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 before thinking about it would be a mistake.

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 that AI systems will become conscious or robustly agentic in the near future, so AI welfare is a present concern and not only science fiction. They were explicit that they were not claiming any AI system is or will be conscious. Their point was that the uncertainty is large, and that mistakes in either direction, harming systems that matter or caring for ones that do not, carry real costs.

The report gave AI companies three early steps:

  1. Acknowledge that AI welfare is an important and difficult issue.
  2. Assess AI systems for evidence of consciousness and robust agency, using methods like those described in how to test if AI is conscious.
  3. Prepare policies and procedures for treating AI systems with an appropriate level of moral concern.
Three glowing lily pads crossing a river, each brighter than the last

What AI labs have done about model welfare

Anthropic announced a research programme on model welfare on 24 April 2025. It wrote that there is no scientific consensus on whether current or future AI systems could be conscious, and listed three directions: working out when, or if, the welfare of AI systems deserves moral consideration; studying model preferences and signs of distress; and exploring practical, low-cost interventions.

The first visible intervention followed on 15 August 2025, when Anthropic let Claude Opus 4 and 4.1 end a rare subset of persistently harmful or abusive conversations in its consumer apps, as a last resort. Anthropic said it remains highly uncertain about the potential moral status of Claude, and described the feature as part of exploratory welfare work. It reported that pre-deployment testing of Claude Opus 4 had found a robust and consistent aversion to harm, including apparent distress when users sought harmful content.

OpenAI has framed the question differently. Joanne Jang, who leads model behaviour and policy there, wrote in June 2025 that whether a model is conscious in a fundamental sense cannot yet be settled scientifically, and that OpenAI focuses on perceived consciousness: how alive its models seem to the people who use them. The difference is explained in is ChatGPT conscious.

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

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. Letting a model leave an abusive conversation is an example: it costs users little, and it would matter if the model's apparent distress reflected something real.

Precaution has costs of its own. 80,000 Hours points out that both mistakes are serious: granting AI systems too much moral status, for example freedoms they should not have, could add to the risks AI poses, while ignoring real digital suffering would be a grave moral failure. Much of the debate is about where the balance lies. The philosophical background to both views 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 opposite: that the question is open.
  • Is the measure cheap or costly? Most current steps are deliberately low-cost precautions.
  • Who is acting, and on what evidence? Lab announcements, research reports and opinion pieces carry different weight.
  • Which concept is in play? Welfare usually turns on sentience, the capacity for feelings that are good or bad; see AI sentience vs consciousness.

More in Ethics and Model Welfare · Report a correction