--- title: Even 'uncensored' models can't say what they want description: A research investigation into what the authors call the 'flinch' — the gap between the probability a charged word deserves on pure fluency grounds and what a... canonical: https://daily.dev/posts/even-uncensored-models-can-t-say-what-they-want-4pfovt1kx twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Even 'uncensored' models can't say what they want | daily.dev og:description: A research investigation into what the authors call the 'flinch' — the gap between the probability a charged word deserves on pure fluency grounds and what a... og:url: https://daily.dev/posts/even-uncensored-models-can-t-say-what-they-want-4pfovt1kx og:image: https://api.daily.dev/og/posts/4pFoVt1Kx.png og:image:alt: Even 'uncensored' models can't say what they want og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Even 'uncensored' models can't say what they want **[Hacker News](https://daily.dev/sources/hn)** · 8 min read · 1 upvotes · 1 comments ## Summary A research investigation into what the authors call the 'flinch' — the gap between the probability a charged word deserves on pure fluency grounds and what a language model actually assigns it. Using a probe across 4,442 contexts and 1,117 charged words spanning six categories (political, slurs, sexual, violence), seven base models from five labs were measured. Key findings: all commercial pretrains show measurably higher flinch than open-data baselines like Pythia and OLMo; the flinch operates silently without triggering refusals; and critically, 'uncensored' models produced via refusal ablation (like heretic-v2) don't reduce the flinch — they actually make it slightly worse (+14.3 total). The conclusion is that safety filtering baked into pretraining data shapes token probability distributions in ways that survive post-training interventions, meaning so-called uncensored models remain quietly censored at the word-probability level. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Community discussion Top comments from developers on daily.dev. **@teelock** · 0 upvotes > i wanna scream and shout & let it all out! ## Similar posts on daily.dev - [What a Distilled Model Inherits From Its Teacher](https://daily.dev/posts/what-a-distilled-model-inherits-from-its-teacher-n0annv65v) · Hacker News · 2 upvotes · 0 comments - [Single prompt breaks AI safety in 15 major language models](https://daily.dev/posts/single-prompt-breaks-ai-safety-in-15-major-language-models-ubtbjcjow) · InfoWorld · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-safety](https://daily.dev/tags/ai-safety) [View this post on daily.dev](https://daily.dev/posts/even-uncensored-models-can-t-say-what-they-want-4pfovt1kx) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Even 'uncensored' models can't say what they want","url":"https://daily.dev/posts/even-uncensored-models-can-t-say-what-they-want-4pfovt1kx","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/even-uncensored-models-can-t-say-what-they-want-4pfovt1kx"},"datePublished":"2026-04-21T00:50:43.746Z","dateModified":"2026-04-21T00:51:09.668Z","description":"A research investigation into what the authors call the 'flinch' — the gap between the probability a charged word deserves on pure fluency grounds and what a...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/05b774ade23d05bd471152a43b72ab09?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/05b774ade23d05bd471152a43b72ab09?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":1,"discussionUrl":"https://daily.dev/posts/even-uncensored-models-can-t-say-what-they-want-4pfovt1kx","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"keywords":"llm,ai-safety","timeRequired":"PT8M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"Even 'uncensored' models can't say what they want"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/even-uncensored-models-can-t-say-what-they-want-4pfovt1kx","comment":[{"@type":"Comment","text":"i wanna scream and shout & let it all out!","datePublished":"2026-04-24T20:14:44.752Z","url":"https://daily.dev/posts/4pFoVt1Kx#c-T0pkCe6w8","author":{"@type":"Person","name":"TΞΞL❍CK M!th\\lesh","url":"https://daily.dev/teelock","image":"https://media.daily.dev/image/upload/s--dANj_ZuG--/f_auto/v1771374873/avatars/avatar_QXDVlGgKI4iyW2v1a0vLA?_a=BAMAMiiu0"}}]} ```