The short version
Parts of the Wire are written by a machine. We think you are entitled to know which parts, what a human does about it, and how it fails, rather than a line at the bottom of the page saying "AI-assisted" and leaving you to guess.
1. What is machine-generated
| Part of the Wire | Machine's role | Human's role |
|---|---|---|
| Story selection and clustering | Reads incoming reports, judges relevance, groups reports about the same event into one story | Sets the rules and the taste model; reviews outcomes; kills bad clusters |
| Story summaries | Writes the summary you read under a headline, from the source article | Reviews; the headline and the link are the publisher's own |
| The desk memo | Written end to end by a model from the day's stories, rewritten through the day | Reviewed by the publisher; the format, the constraints and the prohibitions are human-set |
| Deal extraction | Reads reports and pulls out parties, price, structure, territory, rights and status into a ledger row | Reviews rows; adjudicates disputed values; every claim keeps the sentence it came from |
| Entity matching | Matches company names and aliases in text to the register | Human-curated gazetteer; exact matching only, because fuzzy matching puts one company's facts on another's record |
| Company one-liners on cards | Drafted from register and encyclopedia data | Reviewed |
2. What is not machine-generated
- Headlines of third-party stories. Those are the publisher's own words, reproduced.
- Quotations. Reproduced, not paraphrased.
- Figures in a filing. Taken from the filing.
- Market data. Supplied by a data source.
- The register's core company records. Compiled from public registries.
- The editorial rules, the taste model, and this policy.
3. What we do about the fact that models are wrong sometimes
3.1 Every machine-written claim is tied to a source. A ledger row keeps, per claim, the report and the sentence the claim was taken from. This is the single most important control we have: it means a hallucinated number has nowhere to hide, because the sentence either says it or it does not.
3.2 The model is not allowed to be the only witness. A deal reaches the ledger on the strength of the reporting behind it, not on the model's confidence. Where reports disagree, the row is flagged as disputed and shows the disagreement.
3.3 Human review before publication. Nothing a reader or a third party submits publishes automatically. Machine-generated editorial output is reviewed by the publisher before and after it ships, and the desk memo is read every time it is rewritten.
3.4 Corrections are structural, not cosmetic. Where a published figure changes, the row records what it said before. See Editorial Standards and Corrections.
4. Where it still fails, honestly
We would rather tell you the failure modes than let you find them.
- A summary can be fluent and wrong. Machine-written prose does not signal its own uncertainty. A summary that misreads a source reads exactly like one that did not.
- A deal that ends can be invisible. The extraction is built around transactions happening. An abandoned deal does not always parse as an event, so the absence of a row is not evidence that nothing happened.
- Clustering can merge two events or split one. Two lawsuits about the same matter can become two stories; two reports of different matters can become one.
- Entity matching can miss. A company with an unusual name, or one newly formed, may not be tagged even though it is in the register.
- The register's coverage is uneven. Public companies are well covered; a private producer in a small market may have a thin record or none.
None of this is a reason to distrust the Wire. All of it is a reason to click the source before you act on a number.
5. Which systems
Large language models supplied by Anthropic PBC are used for summarisation, clustering, extraction and the desk memo. Development of the Wire itself was also AI-assisted.
Anthropic is not affiliated with, and does not endorse or sponsor, the Wire.
Your data is not sent to them. The models operate on published source material — articles, filings, registry records. Subscriber account data, billing data and taste profiles are not sent to any model provider.
6. Your content is not training data
We do not use anything you submit — a correction, a bug report, an enquiry — to train a model, and we do not permit our providers to train on it.
Equally, you may not use the Wire's content to train a model: see clause 3.4 of the Acceptable Use Policy and our express reservation of rights against text and data mining in clause 11.3 of the Terms.
7. Automated decisions about you
The Wire ranks and orders stories for you automatically, based on a taste profile that lives in your own browser. That affects the order in which you read things and nothing else. We make no automated decision that produces legal effects for you or similarly significantly affects you, within the meaning of Article 22 of the UK GDPR.
8. Regulatory position
[CONFIRM: the position stated in this section.] We consider the Wire to be a deployer, not a provider, of general-purpose AI systems, and its use to be limited-risk under the EU AI Act, attracting transparency obligations rather than high-risk ones. This document, and the disclosure carried on every page, are intended to discharge the obligation to tell you that you are reading machine-generated content. This assessment should be confirmed against the Act's applicable dates.
9. Changes
We will update this statement when what the machine does changes. Its date is at the top.
End of AI Use and Transparency Statement.