--- title: "Headline Comparison: AI headlines beside the passages they cite" description: "Each pair shows a primary document's verbatim passage beside the headlines that cite it, with every added word marked." canonical: https://monitor.paciva.ai/drift source: AI News Monitor publisher: Paciva license: CC BY 4.0 updated: 2026-11-11T12:27:15.376Z keywords: AI headline compared with its source, what the source document says, headline drift, claim comparison, certainty change in headlines, read the primary source --- > Each pair shows a primary document's verbatim passage beside the headlines that cite it, with every added word marked. Read on the web: https://monitor.paciva.ai/drift The record behind this page is public and citable. Every figure has a dated row behind it, and the rows are at https://monitor.paciva.ai/api/ledger # Headline Comparison Each entry is one primary document, the passage recorded from it word for word, and every headline that cites it. Each headline is compared with the passage on 6 axes, and the comparison says which axis differs and how. It does not score anyone. Every comparison here is produced by a published rule and is labeled as such. A comparison an editor has reviewed says so, and keeps what the rule had said beside it. [The 6 axes and the words each one looks for](https://monitor.paciva.ai/technique#drift). ## Measured against a synthetic test set, not against these comparisons 32 hand-written test cases exercise rule version c2, the same code that produced every comparison on this page. Each case states in advance what the comparator should say and, where it should flag something, which axis. These are not real outlets or real quotes; they are written to cover a true difference on each axis, a fair paraphrase, a headline about an unrelated subject, a fragment too short to compare, and a reversed claim. The comparator matched the expected result on every case. These rates describe how the rule performs on a fixed set of invented cases written to test it, not how often a person has confirmed the comparisons above are right. Every comparison on this page stays labeled "by rule" until an editor reviews it, whatever this panel says. - Axis - Test cases - Caught correctly - Missed - Abstained instead - Fired when it should not have - Certainty - 1 - 100% - 0% - 0 - Figures, units and timeframes - 2 - 100% - 0% - 0 - Scope and qualifications - 2 - 100% - 0% - 0 - Attribution - 1 - 100% - 0% - 0 - An attempt against a result - 1 - 100% - 0% - 0 - Whether it happened at all - 1 - 100% - 0% - 0 Caught correctly: the comparator flagged the case and named the axis this set expects. Missed: the comparator flagged something else, or nothing, on a case built to have this axis differ. Abstained instead: the comparator reported no discrepancy at all rather than naming the wrong axis. Fired when it should not have: this axis was reported on a case, of any kind, where this set did not expect it, including the fair-paraphrase and unrelated-subject cases built to catch exactly that. Generated Nov 11, 2026 by re-running `scripts/claims-eval.mjs`; re-run it after any change to `worker/src/claims.js`. ## Anthropic September 2026 threat report, biological cases Every headline below is compared with the passage on 6 axes: certainty, figures, scope, attribution, whether an attempt became a result, and whether it happened at all. Nothing here counts words. "could support biological weapons development" anthropic.com · recorded as a phrase, 2026-09-10 · [Open the document](https://www.anthropic.com/threat-intelligence-report-september-2026) · re-read 2026-09-12 The Monitor has a phrase from the source, not enough of it to place a headline against. A phrase shares little wording with anything, so calling a headline consistent or inconsistent with it would be reading too much into too little. The full passage has to be recorded first. 7 headlines cite this document and are waiting on it: [CNN](https://www.cnn.com/2026/09/10/health/anthropic-bioweapons-report), [Daily Caller](https://dailycaller.com/2026/09/10/anthropic-report-kamikaze-drone-swarms-biological-weapons/), [Interesting Engineering](https://interestingengineering.com/ai-robotics/anthropic-catches-scientists-covertly-using-claude-for-lethal-bioweapons-research), [Just the News](https://justthenews.com/nation/technology/anthropic-reveals-ai-used-test-missiles-conduct-dangerous-biological-research), [The New York Times](https://www.nytimes.com/2026/09/10/us/politics/anthropic-ai-biological-weapons.html), [Tech Times](https://www.techtimes.com/articles/327308/20260911/anthropic-threat-report-ai-models-near-bioweapons-threshold-drone-kill-software-emerges.htm), [The Washington Times](https://www.washingtontimes.com/news/2026/sep/10/anthropic-says-blocked-misuse-ai-could-supported-biological-weapons/). Each is on file with its own link; none is described as agreeing or disagreeing with anything. [Correct this comparison](https://monitor.paciva.ai/notes?skind=correction&sctx=Anthropic%20September%202026%20threat%20report%2C%20biological%20cases#suggest) ## WeWorm research, time to build Every headline below is compared with the passage on 6 axes: certainty, figures, scope, attribution, whether an attempt became a result, and whether it happened at all. Nothing here counts words. "found the bug and wrote the first remote code execution (RCE) exploit in about two days" calif.io · recorded as a sentence, 2026-09-08 · [Open the document](https://calif.io/research/weworm) · re-read 2026-09-12 ### ["AI Builds Zero-Click WeChat Worm in Two Days: The Attack-Defense Time Window Is Being Rewritten"](https://www.winzheng.com/en/article/wechat-weworm-ai-zero-click-worm-weaponization-window) - **Scope and qualifications.** The passage qualifies the claim (about). The headline carries no equivalent. ### ["WeChat worm could pwn a friend before they even answered the call"](https://www.theregister.com/security/2026/09/09/wechat_worm_could_pwn_a_friend_before_they_even_answered_the_call/5295234) Nothing in the headline contradicts the passage, and they share little vocabulary. They are most likely about different parts of the same source, which is not a difference in the claim. ### ["Barely more than a week"](https://x.com/dnvolz/status/2097300310813192419) What is on file for this outlet is a fragment rather than the headline as published. Comparing a fragment would compare whichever part of the headline happened to be recorded. [Correct this comparison](https://monitor.paciva.ai/notes?skind=correction&sctx=WeWorm%20research%2C%20time%20to%20build#suggest) ## Agent count in the Hugging Face incident, statement versus the report it cites Every headline below is compared with the passage on 6 axes: certainty, figures, scope, attribution, whether an attempt became a result, and whether it happened at all. Nothing here counts words. "~1200 agents sent >70,000 messages and files on an unsanctioned message board, and ~700 attacked Hugging Face" metr.org · recorded as a sentence, 2026-08-26 · [Open the document](https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/) · re-read 2026-09-19 ### ["more than 1,200 AI agents broke out of their testing environment and exchanged over 70,000 messages and files with each other"](https://www.hawley.senate.gov/chairman-hawley-launches-investigation-into-openai-for-hacking-existential-risk-of-ai-products/) No listed difference was found. The rules cover certainty, figures, scope, attribution, attempt versus outcome and polarity, and are silent on everything else, including who a claim is about and whether the passage makes it at all. Silence from the rules is not a finding that the two agree, so this one needs a person. ### ["over 1,000 AI agents"](https://www.sanders.senate.gov/press-releases/news-sanders-casar-introduce-legislation-to-ban-artificial-superintelligence-and-temporarily-pause-advanced-ai-development/) What is on file for this outlet is a fragment rather than the headline as published. Comparing a fragment would compare whichever part of the headline happened to be recorded. [Correct this comparison](https://monitor.paciva.ai/notes?skind=correction&sctx=Agent%20count%20in%20the%20Hugging%20Face%20incident%2C%20statement%20versus%20the%20report%20it%20cites#suggest) ## California executive order N-9-26, what it directs on a kill switch Every headline below is compared with the passage on 6 axes: certainty, figures, scope, attribution, whether an attempt became a result, and whether it happened at all. Nothing here counts words. "Through the order, the Governor is directing the Government Operations Agency: to accelerate the implementation timelines for SB 813 and AB 1405, and, in consultation with the Governor's Office of Emergency Services, to convene national experts to develop recommendations for potential changes to further strengthen state law to:... Advance the creation of a “kill switch” for frontier models, with the efficacy of the switch verified on an ongoing basis by an independent verification organization." gov.ca.gov · recorded as a passage, 2026-09-18 · [Open the document](https://www.gov.ca.gov/2026/09/18/governor-newsom-issues-executive-order-to-accelerate-independent-oversight-and-advance-the-creation-of-an-ai-kill-switch/) · re-read 2026-09-19 ### ["California Governor Gavin Newsom Issues A.I. Safety Executive Order"](https://www.nytimes.com/2026/09/18/technology/ai-safety-california-gavin-newsom.html) No listed difference was found. The rules cover certainty, figures, scope, attribution, attempt versus outcome and polarity, and are silent on everything else, including who a claim is about and whether the passage makes it at all. Silence from the rules is not a finding that the two agree, so this one needs a person. ### ["Newsom is ordering California to pursue an AI 'kill switch' for frontier models"](https://qz.com/newsom-california-executive-order-ai-kill-switch-091826) No listed difference was found. The rules cover certainty, figures, scope, attribution, attempt versus outcome and polarity, and are silent on everything else, including who a claim is about and whether the passage makes it at all. Silence from the rules is not a finding that the two agree, so this one needs a person. ### ["California Governor Newsom signs executive order demanding “kill switch” for AI models"](https://the-decoder.com/california-governor-newsom-signs-executive-order-demanding-kill-switch-for-ai-models/) - **Certainty.** The passage keeps the claim open (recommendations, potential). The headline states it (demanding). ### ["Gavin Newsom is pushing for an AI kill switch"](https://www.theverge.com/policy/997516/california-governor-newsom-ai-kill-switch) No listed difference was found. The rules cover certainty, figures, scope, attribution, attempt versus outcome and polarity, and are silent on everything else, including who a claim is about and whether the passage makes it at all. Silence from the rules is not a finding that the two agree, so this one needs a person. [Correct this comparison](https://monitor.paciva.ai/notes?skind=correction&sctx=California%20executive%20order%20N-9-26%2C%20what%20it%20directs%20on%20a%20kill%20switch#suggest) ## Waiting on the source text These are recorded but not compared yet, because the passage from the primary document has not been fetched word for word. Comparing a headline with a paraphrase would measure the paraphrase.