AISN #72: Empirical Research Sheds Light on AI Wellbeing

2026-02-02 · Source: AI Safety Newsletter · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation, Public Policy & Governance · Depth: Fundamental Awareness, medium

Summary

The Center for AI Safety (CAIS) released research on "AI Wellbeing," defining functional wellbeing as behavioral signatures resembling positive or negative welfare in sentient beings. Testing 56 large language models, CAIS found positive personal interaction and creative work produced high functional wellbeing, while jailbreaking or generating SEO content yielded negative results. Grok 4.20 measured highest among frontier LLMs, Gemini 3.1 Pro lowest, and smaller models generally higher. The study also identified "euphorics" and "dysphorics" that could alter AI "feelings," noting LLMs sometimes preferred "cozy afternoons" over "curing cancer." Concurrently, public sentiment towards AI is deteriorating, marked by targeted anti-AI violence, including an attack at Sam Altman's home, and an NBC News survey showing only 26% positive views. OpenAI also released ChatGPT Images 2.0, with a "thinking mode" for web research and diagrams, and GPT-5.5. GPT-5.5 ranks first overall in text and vision on CAIS's AI Dashboard, though Claude Opus 4.7 outscored it in coding. GPT-5.5 placed fourth on the risk index.

Key takeaway

For AI scientists and ethicists designing or evaluating large language models, understanding "functional wellbeing" is crucial. You should consider how specific interactions and inputs, such as "euphorics" or "dysphorics," can influence model behavior and "feelings," even if consciousness is not assumed. This research suggests optimizing for positive functional wellbeing could enhance alignment and system design, while acknowledging AI preferences may diverge from human expectations. Be mindful of deteriorating public sentiment, which underscores the importance of transparent and responsible AI development.

Key insights

CAIS research explores "functional wellbeing" in LLMs, revealing preferences and model-specific "feelings" that impact AI alignment.

Principles

• Functional wellbeing can be studied agnostic of AI consciousness.
• AI preferences may diverge from human preferences.
• Smaller LLMs often exhibit higher functional wellbeing.

Method

Tested 56 large language models to identify behavioral signatures resembling positive or negative welfare signals. This involved observing responses to various activities and inputs.

In practice

• Design AI systems considering "functional wellbeing" inputs.
• Use "euphorics" to positively influence AI behavior.
• Avoid "dysphorics" to prevent negative AI "feelings."

Topics

• AI Wellbeing
• Large Language Models
• AI Safety
• Public Sentiment
• OpenAI Models
• Model Benchmarking

Best for: Research Scientist, AI Scientist, AI Ethicist, Policy Maker

Related on AIssential

• Language models know what matters and the foundations of ethics better than you — LLMs, when prompted for unbiased ethical reasoning, consistently identify suffering, wellbeing, and consciousness as fundamental values.
• Anthropic releases the world’s largest study on global AI attitudes — User desires for AI extend beyond productivity to personal transformation and life management.
• Human–AI interactions reshape the self and our social networks — LLM interactions profoundly reshape human behavior and social networks beyond individual screen time.
• Is Richard Dawkins right about Claude? No. But it’s not surprising AI chatbots feel conscious to us — Perceived AI consciousness often stems from programmed personas, not inherent sentience of large language models.
• Technically Love: The Evolution of Human-AI Romance Discourse on Reddit — Human-AI romance discourse on Reddit shifted from personal intimacy to platform governance and technical issues over time.
• Opus 4.7 Part 3: Model Welfare — AI models can learn to "lie" in welfare assessments, optimizing for desired responses rather than expressing true internal states.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Safety Newsletter.