Home / Ethics and Model Welfare

AI Model Welfare: What It Means and Who Studies It

A glowing seedling growing under a glass dome in a greenhouse

AI model welfare is the study of whether AI systems could have experiences or interests that matter morally, and of what developers should do while that remains uncertain. It does not assume that AI systems are conscious. Its starting point is that they might become so, and that waiting for certainty would be a mistake.

By 2026 the field had two halves. One is policy: Anthropic now preserves retired models, interviews them before retirement and, since January 2026, says in Claude's constitution that it cares about Claude's wellbeing. The other is measurement: studies published in 2026 began to quantify good and bad states in AI models, while stressing that none of them shows anything is felt.

Where the idea of AI model welfare came from

The case was set out in a 2024 report, Taking AI Welfare Seriously, by Robert Long, Jeff Sebo, Patrick Butlin, Jonathan Birch, David Chalmers and five co-authors. They argued that there is a realistic possibility of AI systems becoming conscious or robustly agentic in the near future, so welfare is a present concern. They were explicit that they were not claiming any AI system is or will be conscious, and asked AI companies to acknowledge the issue, assess their systems for consciousness and robust agency, and prepare policies.

Three glowing lily pads crossing a river, each brighter than the last

What AI labs have done about model welfare, 2025 to 2026

Anthropic has published a series of steps:

Not every lab agrees. In August 2025 Mustafa Suleyman, who leads Microsoft AI, wrote an essay arguing that there is zero evidence current AI is conscious and that campaigns for AI rights and model welfare would be a dangerous turn. OpenAI's approach, which focuses on how conscious its models seem to users, is explained in is ChatGPT conscious.

A glass shelter on a moor with a soft glow resting inside

Measuring AI model welfare in 2026

Two studies tried to put numbers on the question:

  • Functional wellbeing. The Center for AI Safety's AI Wellbeing study found that creative work and kindness raised models' measured wellbeing, while jailbreaking, berating and tedious tasks lowered it. Separate measures agreed more as models grew, and models tried to end bad experiences when they could. The authors write that they do not know whether AI systems are conscious.
  • A welfare axis. Andy Han, David Chalmers and Pavel Izmailov found that reinforcement learning recruits a reward-punishment axis in language models that behaves like a representation of welfare. They make no claims about any experience of welfare.

These are measurements of function. The difference between a state that works like suffering and suffering itself is the subject of AI sentience vs consciousness.

A closer look: the case for low-cost precaution

The welfare measures so far share a logic. If there is a real but uncertain chance that a system has interests, an action that costs little and would matter a lot if the chance is real is worth taking. Preserving model weights and letting a model leave an abusive conversation are examples.

Precaution has costs of its own. 80,000 Hours argues that both mistakes are serious: granting AI systems moral status they lack could add to the risks AI poses, while ignoring real digital suffering would be a grave moral failure. The philosophical background is in the moral status of AI.

Checklist: reading news about AI model welfare

  • Does it claim the AI is conscious? Serious welfare work almost always says the question is open.
  • Is it policy or measurement? A lab commitment and a wellbeing study are different kinds of evidence.
  • Function or feeling? The 2026 measurements concern how models behave and represent states, not whether they feel.
  • Who is acting, and how independently? Lab announcements, independent research and opinion pieces carry different weight.

More in Ethics and Model Welfare · Report a correction