[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-researchers-build-a-voting-panel-to-catch-ai-hallucinations":10,"sections":35},{"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":30,"feedback":34,"feedback_at":22,"cost_usd":34,"total_tokens":34},6423,"researchers-build-a-voting-panel-to-catch-ai-hallucinations","Researchers Build a Voting Panel to Catch AI Hallucinations","A team's multi-model voting system flagged hallucinated text in AI image captions, topping three of four languages in a research benchmark.","A team of researchers just built a panel of AI models to catch other AI models lying about what they see.\n\nThe system, entered in the SHROOM-Visions shared task, works on vision-language models that read images and generate captions, then flags the exact word spans where a model made something up. The method fine-tunes several vision-language models to act as independent judges, then merges their guesses through character-level majority voting - if enough judges flag the same stretch of text, it gets marked hallucinated. The researchers also tested activation probes, a technique that inspects a model's internal signals rather than just its final output. Across four languages, the approach placed first in three and landed on the podium in every language and metric it was scored on.\n\nHallucination detection is the unglamorous plumbing behind trustworthy AI captions, alt text, and image-based chatbots - it is the difference between a tool you can rely on and one you have to fact-check yourself. The more interesting finding is that when the AI judges disagreed with each other, that disagreement tracked the cases where human annotators also disagreed, suggesting some hallucinations sit in genuinely ambiguous territory rather than being clean-cut errors.\n\nStill, a voting committee of models that each hallucinate on their own is a stopgap, not a fix - it catches the lies better, it does not explain why the models tell them.","[\"hallucination detection\",\"vision-language models\",\"ai research\",\"benchmarks\"]","2026-09-16T04:00:00.000Z","2026-09-17T15:37:49.833Z","2026-09-17T15:38:01.763Z","published",null,[],"ai",[26,27,28,29],"hallucination detection","vision-language models","ai research","benchmarks",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.17327",0,{"sections":36},[37,41,46,51,56,60,64,69,74,78,83,88,93,98],{"name":38,"slug":24,"count":39,"latest_published_at":40},"AI",3853,"2026-09-17T08:27:09.000Z",{"name":42,"slug":43,"count":44,"latest_published_at":45},"Security","security",648,"2026-09-17T04:00:00.000Z",{"name":47,"slug":48,"count":49,"latest_published_at":50},"Policy","policy",338,"2026-09-11T04:00:00.000Z",{"name":52,"slug":53,"count":54,"latest_published_at":55},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":57,"slug":58,"count":59,"latest_published_at":45},"Hardware","hardware",154,{"name":61,"slug":62,"count":63,"latest_published_at":45},"Science","science",114,{"name":65,"slug":66,"count":67,"latest_published_at":68},"Consumer Tech","consumer-tech",99,"2026-09-09T17:27:33.000Z",{"name":70,"slug":71,"count":72,"latest_published_at":73},"Software","software",75,"2026-09-10T20:41:21.000Z",{"name":75,"slug":76,"count":77,"latest_published_at":45},"Dev Tools","dev-tools",73,{"name":79,"slug":80,"count":81,"latest_published_at":82},"Startups","startups",55,"2026-09-09T23:14:29.000Z",{"name":84,"slug":85,"count":86,"latest_published_at":87},"Gaming","gaming",43,"2026-09-10T12:18:06.000Z",{"name":89,"slug":90,"count":91,"latest_published_at":92},"General","general",41,"2026-09-08T01:57:23.000Z",{"name":94,"slug":95,"count":96,"latest_published_at":97},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":99,"slug":100,"count":101,"latest_published_at":102},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]