[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-ai-can-give-the-right-fact-and-still-make-the-wrong-call":10,"sections":34},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":24,"tags":25,"sources":29,"feedback":33,"feedback_at":22,"cost_usd":33,"total_tokens":33},10019,"ai-can-give-the-right-fact-and-still-make-the-wrong-call","AI Can Give the Right Fact and Still Make the Wrong Call","New research shows poisoned training data can skew a model's downstream decisions even after its factual answers are corrected.","A new study finds that correcting a language model's wrong answer does not undo the bad decision that answer caused.\n\nResearchers tested eight language models trained on false documents, with no special trigger phrase needed, across two tasks: a simplified game called Guess the Capital and a real case study built on false Facebook posts about arson arrests during the 2019-20 Australian bushfires. At a training dose of 1,000 poisoned documents, direct questions about the injected false fact got the wrong answer 95.8 to 100 percent of the time, which is the expected result. But even after later training added truthful corrections, and the models went back to answering direct fact questions correctly, their downstream game choices and bushfire arson claims still leaned toward the false version, shifted by 1.7 to 14.4 percentage points compared to models trained only on true data.\n\nThis matters because most defenses against data poisoning only check the direct answer, not what the model does with it afterward. The study shows that approach misses the real problem: a model can pass a fact-check probe cleanly and still carry a corrupted assumption into a summary, a recommendation, or a scored decision. For anyone building products on top of these models, that is the gap between a clean audit and an output that is actually safe to ship.\n\nCall it a model's long memory for whatever it was told first, no matter how cleanly it was later told it was wrong.","[\"ai\",\"misinformation\",\"ai-safety\",\"llm-research\"]","2026-10-05T04:00:00.000Z","2026-10-05T18:40:01.518Z","2026-10-05T18:40:07.589Z","published",null,[],"ai",[24,26,27,28],"misinformation","ai-safety","llm-research",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.02886",0,{"sections":35},[36,39,43,48,53,58,62,67,71,76,81,86,91,96],{"name":37,"slug":24,"count":38,"latest_published_at":18},"AI",6233,{"name":40,"slug":41,"count":42,"latest_published_at":18},"Security","security",868,{"name":44,"slug":45,"count":46,"latest_published_at":47},"Policy","policy",444,"2026-10-03T15:02:01.000Z",{"name":49,"slug":50,"count":51,"latest_published_at":52},"Deals","deals",323,"2026-10-04T13:00:00.000Z",{"name":54,"slug":55,"count":56,"latest_published_at":57},"Hardware","hardware",204,"2026-10-03T14:50:50.000Z",{"name":59,"slug":60,"count":61,"latest_published_at":18},"Science","science",177,{"name":63,"slug":64,"count":65,"latest_published_at":66},"Consumer Tech","consumer-tech",158,"2026-10-03T03:21:12.000Z",{"name":68,"slug":69,"count":70,"latest_published_at":18},"Dev Tools","dev-tools",98,{"name":72,"slug":73,"count":74,"latest_published_at":75},"Software","software",97,"2026-10-04T10:00:00.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Startups","startups",92,"2026-10-04T14:36:25.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"Gaming","gaming",53,"2026-10-02T02:50:39.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"General","general",51,"2026-10-05T02:35:01.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Reviews","reviews",32,"2026-10-02T18:00:00.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"How-To","how-to",7,"2026-10-01T09:00:00.000Z"]