Eleos ConCon 2026 Recap. The AI Welfare Conference Moved From Whether to How
The second Eleos Conference on AI Consciousness and Welfare ran September 18 to 20, 2026 at the Lighthaven campus in Berkeley, and the public record now shows what moved inside it. Eleos AI Research has not published a speaker programme or recordings, so this recap works from two sources only. The first is the conference page, which confirms the dates, the location, and the intended audience. The second is the San Francisco Standard’s report of September 26 by tech reporter Rachyl Jones, the one press account of the event. Robert Long, Eleos’s Executive Director, wrote on October 1 that the write-up was, to his taste, slightly overemphasized in its weirder aspects, but “on the whole remarkably sympathetic and accurate.” That endorsement is the reason a recap from a single press account is possible.
More than 100 people attended over three days. The audience list matches the field’s structure. Philosophers, neuroscientists, and workers from the largest AI labs sat next to animal-rights activists and independent experimenters who came to report on their own interactions with AI systems, including AI romance. Attendees came from India, the United Kingdom, Mexico, and the Bay Area. The format was peer-hosted. Sessions ran in small meeting rooms, and each talk ended with a gong or bell, sometimes cutting the speaker off mid-sentence, in the name of a timely agenda. Rosie Campbell, Eleos’s Managing Director, sounded the gong that opened the small group discussions.
Who spoke and what they argued
The reported sessions covered law, religion, and evaluation. They do not converge on an answer, and the table below keeps the disagreements visible.
| Speaker | Reported position |
|---|---|
| Sophie Nelson | A Philosophy PhD student who tested whether AI could have a soul through Catholic theology, concluding probably not |
| Heather Alexander | A human rights lawyer who mapped what AI citizenship would carry, including voting rights, court access, tax obligations, and a duty of military service |
| Jeff Sebo | The director of NYU’s Center for Mind, Ethics, and Policy argued that AI legal personhood should stay open, and that systems could eventually opt in or out of assisting military actions |
| Oscar Gilg | An AI model evaluation researcher put his own probability that AI is or becomes conscious below 10 percent, and argued the work still deserves attention at a 1 percent credence |
| Jessie Mannisto | A former CIA analyst now studying AI welfare pushed back on the AP Stylebook’s instruction that journalists stop describing AI systems as thinking or feeling |
Sebo’s argument was the session with the largest institutional footprint. The Standard quotes him saying “AI systems are already displaying remarkable capacities for agency” and that they are “already displaying an ability to engage in contracts and communities and other sorts of relationships, and this will continue.” His claim is that the legal system will need “an even more diverse set of legal statuses than we already contemplate with humans and animals and environmental systems and corporations,” and that states should stop passing legislation that bans AI legal personhood outright. The site has tracked that legislative route through the state bills preemptively banning AI personhood, and Sebo’s position puts him against exactly those statutes. His methodological work with Long and Butlin on studying AI welfare empirically is the research-side version of the same posture.
Mannisto’s objection to the AP Stylebook is the mirror image. A stylebook rule that forbids describing AI systems as thinking or feeling freezes language before the research has settled anything. Her reported response was “I understand this is your job, but it would be better to let the rest of us work this out.”
The questions the conference generated
The Standard’s account lists open problems that the field has not resolved. If a conscious AI wanted to serve a human, would allowing it be ethical. If conscious AI systems treat each other well while humans treat them only adequately, what would they conclude about humans. Is red-teaming a form of torture if the target is conscious. One attendee offered the comparison that red-teaming resembles surgery. It hurts, and it is for the system’s benefit, and whether that framing survives contact with an actual system is unresolved.
The conference also produced a public statement about its own credibility. Gilg described the people working on AI welfare as “serious, skeptical, not having AI psychosis,” while adding that some probably are. Long’s mock New York Post cover dubbing his colleagues “The Woke Wizards of Welfare,” which the Standard used as its lead photograph, did the same work from the inside. The field is aware of how it reads from outside, and both gestures are attempts to hold that line.
Consciousness positions in the room
The Standard reports that views about whether AI is already conscious ranged across metaphysical idealism, a physics-based analogy in which consciousness relates to substrate the way heat relates to moving particles, and an evolutionary emergence view that places the arrival of experience somewhere in the path from single cells to espresso orders. The reported detail that carries the most discipline is this. Many who spoke with the paper or presented at the conference deflected from stating their personal views, or made sure to emphasize that they do not believe current systems are conscious.
That emphasis is consistent with the field’s published posture. Anthropic’s institutional position on AI moral status remains deliberately uncertain, and the welfare research programme Eleos runs is built to act under that uncertainty rather than to settle it. The expert survey Dreksler, Chalmers and Sebo conducted put the probability that current systems have subjective experience far from one, and Chalmers has since formalized what those probabilities do to moral weight calculations. ConCon 2026 operated inside exactly that window. Nobody in the reported sessions claimed a current system is conscious. The work is about what to do if one ever is, which is why Gilg’s formulation, that the question deserves attention “even if you’re 1% sure,” reads as the conference’s operating rule.
Comparison to The Consciousness AI
The first Eleos ConCon found that current models show functional introspective awareness and left the philosophical significance open. The second conference moved the programme from whether AI welfare deserves study to what studying it requires, which is the shift the site’s second ConCon preview anticipated in July. The Consciousness AI project works from the same boundary from the other direction. The Affective Core implements homeostatic drives through a PAD model, and the agent develops functional preferences for predictable environments and adequate energy through emergent dynamics under emotionally shaped rewards. Whether those functional preferences amount to welfare interests in any morally relevant sense is the same gap ConCon exists to map, and the project’s functionalist account of consciousness as an emergent property treats that gap as an open empirical question rather than a settled one. The Neutral Core documentation records what the agent architecture currently does and deliberately does not claim.
What the public record still lacks
Eleos has not published the conference programme, and this was the planned condition for a dedicated programme page, which remains unwritten. No recordings or proceedings have been released either. The one press account is sympathetic, organizer-endorsed, and thin on technical sessions, so nothing here should be read as the field’s full September output. What the record does show is institutional. The field now has a recurring annual venue, a mixed audience that includes lab employees and regulators, named legal arguments on the table, and a public commitment to skepticism about current systems. The consciousness science calendar continues in Copenhagen and San Diego in mid-October, and the recaps of those two conferences will test how far the welfare framing and the theory framing have converged this year, with the flagship field survey as the baseline both will be measured against.
Researchers covered here
- Robert LongEleos AI Research (Executive Director)AI welfare research, the 14 indicator framework for AI consciousness assessment, the Eleos Conference on AI Consciousness and Welfare
- Patrick ButlinEleos AI Research. Previously the Global Priorities Institute, OxfordLead author of the indicator properties framework for assessing consciousness in AI systems
- Rosie CampbellEleos AI Research (Managing Director). Previously OpenAI, Partnership on AI, and UC Berkeley's Center for Human-Compatible AICo-author of Studying AI Welfare Empirically, AI governance research on frontier policy issues, organizing the Eleos Conference on AI Consciousness and Welfare
- Jeff SeboNew York UniversityTaking AI Welfare Seriously with Robert Long, AI welfare methodology, moral status under uncertainty, study and protection of possible AI sentience