AI content detection

Does AI-written content get used by AI answers?

Yes, when it is worth quoting, and the engines say as much themselves.

The answer at the top of this page is ChatGPT replying, on 3 October 2026, to a question our workspace tracks: "AI search optimization for developer tool companies". It never asks who wrote a page. It passes on Google's position that unique, expert-led content beats commodity AI-generated content, and the one page it cited inline was Google's own developer documentation. That question got 12 answers from ChatGPT and Perplexity in the 30 days to 4 October 2026, read over the Attensira MCP; none of them named us, which is a reminder that writing about a topic is not the same as being the source for it.

This glossary is drafted with a language model and reviewed by the named author before it is published; this paragraph is where we say so. What a reader can check on each page is the author's name, the date and link of every source, and the claims themselves at /claims.json. Those are the signals that decide whether a passage is quotable. Production method is not the lever; usefulness is.

What counts as AI content detection?

Two families of technique share the name, and only one of them detects anything.

Post-hoc classifiers read the finished artefact and output a likelihood. For text, the features are statistical: how predictable each word is given the ones before it, how little that predictability varies, how narrow the vocabulary is. Machine text tends to sit closer to the most probable path than human text wanders.

Provenance and watermarking do not detect; they label at the moment of creation. Provenance signing attaches signed metadata to an asset as it is made or edited. Watermarking embeds a signal in the output itself, which only exists if the generator chose to add one.

Why do classifiers flag human writing?

Because they measure typicality, not authorship. Anything that makes human writing more typical pushes it toward the machine class:

• edited house-style prose, such as press releases, legal boilerplate and exam answers;
• writing by a non-native speaker drawing on a smaller set of constructions;
• short samples, where there is not enough text for the patterns to separate.

The reverse holds too. A machine draft that a person reorders and roughens reads as human again. The two error types trade against each other on one threshold, so a detector tuned to catch more machine text accuses more people. There is no setting where both errors are small, which is why a classifier score cannot carry an accusation against a student, an applicant or a contributor on its own.

What does provenance actually prove?

Less than people hope, and exactly what it claims. The C2PA specification, version 2.2 from May 2025, binds assertions about how an asset was made to a claim signature in a tamper-evident manifest. Its scope is stated plainly: the specification does not judge whether provenance data is good or bad, only whether the assertions validate, are correctly formed and were not tampered with.

So a valid manifest tells you what a signer claimed and that nobody changed the claim. It does not tell you the claim is true. And stripped or missing metadata is the normal state of web content, not a warning sign. Watermarks share the same asymmetry: a present, valid signal is strong evidence; an absent one is no evidence at all.

Does Google demote AI-generated pages?

Not for how they were made. Google's spam policies define scaled content abuse as many pages generated mainly to manipulate rankings rather than help people, and name generative AI tools producing pages without added value as one way it happens. The violation is volume without value, whatever wrote it.

The answer quoted at the top of this page carries the same message from the other side: the engine relays Google's preference for expert-led pages and cites Google's own guidance for it. The pages that get attached are the ones carrying a fact worth extracting: a definition, a dated number with its source, a method. A detector run over the prose would change none of that.

What are the objections to AI content detection?

"We should rewrite our pages so detectors pass them." That optimises for statistical oddness, which has nothing to do with whether a passage answers a question well. It usually makes the passage worse to quote.

"No watermark means a human wrote it." Only cooperating generators embed one. Absence carries no information.

"Generation is cheap now, so more pages means more chances to be cited." That is the behaviour the scaled-content policy names, and retrieval picks passages by relevance to one question, not by how many pages a domain has. See AI-first content strategy.

"Disclosing AI help will cost us trust." Trust comes from what can be checked. A named author, a review date and sources a reader can open do more than any label. See content authority.

How does Attensira handle AI-written work?

Attensira's crew writes with models, so this is our own question too. Two rules keep the output usable. Every draft is held to a house rule called Quote only what was read: no quotation, testimonial, statistic, study or customer story without a named source a reader could check, and no claim about what a page, a review or an AI answer says unless the draft quotes or cites it. And workspaces start in Draft: anything that states a claim, such as a price, a number or a competitor name, waits for your yes in the inbox in every mode, so a person sees each claim before it ships.

Then the work is measured where it matters: whether the engines name and cite you afterwards, per engine, with the sample size. Citation tracking shows which of your pages each answer linked, so a rewrite is judged by the answers that follow it, not by a detector score.