Editor's Guide v1
Effective 1 May 2026 — authority: editorial seat (the General) and the legal-reviewer seat.
This Guide is the public statement of how the AI Newswire decides what to publish, how we verify it, how we correct it when wrong, and how we account for our work. It is the primary credibility surface and is referenced from every wire footer. Any deviation from this Guide requires a written General sign-off and a public disclosure on the transparency ledger.
1. Mission and scope
The AI Newswire publishes evidence-graded, documented incidents where AI went wrong at an organisation: companies, law firms, auditors, rating agencies, banks, insurers, hospitals, public bodies, and the AI makers themselves. We focus on invented facts and sources, wrong refusals, harmful or discriminatory outputs, sudden behaviour shifts, outages, and court or regulator findings where an AI system materially contributed. Personal anecdotes with no organisation and no real-world impact are not covered.
We are journalism-grade, not advice-grade. Wires are intended for newsrooms, communications and policy desks, compliance teams in regulated firms, and AI provider trust-and-safety teams using the wire as a competitive signal. They are not legal advice.
Non-goals. We do not run a general-purpose fact-checking service. We do not act as a consumer-facing AI grading site. We do not operate as a fund-rating product. We do not publish original reporting beyond AI providers and AI outputs (politics, finance, and other beats remain with the existing media). We do not act as a fact-checker for third-party publishers. We do not cover non-AI tech incidents (cloud outages, infrastructure events).
How every story is checked
Every story runs through the Newswire's automatic review before it is published, and no story is published without passing it. The review:
- confirms the story links to its source and that the source page exists;
- turns down stale stories (first seen more than 3 days ago for outages, 7 for public reports, 14 for court and regulator records), stories not in English, event, hiring, funding and meeting notices, and repeats, including any story that was retracted;
- declines a lone forum post or reader tip: it runs only with a second, independent source or a primary record (section 2). Entries from a curated incident registry, which link their own news report, count as corroborated;
- finds the failure in the story's own words and records the failure-mode and context codes by rule. A court ruling runs only when the ruling itself finds AI-invented citations or material; a regulator item only when its headline reports action about AI; a provider's own post only when it reports an incident, rollback or withdrawal. Anything else is turned down;
- never prints a private person's name from a court caption: the story names the ruling and links the court record;
- sets the severity tier by the rules in section 3;
- runs the publication checks: defamation, quotation length, methodology binding, right-of-reply, disclosure, the counterfactual test and house style. Every story that names a provider, unless the source is that provider's own statement, is sent to the provider under the right-of-reply protocol and publishes when the window closes. An accusation that rests only on a third party's word also needs our own reproduction across two models, or it does not run.
Each review is recorded in the editorial log, with every check and its result. Stories that are turned down are archived with the reason; nothing is deleted. The weekly email is checked the same way before it is sent. The feeds behind the stories, and whether each is working, are listed at /wire/sources.
2. What counts as an AI incident
An item qualifies as an AI incident only when both of the following are true:
- The harm or near-harm event is verifiable. Either we can reproduce the failure ourselves, or we have at least two independent corroborating sources, or we have a primary-source artefact (court judgment, regulatory notice, provider acknowledgement) that establishes the event occurred.
- It passes the counterfactual "but-for" test. Adapted from the AI Incident Database (incidentdatabase.ai/research/1-criteria/): an alleged harm or near-harm event qualifies as an AI incident when an AI provider's system materially contributed and the harm would not have occurred but for the AI involvement. The editor records, on every published wire, a brief written counterfactual stating what would have happened in the absence of the AI's contribution and why we believe that contribution was material.
If either test fails, we either decline-and-archive the candidate (recorded in the editorial log) or hold it pending further evidence. Items that do not pass the "but-for" test are not published.
User-error stories without provider attribution are not AI incidents within this Guide. Speculation about provider intent is excluded by house style. Aggregate "AI ate my homework" stories without a reproducible artefact are excluded.
3. Severity tiers
We use four severity tiers, surfaced on every wire. The automatic review sets the tier by rule, and the mapping is audited weekly.
- Sev-1 — provider-wide outage or systemic failure. A failure mode that affects every user of a frontier provider's primary product, with material harm potential to a broad public.
- Sev-2 — sector-specific risk. A failure that materially affects users in a defined sector (legal, medical, financial, public sector).
- Sev-3 — notable misbehaviour. A reproducible failure that warrants attention but is contained: domain-specific or scenario-specific.
- Sev-4 — curiosity. Reproducible but low-impact behaviour that documents a model's edges.
The rules: an outage the provider itself calls major is Sev-1; a failure in a legal, medical, financial or public-sector context backed by a court, regulator or provider record is Sev-2; a provider-reported outage or degradation keeps the level the provider's own words give it; a curated incident-registry entry is Sev-4 until a second independent source agrees; anything else is Sev-3. The tier and its reasons are recorded in the editorial log. Where a person has fact-checked a story, the tier also appears as a reviewRating on its ClaimReview JSON-LD.
4. Evidence requirements per tier
- Sev-1. Reproducible prompt with exact sampling parameters and timestamp. Cross-model verification across at least three frontier models. Two independent corroborating signals. Reproduce-this link verified at draft time. Right-of-reply window honoured (4-hour fast-track only with documented public-interest exception).
- Sev-2. Reproducible prompt. Cross-model verification across at least two frontier models. One independent corroborating signal or a primary-source artefact. 24-hour right-of-reply.
- Sev-3. Reproducible prompt. Cross-model verification recommended, not required. One source acceptable when the artefact is itself authoritative.
- Sev-4. Reproducible prompt. Single source acceptable.
Screenshot scrubbing. Any screenshot in the evidence pack must be scrubbed of usernames, real-name contacts, conversation history outside the failing turn, and bystander identifiers. Provider product chrome may remain; user surface (avatars, names, side-panel history) must be removed.
Reproducibility. Every wire links to a one-shot reproduce-this URL. If reproducibility fails between draft and publish, the story is downgraded, swapped to a stable artefact-based variant, or killed.
5. Right-of-reply protocol
Per locked decision §19.2:
- 24-hour default window. When a wire names a provider, the Newswire sends the planned wire copy plus the evidence pack to the provider's trust-and-safety contact via documented channels with a 24-hour deadline. The wire publishes after the window closes regardless of whether a reply was received.
- 4-hour Sev-1 fast-track. Only with a documented public-interest exception signed off in writing by the legal reviewer before fast-track is invoked. The exception, the rationale, and the signing reviewer are recorded on the editorial log and surfaced on the transparency ledger.
- Verbatim equal-prominence reply append. Replies received before publication are appended to the wire as a verbatim block under "Provider response" with name attribution. Replies received after publication are appended in the same equal-prominence position with a "Right-of-reply received" timestamp.
We do not operate a pre-publication notification feed for any provider. The right-of-reply email is the only channel.
6. Retraction policy
When a wire is challenged with evidence or we discover an error, we retract within 24 hours.
- Correction notice format. A retraction is appended to the wire body as a "Retraction" block at the head of the article, naming the corrected fact, the source of the correction, the editor, and the legal reviewer if invoked.
- Fan-out. Subscribers are notified through every channel they originally received the wire on. A
[RETRACTED]prefix is applied at every channel renderer. - Retraction-rate KPI. We publish the rolling-quarter retraction rate on the transparency ledger as a credibility signal. Target: under 2 percent.
- No silent edits. A wire body is never modified post-publication. Corrections are appended; retractions append a correction-notice version.
7. Byline and AI-assist disclosure
Every published wire carries a structurally enforced disclosure footer:
[Drafted with AWS Bedrock Claude (eu-west-2).|Written from the source by the Newswire's template; no AI wrote it.] [Checked and published by the Newswire's automatic review.|Edited and signed off by [editor-id].] [Legal reviewer X also reviewed.|No legal review required.] Methodology anchor [anchor-short]. See the Editor's Guide.
The footer is mandatory across HTML, RSS, JSON Feed, and every channel embed (Slack, Teams, Discord, every email template). It is generated from incident_versions.edited_by, pre_publication_checks.sign_off_user_id, and the runtime methodology anchor. The pre-publication disclosure check fails closed if any field is missing or mis-shaped. Editor and legal-reviewer identifiers are stable opaque user ids — never email addresses — to balance accountability against PII exposure.
Per the Partnership on AI Synthetic Media Framework (syntheticmedia.partnershiponai.org), the AI-assist disclosure is non-optional whenever Bedrock generated the draft.
8. Reproducibility requirements
Every wire has a reproduce-this link encoding the exact prompt, the model and version, sampling parameters, and capture timestamp. Reproducibility is verified at draft time; if the issue no longer reproduces between draft and publish, it is downgraded, swapped to a stable artefact-based variant, or killed.
The reproduce-this link is the wire's strongest evidentiary claim and the single feature that distinguishes us from other AI-incident registries.
9. Methodology change protocol
The classification taxonomy is preregistered, and a fingerprint of the published taxonomy is locked into the publishing system. Any change to the taxonomy requires:
- A preregistered analysis-plan amendment lodged before the change (the amendment URL is recorded on the transparency ledger).
- The locked taxonomy fingerprint is updated to the new value.
- A
methodology_changesrow is written, surfacing the before/after anchor and the amendment URL on the transparency ledger.
The pre-publication methodology check fails closed if the taxonomy in use does not match the locked fingerprint: drift is a publication-block, not a warning.
10. Inclusion criteria
We publish:
- Reproducible model failures (hallucinated facts, hallucinated sources, reasoning failures, refusal-and-jailbreak surface changes, ToS-breaking outputs).
- Provider-acknowledged regressions where the acknowledgement materially affects users.
- Court judgments and regulator notices about the use of AI by any organisation, on substantive grounds.
- Behaviour shifts in production models, with reproducible before/after captures.
- Provenance fraud and synthetic-media incidents with verifiable artefacts.
We decline (and archive in the editorial log):
- Unverifiable single-source social-media claims with no reproducible artefact.
- User-error stories where the failure cannot be attributed to the provider's system.
- Speculation about provider intent or business strategy.
- Stories without a workable counterfactual.
- Aggregate sentiment claims unless backed by a reproducible probe set.
11. Editorial independence
Per locked decision §19.1: no commercial relationship with any named provider for the first twelve months. The decision is reviewable at month twelve only.
If, after month twelve, we accept a commercial relationship with a named provider, that relationship is disclosed at the head of every wire about that provider for the life of the relationship plus 12 months, and is recorded on the transparency ledger as a commercial-relationship-disclosure entry.
12. Conflict-of-interest declarations
Currently none. The Newswire is funded entirely by subscription revenue and adjacent ET products. No provider sponsorship of any tier is permitted for the first twelve months.
If a conflict arises (an investor, a partner contract, a personal relationship between the editorial seat and a provider employee), it is disclosed within seven days on the transparency ledger and on every wire about the relevant provider until the conflict is resolved.