing. That edit takes the same twenty minutes and moves the metric that matters.

So the real cost is not the subscription. It is that the detector loop occupies the editorial slot — the one pass a human makes over the draft — and spends it on the wrong objective. One pass, two possible uses, and only one of them is graded.

When a detector is still worth running

Not never. Two legitimate cases:

• A client or employer contractually requires a score. Then it is a compliance artefact, not a quality signal, and you should treat it as such — including flagging the false positive risk to whoever set the requirement.
• As a crude sameness check across a batch. If everything you produce scores identically, that is weak evidence your inputs are too uniform. Weak evidence, low cost — but a similarity check against your own back catalogue does the same job more honestly.

What it is not worth doing: rewriting good work until an unvalidated tool feels better about it. That is time taken from the one thing that does move distribution — the editorial pass where a human adds something.

FAQ

Could platforms be running detectors internally without saying so?

Possible, but the published language points elsewhere. LinkedIn described training a model on posts human editors labelled generic versus original — that is a quality classifier, not an origin detector. TikTok's approach to provenance runs through Content Credentials and invisible watermarking, which identify AI-generated media by embedded signals rather than by statistical text analysis. Neither resembles a consumer AI-text detector.

My post got flagged by a detector. Should I rewrite it?

Ask what the flag would change. If no platform in your chain uses that score, the flag has no consequence. Better use of the same twenty minutes: check the post against the four-item list above. If it fails those, rewrite — for that reason, not for the score.

Does declaring AI use hurt my reach?

Disclosure is increasingly required rather than optional, particularly in commerce contexts, and platforms describe penalties for failing to label rather than for labelling. The genuinely new variable is on the audience side: TikTok is testing a control that lets viewers choose how much AI-generated content they see, which we covered in TikTok's account-level AI spam detection. Label honestly — then make the content good enough that the label is not the deciding factor.