[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-study-finds-no-gate-decides-what-ai-models-can-report":10,"sections":40},{"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":30,"tags":31,"sources":35,"feedback":39,"feedback_at":22,"cost_usd":39,"total_tokens":39},5151,"study-finds-no-gate-decides-what-ai-models-can-report","Study Finds No Gate Decides What AI Models Can Report","New research finds no gate controls what a language model can report, only a mid-depth attention window that gathers hidden values into words.","A new interpretability study finds that no internal gatekeeper decides which facts a language model can put into words; attention does that job instead.\n\nThe researchers tested open-weight language models using Jacobian-based lens analysis on a benchmark whose five task variants share identical context but differ in whether a hidden value is actually needed to answer. They were hunting for a 'gate': some mechanism that would only let a value into the model's reportable workspace when the task demanded it. They didn't find one. A single shared linear decoder pulled the hidden value out of every task variant, including the control that never needed it, at 6.4 to 9.0 times its selection-corrected floor. What did show up was a mid-depth window of layers where attention, not the feedforward sublayers, gathers the value into a form the model can later report: transport through that window measured at least 17 times higher than at shallower layers, and the window landed at the same fractional depth in both a 64-layer hybrid model and a 62-layer dense model from an unrelated family.\n\nThat's a meaningful correction for anyone using interpretability probes to claim a model 'knows' or 'doesn't know' something: presence of information tracks depth, not need. The paper also flags a calibration problem worth taking seriously: three different readout components landed within 12 percent of each other on raw scores while differing 7.4 times in how much they actually swayed the model's answer, meaning today's readability tests can't reliably tell 'the model has this' apart from 'the model is using this.'\n\nIn other words, catching a model thinking out loud says less about what it's doing with an idea than where in its layer stack that idea happens to live.","[\"ai\",\"interpretability\",\"language models\",\"arxiv\"]","2026-08-18T04:00:00.000Z","2026-08-18T07:15:04.837Z","2026-08-18T07:15:16.659Z","published",null,[24],{"id":25,"reviewer":26,"round":27,"reason":28,"status":29},"editor-r1","editor",1,"Add the specific effect-size numbers from the source (the 6.4-9.0x floor multiple, the 17x transport threshold, and the 12%\u002F7.4x calibration caveat figures) instead of vague phrases like 'well above' and 'at least' — the concrete magnitudes are available in the source and needed to meet the concrete-specifics bar.","resolved","ai",[30,32,33,34],"interpretability","language models","arxiv",[36],{"name":37,"url":38},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.15022",0,{"sections":41},[42,46,50,55,60,65,70,75,80,84,89,94,99,104],{"name":43,"slug":30,"count":44,"latest_published_at":45},"AI",3293,"2026-08-20T04:00:00.000Z",{"name":47,"slug":48,"count":49,"latest_published_at":45},"Security","security",435,{"name":51,"slug":52,"count":53,"latest_published_at":54},"Policy","policy",210,"2026-08-19T09:32:27.000Z",{"name":56,"slug":57,"count":58,"latest_published_at":59},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":61,"slug":62,"count":63,"latest_published_at":64},"Hardware","hardware",140,"2026-08-19T18:25:42.000Z",{"name":66,"slug":67,"count":68,"latest_published_at":69},"Consumer Tech","consumer-tech",95,"2026-08-18T16:05:00.000Z",{"name":71,"slug":72,"count":73,"latest_published_at":74},"Science","science",90,"2026-08-19T18:41:02.000Z",{"name":76,"slug":77,"count":78,"latest_published_at":79},"Software","software",73,"2026-08-18T07:51:50.000Z",{"name":81,"slug":82,"count":83,"latest_published_at":18},"Dev Tools","dev-tools",69,{"name":85,"slug":86,"count":87,"latest_published_at":88},"Startups","startups",47,"2026-08-19T19:13:46.000Z",{"name":90,"slug":91,"count":92,"latest_published_at":93},"Gaming","gaming",41,"2026-07-09T04:00:00.000Z",{"name":95,"slug":96,"count":97,"latest_published_at":98},"General","general",33,"2026-08-18T22:18:13.000Z",{"name":100,"slug":101,"count":102,"latest_published_at":103},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":105,"slug":106,"count":107,"latest_published_at":108},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]