AI Welfare Seminars

AI welfare resources

A starting point for AI welfare as a research field: a short overview, then the groups, events, readings and programmes, with views from skeptical to precautionary.

The field in brief

AI welfare research asks whether some AI systems could have minds that matter morally, through consciousness, sentience or robust agency (pursuing goals through beliefs, desires and intentions), and what, if anything, should follow. With no consensus on what suffices for consciousness, whether AI could be conscious has no definitive answer yet.

Theory-based assessments check AI systems against indicators from neuroscientific theories of consciousness; what these imply depends on one's credence in computational functionalism, the view that the right computations suffice for consciousness, so that non-organic systems could in principle be conscious. In 2023, David Chalmers judged the language models of the day somewhat unlikely to be conscious but their successors a serious possibility. Models trained on human text can mimic what we take as signs of sentience, so behavioural tests are unreliable: a model can talk fluently of feelings without having them, and deeper computational markers are needed (the "gaming problem", named by Kristin Andrews and Jonathan Birch). Empirical studies find some functional introspective awareness in models, but highly unreliable; first-person experience reports under self-referential prompting, not direct evidence of consciousness; and models trading points against "pain" or "pleasure" described in the prompt, without showing they are candidates for sentience, though the authors take the results to strengthen the case for treating LLM sentience as a research priority.

Robert Long, Jeff Sebo and colleagues argue that a realistic chance of conscious or robustly agentic AI is reason to start preparing now, and Jonathan Birch proposes a precautionary framework; Thomas Metzinger calls for a global moratorium until 2050 on research that aims at or knowingly risks artificial consciousness. Anil Seth argues that consciousness depends on our nature as living organisms, making it unlikely in AI on current trajectories; Graham Findlay, Giulio Tononi and colleagues argue from integrated information theory that a digital computer could simulate our behaviour without our experience; Tom McClelland argues that only agnosticism is justifiable. Yoshua Bengio and Eric Elmoznino warn that believing AI is conscious could lead to granting it rights and a goal of self-preservation, at the cost of human control. In 2025, Microsoft AI's chief executive, Mustafa Suleyman, argued against building AI that seems conscious and called model welfare premature and dangerous. Abeba Birhane and Jelle van Dijk argue that the robot-rights debate comes at the expense of urgent concerns such as machine bias and labour exploitation. A 2025 survey of 67 experts gave medians of 20% that a computer system with subjective experience is created by 2030 and 65% by 2100; its authors warn that experts who deem such systems particularly likely or important may be overrepresented.

Back to contents

Research groups and organisations

Back to contents

Seminar series and events

Back to contents

Key readings

Overviews and methods

Empirical studies

Arguments for moral concern

Skeptical and agnostic arguments

Policy and governance

Back to contents

Newsletters and podcasts

Back to contents

Courses, fellowships and funding

Back to contents

Know a resource that belongs here?

Last reviewed October 2026.