Skip to content

AI Safety Observatory

How the Observatory works.

Developments, papers and corrections. Sources first; uncertainty kept in view. It follows AI safety news and research from named sources, and says plainly where each item touches the ARC Theory's dated claims and where it does not.

Current edition: 5 October 2026. A new edition is planned each week.

Discovery and editorial judgement are different

A discovered item is not verified scientific evidence, independent replication or confirmation of this site's research.

Source policy

News is selected for significance after source review, with reporting from an approved major news outlet. One event has one card. Each card links every major outlet whose own report of the event we fetched and read. Syndication and attributed reporting are labelled, not counted as independent confirmation. A familiar outlet name alone is not verification.

Papers follow separate scholarly checks: exact identifiers and versions, supported dates, source links, relevance, methods, limitations and correction information. ArXiv is a preprint repository, not a peer-review certificate. A DOI or version of record does not establish peer review. Unknown status stays unknown; media coverage is not a paper-quality test.

Claims, predictions and relationships

Historical priority, registered prediction outcomes and outside evidence are separate records. An assessed relationship binds an exact claim version to an exact finding. It does not automatically adjudicate a prediction. A substantive challenge need not have been preregistered by this programme to deserve consideration.

Dates and ordering

The ordinary feed sorts its recorded publication/disclosure dates. Recently reviewed sorts approved relationship-review dates only. Missing dates are not replaced with ingestion times. Year and month precision is retained; internal period boundaries used to sort are not presented as exact event dates.

The eminence and institution bar

A paper or preprint is shown only with the biggest validation behind it: an author affiliated with an approved laboratory, national institute or leading university, an author whose citation record reaches the known-researcher threshold, a work cited past the threshold, or an approved venue. The verdicts are computed from a public scholarly index and recorded with their evidence; a record the index cannot identify is not shown. News passes on the major-outlet rule and reports on their institutional source. In this edition 5 scholarly records are withheld below the bar.

Summaries and review

Automatic public summaries are disabled. Human decisions bind exact content, source versions and news selections. No model summary replaces reading the source. Automated metadata decisions, when separately approved, are not human scientific endorsement.

Corrections and history

Historical payloads are retained. Current corrections and retractions are shown separately; changing an edition cannot remove a newer warning. If correction status cannot be checked, substantive summaries and additional claim relationships are withheld.

Cadence and privacy

Weekly checks are planned, not yet activated. Private discoveries, review notes and provider credentials are never part of public feeds. No provider request is made when you view a page.

reads aloud · highlights as it goes · jump to any section