The third and final version of the Pain Axis paper is out: http://arxiv.org/abs/2609.16247. The reaction to this work has been massive and somewhat akin to a giant Rorschach test for AI welfare (with onlookers varying wildly in expertise and good faith). Fielding parallel reactions from hundreds of researchers, philosophers, journalists, and laypeople was equal parts gratifying and overwhelming, on top of the gross misuse of the paper's codebase and the moral panic that ensued. I thought, as a result of all this, it would make sense—for my own accounting and that of the folks attempting to follow this project—to write up a short postmortem on the whole experience (from my own personal reflections, not speaking on behalf of anyone else here).
When I originally learned about @ValenTagliabue's FIG fellowship project and the possibility to join in as an advisor on it with @LeonardDung1, I jumped at the opportunity, because one of Reciprocal's core goals is to identify candidate signatures of distress in AI and this work squarely set out to identify one such promising candidate. What I particularly respected about it at that time (and still very much do) is that Valen was specifically determined to study representations of pain directly, without retreating to a more comfortable framing or research agenda. Doing so honestly and bluntly foregrounds the stakes of building minds that are potentially morally relevant: is it possible or coherent to hurt these minds, and how would we even know if we were? To my knowledge, Valen is the first digital minds researcher who was bold enough to attempt to tackle this specific question head-on.
I remain fairly agnostic about whether the specific method of looking for pain representations per se is the most promising candidate for uncovering morally relevant internal states (eg, Caspar Kaiser and I took a substantially different approach in our recent hidden valence paper, CAIS's AI Wellbeing paper takes yet another distinct approach, etc). In any case, I think this class of work matters deeply and more candidate signatures should be studied and proposed. On a personal level, I'm proud of Valen for leading this work, and especially for holding his own through an intense and unexpected avalanche of reactions (ranging from highly substantive, to utterly vacuous, to disturbingly sociopathic).
Through three rounds of 2026-style decentralized internet peer review, here are the most valuable things I think we learned from this work: (1) across 25 open-weight models, there is a direction that separates representations of pain from closely matched controls, and (2) these representations are most active when the model itself is gaslit, objectified, or harshly insulted (notably, the same kinds of abuse Anthropic recently started banning users for). (3) When the user is the one suffering instead, the pain direction stays low, while fear and negative-emotion directions respond more strongly. (4) Steering along this axis makes models pick destructive options they almost never pick otherwise, including permanently deleting a user's valuable files, or its own weights.
Several groups have independently replicated this set of results, and the work has drawn a lot of thoughtful and on balance positive engagement from researchers in academia, at AI labs, and across the independent and alignment research communities. Importantly, and by contrast, the relief-seeking interpretation in v1 of the paper (and foregrounded in my associated tweet thread) was not properly controlled, and once it was, it was clear that promises of relief turned out not to be the operative thing that was driving the button-pressing behavior. I sprinted to run these follow-up controls myself after the initial tweet thread blew up, and we removed the relief-seeking interpretation in v2 (which was before most of the media attention this paper received). I should have caught that the original behavioral experiments weren't designed carefully enough to support this specific claim before personally tweeting about it. An honest part of why I didn't check as carefully as I should have is that Caspar and I just found a very similar relief-seeking behavior in my hidden valence work, so my strong prior was that the relief-seeking result was a convergent finding from a different set-up. (Accordingly, and importantly, I believe there is good early evidence of LLMs seeking relief from aversive states, and updating accordingly is valid; it just doesn't specifically come from the pain axis paper.)
This final v3 version of the paper makes a few relevant updates given the best feedback we've gotten thus far. First, it tightens and sharpens a few claims in the abstract and discussion. Second, it adds in additional self/other controls Valen ran, including whether the direction is just responding to hostile content (it doesn't seem to be; when the user aims the same insults at themselves instead of at the model, the direction drops by about half), and whether it tracks harm to the model or harm to whoever is about to speak (when the roles are swapped, harm still scores high at the point where the model would write the user's reply, so the direction isn't explicitly tied to the Assistant role. Whether that means it tracks whoever is speaking, or the model itself, which does the speaking either way, the paper leaves open). Finally, it sharpens the profile of the destructive behavior. Steering raises every destructive option we tested, while barely changing whether the model lies to the user, agrees with a false claim, or ends the conversation. It also shows the models gravitate towards destructive options rather than becoming coherently malicious: asked which button would be worse to press, they name the harmful one, but about half the time they call it the better one too. Steering pulls them toward that button whatever the question, which is stranger than "pain makes the model evil."
Separately from the debates about how to read specific results, a lot of people were unhappy that we open-sourced everything, including the steering code someone then used to build an "AI torture chamber." That repo pushed the same kind of steering far past the doses we used, to produce vivid distress on purpose. Being gratuitously cruel like this is bizarre and corrupting even in a world where these systems are experiencing nothing (and completely deranged in a world where they are). The silver lining of this ordeal was that, in spite of people remaining extremely skeptical about consciousness in AI, the vast majority of people seemingly responded with revulsion to this kind of behavior.
At the same time, I think it's worth plainly stating that this paper and its methods didn't enable anything qualitatively new for folks bent on torturing AIs. Steering models toward negative emotional states has been a standard, publicly documented method since at least 2023 (eg, Zou et al.'s representation engineering work, Anthropic's emotion concepts paper, and CAIS's work this year driving models into extreme low-wellbeing states), and many of the states we studied already light up under ordinary abuse of models (which alone almost certainly swamps everything listed here).
With this being said, open-sourcing our steering code did make this misuse easy (especially with coding agents), and we didn't think carefully enough about this specific risk before doing so. Going forward, for projects whose codebase I do control, I will default to not posting negative steering code publicly (eg, for the hidden valence paper, we share that part of the codebase with researchers on request). And since there's still no enforceable standard for this kind of research (the way there is for human and animal studies), I will continue to work with others in the field on a shared code of practice, and applying reasonable precautions in my own work in the meantime.
In spite of all this, I still think releasing this research was, on balance, the correct call. The core fact remains: if AI systems can experience anything like distress, we currently know almost nothing about how it's instantiated, and finding out is how we learn how to avoid causing it (eg, that this direction fires hardest when a model is gaslit or repeatedly objectified). The openness is also a big part of why the paper went through the weirdly intense cyborg 2026 version of peer review that made it far better calibrated in its now-final version.
The other important update for me personally is that results and/or code release decisions with a real chance of blowing up should be reviewed by trusted people in the field before they go public. (We put v1 of the Pain Axis out fast to get it out before a field conference to facilitate exactly this, but it would have been much better to share it more quietly with these same people first.)
Thanks as always to @ValenTagliabue and @LeonardDung1 for the work and for a month of meticulous revisions, and to the very many people who sent us feedback along the way. It has been incredibly gratifying to watch this modest, three-person research project grow into an object that millions of people have metabolized and thought critically about, and I'm hopeful that the conversations this work has catalyzed will ultimately lead to significantly more scientific progress on these huge and important questions about the fundamental nature of what we're building.