AI Content Detection Explained: What Detectors Actually Measure

This topic is unusually badly served, because the people writing about it are mostly selling detection, teaching people to evade detection, or reassuring readers that Google does not care. All three skip the same thing: what these tools actually measure.

Once you know that, everything else — the false positives, the evasion advice, the search-ranking question — resolves cleanly.

What detectors actually measure

Two statistical properties, under various names.

• Perplexity — how surprising the word choices are. Language models select high-probability continuations, so their output has low perplexity. Predictable text.
• Burstiness — how much sentence length and complexity vary. Humans write a long winding sentence, then a short one. Models produce more uniform rhythm.

A detector scores text on these and outputs a probability. That is the whole mechanism. There is no watermark, no signature, no hidden marker. It is reading style and inferring origin.

How accurate are they?

The most informative data point is that OpenAI shut down its own AI Text Classifier in July 2023, citing low accuracy. The organisation with the deepest knowledge of how its models generate text concluded it could not reliably detect that text.

Commercial detectors advertise high accuracy figures. Those come from vendor testing on their own datasets, typically unedited AI output versus clearly human writing — the easy case. Real-world text is edited, mixed, and written by people with varied styles, and accuracy degrades sharply there.

Text type	Detector reliability	Why
Raw AI output, unedited	Reasonably good	Low perplexity, uniform rhythm. The easy case
AI output, lightly edited by a human	Poor	Editing raises perplexity quickly
Human writing, plain and clear	Poor — false positives	Clarity looks like predictability
Non-native English speakers	Poor — false positives	Formal learned constructions score as predictable
Technical or legal writing	Poor — false positives	Precision and consistency are the point
Short passages under 300 words	Unreliable either way	Not enough text for the statistics to mean anything
Human-AI collaborative writing	Meaningless	The question has no binary answer

Does Google penalise AI content?

The question most publishers actually care about, and the answer is clearer than the discourse suggests.

Google’s published position is that it rewards helpful, reliable, people-first content regardless of how it is produced, and that using automation to manipulate rankings violates its spam policies. Those two statements are consistent: the target is low-value content produced at scale to game search, not the use of a writing tool.

What actually gets penalised is thin, unoriginal, unhelpful content — which AI makes cheap to produce in volume, which is why the two get conflated. The causal factor is the thinness, not the tool.

• Publishing 400 generated pages targeting keyword variations is the problem behaviour, and it was the problem behaviour when people did it by hand.
• Using AI to draft an article you then research, verify, edit and add genuine expertise to is not distinguishable from good writing, because it is good writing.
• The test is whether the page is worth reading, which is the same test it has always been.

Nor does Google run a detector in the way people imagine. Given that purpose-built detectors are unreliable, a search engine assessing billions of pages would be poorly served by a signal with that error rate. Quality signals are more robust and harder to game.

If you have been falsely accused

A practical section, because this happens frequently and the advice available is poor.

1. Ask which tool was used and what score it gave. Specific numbers, in writing. Vague accusations should not stand.
2. Produce your process evidence. Draft history in Google Docs or Word version history, notes, outlines, browser research history, timestamps. This is the strongest defence available and it only exists if you worked in a tool that records it.
3. Run known-human text through the same detector. Something you wrote years before these tools existed, or a passage from a published book. Flagged results undermine the tool’s authority efficiently.
4. Cite OpenAI’s withdrawal of its own classifier. It is the single most persuasive fact available and it comes from the most credible possible source.
5. Ask for the false positive rate the institution assumes, and how many students that implies are wrongly accused each year. Most have never calculated it.
6. Offer to discuss the content. Someone who wrote a piece can talk about the choices in it. This is better evidence than any score, in either direction.

Point two is worth acting on before you need it. If you write anything that could be challenged, work somewhere with version history. A draft that evolved over six hours across forty revisions is close to unfakeable.

What to do instead, if you run a team

Detection is the wrong lever. If your concern is quality and honesty, there are better ones.

• Write a disclosure policy rather than a ban. Bans drive use underground; disclosure gets you visibility and a conversation about appropriate use.
• Review for accuracy, not origin. Check citations resolve, figures are correct, claims are supported. This catches the actual risk, which is wrongness rather than authorship.
• Require something the model cannot have. First-hand experience, original data, a named source, a specific example from your work. Unfakeable and it improves the output.
• Discuss the work. Five minutes talking about a piece reveals more than any score.

My verdict

AI detectors measure predictability and infer authorship, which is a reasonable heuristic and an unreasonable basis for a decision about a person. They work on the case nobody needs help with, and fail on edited text, plain human writing, and non-native speakers — with the failures landing disproportionately on people who cannot push back.

For publishers the question is a distraction. Google does not penalise AI assistance; it penalises unhelpful content, which it has always done. Write something worth reading, verify what you claim, add what only you know, and the tool you drafted with stops being relevant.

And keep your version history. It is the only evidence that actually works.

For why AI text has these statistical properties, see what ChatGPT is and how it works. For producing content that holds up on its merits, the best AI content creation tools covers the workflow side. The same detection argument now surfaces in recruitment, where candidates write applications with AI and employers reach for screening tools in response, which is covered in best AI tools for HR.

Frequently asked questions