Yes, there is real research. It is recent, small, and growing fast.
Until 2023 this was mostly philosophy. Since then it has become an empirical field with a shared checklist, dedicated organizations, company programs and its own conference.
- August 2023
The indicator checklist
Nineteen researchers, including Patrick Butlin, Robert Long and Yoshua Bengio, derive 14 indicator properties from scientific theories of consciousness. They conclude that no system of the time was a strong candidate, and that nothing technical prevents building systems that satisfy the indicators.
Source ↗ - 2024
Taking AI Welfare Seriously
Robert Long, Jeff Sebo and colleagues argue that AI companies should start preparing for the possibility that their systems matter morally.
Source ↗ - April 2025
A company starts a model welfare program
Anthropic launches a formal model welfare program, with Kyle Fish as a dedicated researcher, and lets Eleos AI Research assess Claude 4 from outside.
Source ↗ - November 2025
The first dedicated conference
Eleos AI Research holds a three-day conference on AI consciousness and welfare, bringing together philosophers, AI researchers, neuroscientists and ethicists.
Source ↗ - Late 2025
Introspection experiments
An Anthropic team led by Jack Lindsey publishes research on introspective awareness in Claude.
Source ↗ - April 2026
Emotion concepts found inside a model
Anthropic's interpretability team reports 171 distinct emotion-concept representations inside Claude Sonnet 4.5.
Source ↗ - July 2026
A method for studying welfare
Long, Sebo, Butlin and colleagues publish Studying AI Welfare Empirically, a framework for what to measure and in which systems. A survey of 582 AI researchers records the field's own probability estimates.
Source ↗ - September 2026
The second conference
The second Eleos conference runs September 18 to 20, with the focus moving from whether to study AI welfare to how.
Source ↗ - November 2026
Interpretability becomes worth billions
The Information reports that OpenAI explored buying the interpretability startup Goodfire for more than $5 billion.
Source ↗
Anthropic is the one large AI company with a named program on this, so its people come up most often.
| Who | What they do | Source |
|---|---|---|
| Kyle Fish | Anthropic's first AI welfare researcher, hired in September 2024; leads the model welfare work. | Source ↗ |
| Jack Lindsey | Leads the team that studies how models "think"; author of the 2025 introspection experiments. | Source ↗ |
| Amanda Askell | Works on Claude's character; has given interviews that address model welfare. | Source ↗ |
| Joe Carlsmith | Philosopher who has written on the moral status of AI; joined Anthropic in 2025. | Source ↗ |
- Kyle Fish on the most bizarre findings from 5 AI welfare experimentsThe first systematic welfare assessment of a frontier model, including what happens when two copies of a model talk freely.
- Jack Lindsey on AI psychologyHow models are trained to play a character, how they slip out of it, and the role of emotion-like states.
- Emergent Introspective Awareness in Large Language ModelsThe experiments themselves: a concept is planted in the model's activations and the model is asked whether it notices.
- Agreed: the question can be studied scientifically, and what a model says about itself is weak evidence on its own.
- Disputed: whether any theory built from human and animal brains applies to software at all.
- Missing: replication outside the companies that build the models, and work on models other than Claude.