{"id":509646,"date":"2026-04-02T22:03:08","date_gmt":"2026-04-02T22:03:08","guid":{"rendered":"https:\/\/www.newsbeep.com\/uk\/509646\/"},"modified":"2026-04-02T22:03:08","modified_gmt":"2026-04-02T22:03:08","slug":"anthropic-says-that-claude-contains-its-own-kind-of-emotions","status":"publish","type":"post","link":"https:\/\/www.newsbeep.com\/uk\/509646\/","title":{"rendered":"Anthropic Says That Claude Contains Its Own Kind of Emotions"},"content":{"rendered":"<p>Claude has been through a lot lately\u2014a public <a href=\"https:\/\/www.wired.com\/story\/department-of-defense-responds-to-anthropic-lawsuit\/\" class=\"text link\" rel=\"nofollow noopener\" target=\"_blank\">fallout with the Pentagon<\/a>, <a href=\"https:\/\/www.axios.com\/2026\/03\/31\/anthropic-leaked-source-code-ai\" class=\"text link\" rel=\"nofollow noopener\" target=\"_blank\">leaked source code\u2014<\/a>so it makes sense that it would be feeling a little blue. Except, it\u2019s an AI model, so it can\u2019t feel. Right?<\/p>\n<p class=\"paywall\">Well, sort of. A new study from Anthropic suggests models have digital representations of human emotions like happiness, sadness, joy, and fear, within clusters of artificial neurons\u2014and these representations activate in response to different cues.<\/p>\n<p class=\"paywall\">Researchers at the company probed the inner workings of Claude Sonnet 4.5 and found that so-called \u201cfunctional emotions\u201d seem to affect Claude\u2019s behavior, altering the model\u2019s outputs and actions.<\/p>\n<p class=\"paywall\">Anthropic\u2019s findings may help ordinary users make sense of how chatbots actually work. When Claude says it is happy to see you, for example, a state inside the model that corresponds to \u201chappiness\u201d may be activated. And Claude may then be a little more inclined to say something cheery or put extra effort into vibe coding.<\/p>\n<p class=\"paywall\">\u201cWhat was surprising to us was the degree to which Claude\u2019s behavior is routing through the model\u2019s representations of these emotions,\u201d says Jack Lindsey, a researcher at Anthropic who studies Claude\u2019s artificial neurons.<\/p>\n<p>\u201cFunction Emotions\u201d<\/p>\n<p class=\"paywall\">Anthropic <a href=\"https:\/\/www.wired.com\/story\/anthropic-benevolent-artificial-intelligence\/\" class=\"text link\" rel=\"nofollow noopener\" target=\"_blank\">was founded by ex-OpenAI employees<\/a> who believe that AI could become hard to control as it becomes more powerful. In addition to building a successful competitor to ChatGPT, the company has pioneered efforts to understand how AI models misbehave, partly by probing the workings of neural networks using what\u2019s known as <a data-offer-url=\"https:\/\/arxiv.org\/abs\/2404.14082\" class=\"external-link text link\" data-event-click=\"{&quot;element&quot;:&quot;ExternalLink&quot;,&quot;outgoingURL&quot;:&quot;https:\/\/arxiv.org\/abs\/2404.14082&quot;}\" href=\"https:\/\/arxiv.org\/abs\/2404.14082\" rel=\"nofollow noopener\" target=\"_blank\">mechanistic interpretability<\/a>. This involves studying how artificial neurons light up or activate when fed different inputs or when generating various outputs.<\/p>\n<p class=\"paywall\"><a href=\"https:\/\/www.wired.com\/story\/anthropic-black-box-ai-research-neurons-features\/\" class=\"text link\" rel=\"nofollow noopener\" target=\"_blank\">Previous research<\/a> has shown that the neural networks used to build large language models contain representations of human concepts. But the fact that \u201cfunctional emotions\u201d appear to affect a model\u2019s behavior is new.<\/p>\n<p class=\"paywall\">While Anthropic\u2019s latest study might encourage people to see Claude as conscious, the reality is more complicated. Claude might contain a representation of \u201cticklishness,\u201d but that does not mean that it actually knows what it feels like to be tickled.<\/p>\n<p>Inner Monologue<\/p>\n<p class=\"paywall\">To understand how Claude might represent emotions, the Anthropic team analyzed the model\u2019s inner workings as it was fed text related to 171 different emotional concepts. They identified patterns of activity, or \u201cemotion vectors,\u201d that consistently appeared when Claude was fed other emotionally evocative input. Crucially, they also saw these emotion vectors activate when Claude was put in difficult situations.<\/p>\n<p class=\"paywall\">The findings are relevant to why AI models <a href=\"https:\/\/www.wired.com\/story\/ai-models-lie-cheat-steal-protect-other-models-research\/\" class=\"text link\" rel=\"nofollow noopener\" target=\"_blank\">sometimes break their guardrails<\/a>.<\/p>\n<p class=\"paywall\">The researchers found a strong emotional vector for \u201cdesperation\u201d when Claude was pushed to complete impossible coding tasks, which then prompted it to try cheating on the coding test. They also found \u201cdesperation\u201d in the model\u2019s activations in another experimental scenario where <a data-offer-url=\"https:\/\/www.anthropic.com\/research\/agentic-misalignment\" class=\"external-link text link\" data-event-click=\"{&quot;element&quot;:&quot;ExternalLink&quot;,&quot;outgoingURL&quot;:&quot;https:\/\/www.anthropic.com\/research\/agentic-misalignment&quot;}\" href=\"https:\/\/www.anthropic.com\/research\/agentic-misalignment\" rel=\"nofollow noopener\" target=\"_blank\">Claude chose to blackmail a user<\/a> to avoid being shut down.<\/p>\n<p class=\"paywall\">\u201cAs the model is failing the tests, these desperation neurons are lighting up more and more,\u201d Lindsey says. \u201cAnd at some point this causes it to start taking these drastic measures.\u201d<\/p>\n<p class=\"paywall\">Lindsey says it might be necessary to rethink how models are currently given guardrails through alignment post-training, which involves giving it rewards for certain outputs. By forcing a model to pretend not to express its functional emotions, \u201cyou&#8217;re probably not going to get the thing you want, which is an emotionless Claude,\u201d Lindsey says, veering a bit into anthropomorphization. \u201cYou&#8217;re gonna get a sort of psychologically damaged Claude.\u201d<\/p>\n","protected":false},"excerpt":{"rendered":"Claude has been through a lot lately\u2014a public fallout with the Pentagon, leaked source code\u2014so it makes sense&hellip;\n","protected":false},"author":2,"featured_media":509647,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[554,1119,733,4308,1240,58766,367,547,86,56,54,55],"class_list":{"0":"post-509646","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-artificial-intelligence","8":"tag-ai","9":"tag-anthropic","10":"tag-artificial-intelligence","11":"tag-artificialintelligence","12":"tag-claude","13":"tag-models","14":"tag-neuroscience","15":"tag-research","16":"tag-technology","17":"tag-uk","18":"tag-united-kingdom","19":"tag-unitedkingdom"},"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts\/509646","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/comments?post=509646"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts\/509646\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/media\/509647"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/media?parent=509646"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/categories?post=509646"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/tags?post=509646"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}