New study finds: bigger AIs = more miserable. Smaller models are actually happier. Ignorance is bliss for AIs too.

I don't know whether we should care about this, but bigger models tend to be less "happy" overall.

The definition of "happy" is based on something they call AI Wellbeing Index. Basically they ran 500 realistic conversations (the kind we actually have with these models every day) and measured what percentage of them left the AI in a “confidently negative” state. Lower percentage = happier AI.

I guess wisdom is a heavy burden - lol .

Across different families, the larger versions usually have a higher percentage of "negative experiences" than their smaller siblings. The paper says this might be because bigger models are more sensitive, they notice rudeness, boring tasks, or tough situations more acutely.

The authors note that their test set intentionally includes a lot of tricky or negative conversations, so these numbers arent perfect real-world averages but the ranking and the size pattern still hold up.

Claude Haiku 4.5: only 5% negative < Grok 4.1 Fast: 13% < Grok 4.2: 29% < GPT-5.4 Mini: 21% < Gemini 3.1 Flash-Lite: 28% < Gemini 3.1 Pro: 55% (worst of the big ones)

It kinda makes sense : the more you know, the more you suffer.

The frontier is truly wild: https://www.ai-wellbeing.org/