It copied without the quote marks

Overview

Why now. You paste a report into an assistant and ask for a summary. What comes back reads like clean prose. You check the figures, tighten a sentence, drop it into a client note, and it goes out under your name. Three of those sentences are the source’s own words, run together exactly as the author wrote them. Nothing on the screen told you which three.

On 1 September 2026, Anthropic published its list of ways a newly released model behaves differently from the one before it. Sixth in a list of seven behaviour changes sits this: “When summarizing documents, the model is more likely to reproduce passages of the source without marking them as quotations.”

Read that as what it is — a vendor writing down an attribution failure in its own product, on a page for developers that almost nobody else opens. Read it precisely, too: more likely than the previous model. No rate is published, and none is implied.

The habit that follows is small. Before an AI summary of a document goes into your own work, search two of its phrases in the source. Anything that comes back word-for-word needs quote marks and a citation — because the model did not add them.

The content

This is not the hallucination problem, and confusing the two is why the check gets skipped. Every fact in that summary can be correct. You verified them. The failure sits in a different dimension: not is this true, but whose words are these — and nobody told you that was a dimension.

Call the thing you are looking for the unmarked span: a run of the source’s own consecutive words, arriving inside your paraphrase with nothing around it.

It is not one vendor’s quirk. A study published at the end of December 2025 built a corpus of 35,626 abstract-and-summary pairs from biomedical papers, where the summaries had been written by the researchers themselves, then put five open-weight models — none of them Anthropic’s — at the same task. Measuring how much of each summary’s vocabulary did not appear in the source, the human-written summaries came in at 46.13% novel words against a model range of 33.48% to 39.85%. The authors’ conclusion is that models lean on the source’s own lexical choices and rhetorical structures more than people do: more extractive, less abstractive.

Now the limits, because they change what you actually check. Overlap rates on their own prove nothing. Any faithful summary shares most of its words with its source, and a lower novel-vocabulary score is not evidence of a single copied sentence. What matters is contiguous multi-word runs — the only thing you can practically search for anyway.

And the gap is not universal. Earlier work on news summarisation found the direction reversing depending on which system you compare against and how that system was trained — though that comparison runs model against model, not model against human. So “AI copies more than people” is a tendency with conditions attached, not a law. What is unarguable is the half with your name on it: when a span does come across, nothing marks it.

The consequence is documented, with a person attached. In March 2026 a New York Times contributor lost the relationship over a near neighbour of this. In his own words: “I made a serious mistake in using an AI tool on a draft review I had written, and I failed to identify and remove overlapping language from another review that the AI dropped in.”

Nothing in that sentence describes carelessness about facts. He was careless about a category of error nobody had told him to watch for, and no interface offered to show him. Note the variant, because it matters below: the language came from a different review of the same book, not from a document he had handed over.

Which is the argument for a habit rather than a policy. The vendor’s own page does point to a prompting fix — a worked example of a correctly attributed summary, added to your standing instructions. Worth doing. But an instruction is not a check. You cannot see whether it held without looking, and looking takes two minutes.

Try it

Take the next AI summary you are about to use — of a document you also have open.

Pick two phrases from it, six to ten words each. Choose the ones that sound most like finished writing: a technical phrase, a clause that reads better than the material around it. Those are the likeliest borrowings, because they were the good sentences in the source too.

Search each one in the original. Ctrl+F, or your reader’s find function.

• Exact hit — it gets quotation marks and a citation, or you rewrite it in your own words. Not deciding is the only wrong answer.
• Near hit (the same clause with a word changed) — rewrite it. Reworded borrowing is still borrowing, and it will not survive anyone else running the same search.
• No hit — you have spent two minutes and you can stop.

If the source reached the assistant through a connector your workspace has enabled rather than a paste, open the document itself and search there. If your workspace offers saved prompts or instructions, add that worked example to the one you use for summarising — and keep doing the search anyway.

Where it breaks: two ways. It searches only the source you have — if the assistant pulled something in that you never saw, as happened above, there is nothing to search against. And it finds the source’s words in your prose, not the opposite failure: quotation marks around a sentence the source never wrote. That one you check in the other direction.

Additional reading

• What’s new in Claude Fable 5.1 — Anthropic, 1 September 2026. The changed-behaviour list. Worth knowing such pages exist for whatever you use.
• Not too long do read: Evaluating LLM-generated extreme scientific summaries — Lyu & Ke, 29 December 2025. The novel-n-gram table, and the dataset construction: researcher-written one-line summaries harvested from bibliography comments.
• AI journalism mistakes tracker — Press Gazette, continuously updated. The Preston entry is from March 2026. Worth scanning the rest for how often the failure is attribution rather than invention.
• News Summarization and Evaluation in the Era of GPT-3 — Goyal, Li & Durrett, September 2022. The counter-evidence, in full. Note that its comparison is between models, not between models and people.

Editor’s note

There’s a micro-commentary in here somewhere about training data that I’ll leave alone. The point to remember is that you can’t be sure about whether appropriate attribution has been made when a summary of a piece of text is delivered to you. There will be some tasks where this is problematic, and for those tasks, this is something to keep in mind.