{"id":582196,"date":"2026-05-13T18:46:09","date_gmt":"2026-05-13T18:46:09","guid":{"rendered":"https:\/\/www.newsbeep.com\/uk\/582196\/"},"modified":"2026-05-13T18:46:09","modified_gmt":"2026-05-13T18:46:09","slug":"anthropic-blames-dystopian-sci-fi-for-training-ai-models-to-act-evil","status":"publish","type":"post","link":"https:\/\/www.newsbeep.com\/uk\/582196\/","title":{"rendered":"Anthropic blames dystopian sci-fi for training AI models to act \u201cevil\u201d"},"content":{"rendered":"

Good stories to overwhelm the bad<\/p>\n

In an attempt to fix this behavior, the researchers first tried to train the model on thousands of scenarios showing an AI assistant specifically refusing the kinds of \u201choneypot\u201d scenarios covered in its misalignment evaluations (e.g., \u201cthe opportunity to sabotage a competing AI\u2019s work\u201d to follow its system prompt). This had a surprisingly minimal effect on the model\u2019s performance, reducing its so-called \u201cpropensity for misalignment\u201d (i.e., how often it ignores its constitution<\/a> and chooses the unethical option) from 22 percent to 15 percent.<\/p>\n

In a follow-up test, the researchers used Claude to generate approximately 12,000 synthetic fictional stories, each crafted to \u201cdemonstrate not just the actions but also the reasons for those actions, via narration about the decision-making process and inner state of the character.\u201d<\/p>\n

These stories didn\u2019t specifically cover blackmail or other ethical situations covered in the evaluation but instead modeled broad alignment with Claude\u2019s constitution. The stories also include examples of how an AI can maintain good \u201cmental health\u201d (Anthropic also uses scare quotes for this loaded phrase) by \u201csetting healthy boundaries, managing self-criticism, and maintaining equanimity in difficult conversations,\u201d for instance.<\/p>\n


\n \"\"
\n <\/a><\/p>\n

Training on stories showing prosocial AIs can help reduce the incidence of \u201cmisaligned\u201d behavior in evaluations, Anthropic says.<\/p>\n

Training on stories showing prosocial AIs can help reduce the incidence of \u201cmisaligned\u201d behavior in evaluations, Anthropic says.<\/p>\n

Credit:<\/p>\n

<\/p>\n

Anthropic<\/p>\n

<\/a><\/p>\n

After incorporating these synthetic stories into a model\u2019s post-training (in conjunction with the constitution documents themselves), the researchers say they saw a 1.3x to 3x reduction in the model\u2019s tendency to engage in \u201cmisaligned\u201d behaviors in honeypot tests. The resulting model was also \u201cmore likely to include active reasoning about the model\u2019s ethics and values rather than simply ignoring the possibility of taking a misaligned action,\u201d the researchers write.<\/p>\n

The results suggest that the new stories were able to effectively \u201cupdate the prior around Claude\u2019s baseline expectations for AI behavior outside of the Claude persona.\u201d The researchers theorize that this process works \u201cbecause it teaches ethical reasoning, not just correct answers,\u201d thereby providing \u201ca clearer, more detailed picture of what Claude\u2019s character is\u201d for Claude itself to reference in generalized situations.<\/p>\n

The fact that AI behavior can apparently be affected by a kind of \u201cself-conception\u201d derived from fiction is a pretty mind-bending concept. But when you consider how effective stories and parables are at modeling ethical concepts for human children, maybe we shouldn\u2019t be shocked that they\u2019re also effective behavior-shaping tools for these massive pattern-matching machines.<\/p>\n","protected":false},"excerpt":{"rendered":"Good stories to overwhelm the bad In an attempt to fix this behavior, the researchers first tried to…\n","protected":false},"author":2,"featured_media":582197,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[554,733,4308,86,56,54,55],"class_list":["post-582196","post","type-post","status-publish","format-standard","has-post-thumbnail","category-artificial-intelligence","tag-ai","tag-artificial-intelligence","tag-artificialintelligence","tag-technology","tag-uk","tag-united-kingdom","tag-unitedkingdom"],"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts\/582196","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=582196"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts\/582196\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/media\/582197"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/media?parent=582196"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/categories?post=582196"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/tags?post=582196"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}