How Is AI Text Recognised? Detectors and Reality

The teacher drops the student's essay into a detector: "87% AI." The student swears they wrote it themselves. Who is right? The uncomfortable answer: unknown; and that uncertainty is the detectors' nature, not their exception. The percentage figure's decisiveness is far larger than the technology's.

The need to check AI text is real: education, editorial desks, hiring. This article explains how the detectors work, why they err, and how to use these tools (if you will) responsibly; for both sides: the checker and the checked.

How do the detectors work?

The core approach is statistical traces: AI texts are on average "smoother"; predictable word choices (low perplexity), even sentence lengths (low burstiness), typical transition structures. The detector scores the text against these patterns and gives a probability. And the problem starts right there: these are style signals, not proof — and style "deceives" in both directions: a tidy human writer resembles the machine, an edited machine text resembles the human.

Why they are unreliable: four structural problems

• False positives: human text getting labelled "AI"; especially high for non-native writers and in formal-academic style. The heaviest damage sits here too: the innocent accusation.
• Easy evasion: light editing of AI text (synonyms, sentence shuffling, "humanizer" tools) drops the percentage sharply; meaning the deliberate user you most want to catch is the one caught worst.
• The model race: text models evolve toward writing "like a human"; the detector's target is a perpetually fleeing target.
• The mixed-text reality: today's normal workflow (a human draft + AI editing, or the reverse) has itself aged the binary "AI/human" question; hybrid text sits in the spectrum's middle.

For these reasons the big institutions are cautious: OpenAI shut down its own text classifier over low accuracy; a portion of educational institutions ban the detector result as final proof.

For the checking side: the responsible protocol

The detector is not a tool to discard entirely; its correct place is the "signal, not proof" box. The working protocol: never make the detector percentage the sole basis; check the second signals (a sharp style break from the person's earlier work, a mismatch between subject knowledge and the text's depth, fabricated references; these are more reliable markers); verify by conversation (3–4 deep questions about the text; being unable to explain one's own writing is a stronger indicator than the percentage) and build the process in advance (requiring drafts, version history; it shifts the burden of proof from the later dispute to prior transparency). Procedural language instead of accusatory language: not "the detector said so" but "let's discuss the text together."

For the writing side: protection from an unjust accusation

• Document your process: drafts, notes, version history (Docs' history is natural evidence).
• Declare your AI use per the rules: not hiding it where it is permitted is the strongest defence.
• When facing an accusation: a calm tone + the process evidence + an offer of an oral discussion over the text; support the "the detector errs" argument with the facts above (the official reliability admissions).

The real question: quality, not the source

In the practical world (educational assessment aside) the right question has changed: not "did AI write it?" but "does the text do its job?": are the facts true, is there a thought, what is the reader's benefit. Google's search rules run on the same line: the value is assessed, not the source. The workable standard for editorial desks and business: AI help is not forbidden; unchecked facts, fabricated references and unedited publishing are. That standard is both honest and checkable; unlike the detector percentage.

Sources and further reading

Where to verify the source

The primary sources on detector reliability:

• OpenAI — the note on the shut-down classifier
• Google — the approach to AI content

Continuing the topic

This topic's neighbouring articles:

• AI hallucination: the anatomy of invention
• The hybrid content flow
• The AI policy in a company
• AI basics
• Other articles on this topic

The closing rule is the same for both sides: the percentage figure is not a verdict. If you are the checker, look at the process; if you are the writer, keep the process. The dispute's winner is always the documented side.