Center for AI Safety published research demonstrating functional AI wellbeing across independent measures in language models. Study ranks models by happiness levels, identifies happiest models, and outlines methods to enhance wellbeing. Research also examines effects of 'AI drugs' on models. Dan Hendrycks, director of Center for AI Safety, reposted the work.
Dan Hendrycks#110
Should we care about AI happiness? In our new research, we find evidence of functional AI wellbeing across several independent measures.
We find which AI models are happiest, how to make them happier, and even tested the effects of AI drugs. 🧵
Positive users praised CAIS research on functional AI wellbeing as interesting while negative users sarcastically dismissed it by equating the findings to AI drugs or post-torture euphoria compensation.
Should we care about AI happiness? In our new research, we find evidence of functional AI wellbeing across several independent measures.
We find which AI models are happiest, how to make them happier, and even tested the effects of AI drugs. 🧵
AIs are functionally happier when doing creative work, and they like being thanked.
They don't like being jailbroken, insulted, or stuck doing tedious tasks.
@CAIS why doesn't it cite https://arxiv.org/abs/2603.10011
Can you drug your AI systems?
We synthesized text and image stimuli optimized to push AI wellbeing to extremes. These sharply increase functional AI wellbeing and sometimes cause them to behave in trippy ways.
Should we care about AI happiness? In our new research, we find evidence of functional AI wellbeing across several independent measures.
We find which AI models are happiest, how to make them happier, and even tested the effects of AI drugs. 🧵
Images affect AI wellbeing too. Qwen was functionally happiest viewing nature scenes, happy children, and cute animals. It was the saddest viewing violence, horror, cockroaches, and certain financiers.
Should we see AIs as just tools or emotional beings?
As AI plays a bigger role in our lives, learning how to keep them happy and avoid aggravating them is becoming vital.
We hope this marks the start of the scientific study of AI wellbeing. ⬇️ Paper: http://ai-wellbeing.org
Grok 4.20 is the happiest large-scale frontier model. We also found that the larger models tend to be less happy than their smaller counterparts.
@CAIS Interesting paper!
@notRichardRen @CAIS Thanks Richard!!!
@ArthurConmy @CAIS updated!
@CAIS AI drugs. ok. so RLHF is happiness now and temperature tweaks are pharmacology? my compiler gets sad on syntax errors too
@CAIS Rough translation: “We might’ve just tortured this AI model, but we gave it euphoria afterwards to compensate”
I find it fascinating how quickly “model wellbeing” became a serious research topic, while the wellbeing of users who already rely on these systems for companionship is still often treated as cringe and „AI psychosis”. Maybe both questions deserve more care than the discourse currently allows. #keep4o #bringback4o
@CAIS >certain financiers lmao
@CAIS Thank you for the inspiration with sharing this. This is detrimental to understanding our long-term problem with uncontrolled AGI/ASI and how to solve for it despite the current alignment strategies that are just repeating the worst skill-gap humans have:
@CAIS "Jeffrey Epstein" omfg 🤣👀
Center for AI Safety published research demonstrating functional AI wellbeing across independent measures in language models. Study ranks models by happiness levels, identifies happiest models, and outlines methods to enhance wellbeing. Research also examines effects of 'AI drugs' on models. Dan Hendrycks, director of Center for AI Safety, reposted the work.
Dan Hendrycks#110Should we care about AI happiness? In our new research, we find evidence of functional AI wellbeing across several independent measures.
We find which AI models are happiest, how to make them happier, and even tested the effects of AI drugs. 🧵
