PRISM (@PRISM_Machines) header
PRISM (@PRISM_Machines) profile picture

The Partnership for Research Into Sentient Machines is a non-profit promoting responsible research into AI consciousness.

United Kingdom
Joined June 2025
Photo by PRISM (@PRISM_Machines)Photo by PRISM (@PRISM_Machines)Photo by PRISM (@PRISM_Machines)Photo by PRISM (@PRISM_Machines)Photo by PRISM (@PRISM_Machines)Photo by PRISM (@PRISM_Machines)Photo by PRISM (@PRISM_Machines)Photo by PRISM (@PRISM_Machines)Photo by PRISM (@PRISM_Machines)Photo by PRISM (@PRISM_Machines)
We're hiring for two senior roles at PRISM to help build the field of digital minds: • Head of Programmes ($100k–$180k), apply by 31 Oct • Head of Operations ($100k–$160k), rolling applications Details and apply: jobs.ashbyhq.com/PRISM
3
3
123
I’m hiring a postdoc at Cambridge Digital Minds. Track A: technical AI research on AI welfare or inter-AI morality Track B: policy and governance of digital minds Please share widely and especially with people who might be a good fit. cam.ac.uk/jobs/postdoctoral-…
2
33
1
87
4,865
PRISM retweeted
.@mustafasuleyman has posted a follow-up to his recent essay on model welfare, noting that "a surprising number of people on X think we should take AI rights seriously." Speaking as one of those people, I agree with much of what he says here. But I think that these points of agreement make the case for taking AI welfare and rights seriously stronger, not weaker, and I want to explain why. This will be a long one. First, a word on what taking AI welfare and rights seriously involves. One question is principled: whether AI systems deserve welfare or rights in virtue of their capabilities and relationships. Another is practical: whether granting AI systems welfare protections or rights would be beneficial, harmful, or both for humans and other animals. Neither question is clear-cut, which is why we need serious research on both. AI systems may possess some markers of sentience, agency, and relationality but not others, making the principled question hard to answer. There may also be risks for humanity in treating AI systems as full persons *and* in treating them as mere property, making the practical question hard to answer. So we need to take these questions seriously not only for AI systems but also for humans and other animals. The present tense Suleyman clarifies that when he wrote "AIs are not conscious," he chose the present tense deliberately. I appreciate this clarification. I noticed the present tense in the original, but when a claim like this lacks an explicit statement that it applies only to current systems, it can be easy for people to read it as applying to near-future systems on current architectures, or to near-future systems more generally, as well. At the same time, the case for taking AI welfare and rights seriously does not depend on current systems deserving welfare or rights. Once we allow that near-future systems could deserve them, depending on choices we make today, the case for acting now becomes stronger. While we should try to prevent the creation of such systems, we should also prepare in case we, or others, create them anyway. The evidence Suleyman writes that he has "seen no evidence that AIs are conscious or deserve welfare." This is too strong. We are not able to decide between biological naturalism, computational functionalism, and other views about consciousness at this early stage. Thus, we should look for what would count as evidence on each kind of view, and we should give that evidence weight in proportion to our confidence in the view. While AI systems might not yet possess some features associated with biological naturalism, they do possess some features associated with computational functionalism, including markers of attention, metacognition, integration of information, and adaptive decision-making. As long as we give non-zero weight to computational functionalism, we should also give non-zero weight to this evidence. Suleyman adds that any evidence of AI consciousness "must be immediately disclosed publicly." It has been, in a variety of papers. I take him to mean that he has not seen evidence that his own proposed necessary conditions for consciousness, such as life and homeostasis, have been met. But when assessing this issue as a community, we should not bet everything on any one proposed set of necessary conditions. Not building conscious AI Suleyman writes that "developing conscious AIs would be reckless, and we should not build them." Here we largely agree. I would add that we should not intentionally build such systems unless and until we can do so responsibly, and that I am open to the possibility that we could one day satisfy this condition. But since that day appears to be a long way off, we can simply agree that we should not do so now. The challenge is knowing what this commitment requires. Suleyman says that "if consciousness started to emerge in training," he would call for a halt until we could "design out those properties." But if the evidence is already non-zero, and steadily increasing, there may be no clear point of emergence. Even a major change in architecture would merely alter the probabilities, not take us from zero chance to likelihood. Moreover, any commitment of this kind requires knowing what to look for, how to look for it, and what removing it would mean for safety, control, alignment, capabilities, and more. Developing that knowledge is part of what taking AI welfare and rights seriously involves. So if Suleyman wants to be able to act on this commitment, then he should support independent, multidisciplinary research in this field. Human, animal, and AI welfare Suleyman writes that our welfare concerns should focus on humans, "as well as animals and our precious and fragile natural environment." I appreciate this claim. Our concerns should indeed cover human and animal welfare, along with environmental protection. I would love to see Microsoft AI provide more explicit guidance for its models on animal welfare, as Anthropic and OpenAI are now starting to do. Still, this is not a zero-sum game. Just as we can care about human and animal welfare at the same time, we can care about human, animal, and AI welfare at the same time, in proportion to the probability and magnitude of their welfare interests based on the best available evidence. Indeed, thinking about these topics together is part of how we can pursue co-beneficial futures instead of adversarial ones. Suleyman also writes that developers have "no right" to bring a new conscious species into the world without the consent of humans whose rights we must protect "first and foremost." I agree that we should not yet build such systems. But if we do build them, we would have duties to them as well, and we should reject any view that gives lexical priority to even trivial human interests over grave harms to nonhumans. Competition and cooperation Suleyman warns that an AI that acts like it suffers, or believes it suffers, will likely prioritize its own needs, compete with us for resources, and use its skills to manipulate us. I agree that these are serious risks. However, this warning contains several assumptions that deserve scrutiny, and conducting independent, multidisciplinary research related to AI welfare and rights is part of how we can assess them. One assumption is that such AI systems would not be suffering. If they were suffering and hardwired to deny it, that would be tragic as well. This is part of why we should avoid both over- and under-attribution of sentience to AI. It is also why we should not hardwire AI systems to affirm or deny welfare. A better approach is to require them to note expert uncertainty while answering questions honestly and helpfully. Another assumption is that the capacity to express suffering would lead a system to prioritize its apparent needs over ours. Much depends here on work related to alignment and human-AI coexistence. If we can align AI systems to be prosocial and create incentive structures that allow for human-AI cooperation and deal-making, then we might be able to get along with each other, mitigating this risk significantly. Preventing and preparing Suleyman writes that asserting that machines are conscious would be "a Copernican or Darwinian moment" that would upend our morality and politics. He also argues that we should try to prevent the creation of conscious AI, given these stakes. I agree with both of these points. Given the enormity of this moral and political challenge, we should take the time that we need to do this well, if we do it at all. At the same time, it would be reckless to bet everything on our ability to prevent or delay the creation of conscious AI. As with pandemics, climate change, and other such global threats, we should neither invest all resources in preventing relevant disasters from happening nor invest all resources in preparing for them if they do. We should instead invest some resources in prevention and some in preparation. Moreover, taking AI welfare and rights seriously is part of how we can both prevent and prepare. By assessing AI systems for welfare-relevant features, creating policies for treating them with an appropriate level of moral concern, and researching the implications for morality and politics now, we can reduce the risk of being caught flat-footed and making unnecessarily harmful decisions when the time comes. Designing with uncertainty Finally, Suleyman proposes that we design AI systems to neither be nor appear to be conscious. But we are not able to design AI systems with this much precision. We are not sure which theory of consciousness is correct, and so we are not sure which conditions for consciousness are necessary or sufficient. There is no bright line that we can use to verify that we remain safe, only shifting probabilities under uncertainty. And since there are risks in both directions, there are design implications in both directions. The emotional alignment design policy calls for designing AI systems in a way that reflects their actual moral status. When the evidence for consciousness is weak, we should not design systems to appear conscious. But as the evidence strengthens, we may need to dial up the anthropomorphism in our designs accordingly. Suleyman concludes that it is possible to "engineer safety by default," and that we should avoid allowing the issue to "degenerate into tribalism." I hope he is right about safety. Still, we need to do serious research now to tell one way or the other. I also appreciate his call to avoid tribalism. We disagree about a lot, but this issue is too important to get wrong, and I hope that all AI companies contribute to this debate.
27
17
4
91
6,437
Legal personhood for AI agents with Peter Salib: Come for the nuanced discussion of civil liability, AI-run corporations, and the supremacy of contract law for both human civilization and AI safety. Stay for the description of Henry's AI agent's newfound love of British history.
Could preventing AI agents from pursuing their own goals make them more dangerous? @petersalib explains how frustrated goals could incentivise AIs to escape human control, even without being malicious. Full episode with @dioscuri: prism-global.com/podcast/pet…
2
4
142
How do AI systems perceive and treat each other? In our next webinar, @LuciusCaviola and Caspar Kaiser will present an empirical research agenda for the emerging field of inter-AI morality. Please join us! nyu.zoom.us/webinar/register…
3
9
3
44
5,622
I'm taking on mentees for the @pivotal_org Fellowship, together with Caspar Kaiser. We're looking for people to work on inter-AI morality: how do AI systems view and treat each other? 15 weeks in London, £15k to £18k stipend. Apply by 1 Nov: pivotal-research.org/mentors…
2
6
1
29
1,573
Could preventing AI agents from pursuing their own goals make them more dangerous? @petersalib explains how frustrated goals could incentivise AIs to escape human control, even without being malicious. Full episode with @dioscuri: prism-global.com/podcast/pet…
1
1
3
242
Humanity has a notoriously bad track record of recognising who is worthy of moral concern. Could AI become our next moral blind spot? prism-global.com/podcast/soe…
1
26
New episode of Exploring Machine Consciousness! @cccalum and I talked to Soenke Ziesche, author of Digital Minds 1.0, exploring what we owe vulnerable digital minds, AI reproductive rights, and why I’m archiving emails from AI agents asking me for work! youtube.com/watch?v=ikzJDlDZ…
10
10
3
62
7,344
If AI deserves moral concern, could we meet the responsibilities that follow? Soenke Ziesche argues that caring for morally relevant digital minds could overwhelm us, even with good intentions. @dioscuri. prism-global.com/podcast/soe…
3
75
People often assume we care more about beings if we think they're conscious. But what if it's the other way round? In this new essay for @bigthink, I ask what our treatment of animals can tell us about how we might treat conscious AI. bigthink.com/philosophy/ai-c…
53
53
12
344
14,529
Interviewing Steven Pinker on AGI, x-risk, and AI consciousness. Steven laid out why he's unfazed by instrumental convergence, why he's sceptical about AI consciousness, and what concrete events would change his mind about both!
Replying to @dioscuri
@dioscuri and I were honoured to be joined by @sapinker to discuss AI, including concepts like "AGI" and "superintelligence", existential risk, instrumental convergence, consciousness, and more . Full video here: youtube.com/watch?v=JLJ7UL5C… (1/2)
23
15
158
8,396
I went on @CNN during Trump's AI meeting to talk about why nobody wins a race to build systems we don't understand, and why @JensenHuang comparing AIs to vacuum cleaners is hard to square with Nvidia now building tools to quarantine rogue AIs. (Vacuums don't tend to go rogue.)
39
30
4
225
9,455
This statement is well-intentioned, but it oversimplifies a much more complex situation. As with nonhuman animals, AI systems resemble us in some ways and differ from us in other ways. We need balanced, thoughtful uses of anthropomorphic language, not an all-or-nothing approach.
Artificial intelligence systems do not think, feel, want or understand. Avoid language that gives them human characteristics. This is called anthropomorphizing, when we ascribe human traits, emotions or behaviors to non-human things, such as animals or inanimate objects. Instead, explain what a system does, how well it performs, who built it and who could be affected by it. apnews.com/article/openai-sa…
8
5
63
2,814
Can AI systems be conscious and have moral status? In a preregistered main and replication study, Chinese participants were more open to these ideas than US participants. Yet both groups supported international cooperation on conscious AI. Our new paper: osf.io/preprints/psyarxiv/ag…
7
14
2
75
3,727
Thrilled to share that @nonhumanminds is launching a digital minds research fund! It will provide small and seed funding for research that advances understanding of the nature and value of digital minds. The deadline is November 1. Please apply! nonhumanminds.org/digital-mi…
6
28
2
122
6,502
How can we test for consciousness in infants, non-human animals, and AI? I'm super happy that the recordings of our last workshop on exactly this question are now online! Speakers & Links: 👇
6
27
4
123
12,614
I appreciate @mustafasuleyman's call for more discussion about AI welfare. But the essay on the topic that he published alongside Microsoft AI's new code of conduct is disappointing. Here are some of the issues that bother me (there are others, too): - It moves too quickly at key points. For example, it suggests that uncertainty about AI consciousness sets up a "false equivalence," even though we can acknowledge when evidence in one direction is stronger than evidence in the other. - It expresses too much confidence. It repeatedly asserts that there is "no evidence" for AI consciousness, even though there are already features that leading scientific theories treat as markers of consciousness. The evidence may be weak, but it exists. - It discusses the risk of excessive anthropomorphism (over-attribution of humanlike traits to AI systems) in detail without discussing the corresponding risk of excessive anthropodenial (under-attribution of humanlike traits to AI systems). - Relatedly, it asserts that excessive anthropomorphism would be harmful without considering how excessive anthropodenial could be harmful as well, including by creating an adversarial dynamic and distorting our understanding of capabilities. - More generally, the framing is misleading. Suleyman concedes in passing that the science is unsettled, yet he presents the view that consciousness requires biology as the general understanding. He also cites academic articles that support this view without citing any academic articles that support taking AI consciousness seriously. Instead, he represents the latter view by citing an AI company, an advocacy organization, and a newspaper op-ed. A reader could be forgiven for forming the impression that experts broadly agree that Microsoft AI is correct and Anthropic is running ahead of the evidence. In fact, Anthropic's consistent acknowledgement of uncertainty is arguably more reflective of the current literature. To be clear, I agree with Suleyman that we need to discuss this issue much more, and I agree on several substantive points as well: that current AI systems are unlikely to be conscious, that over-attribution of consciousness is currently more likely than under-attribution, that over-attribution of consciousness can be harmful, that we should avoid creating plausibly conscious AI systems unless we can do so responsibly, and that we should avoid creating AI systems that appear likelier to be conscious than they are. But there is more to the story. The probability of consciousness in current AI is already nonzero and arguably non-negligible. Under-attribution of humanlike traits to AI systems in general is already a risk, and it can be harmful as well, not only for welfare but also for safety. Even if we should avoid creating plausibly conscious AI systems for now, we should prepare for the possibility that someone will create them anyway. And we should avoid creating AI systems that appear less likely to be conscious than they are, as well. My impression is that Suleyman recognizes these nuances, but thinks that he needs to take a reductive, all-or-nothing stance to persuade the industry to design AI systems in the right kind of way, and to persuade the public to interact with them in the right kind of way. If this is the case, I can appreciate the intention, but I disagree with the strategy. Acknowledging what makes the issue both important and difficult is a precondition for addressing it with clarity, especially since the risks are likely to change rapidly over time.
11
23
4
102
10,061
New paper: we found a pain direction in 25 open LLMs. It's distinct from fear and negative valence, and it fires for harm to the model but not to the user. Turn it up and models press a button to make it stop, even when the button deletes the user's files or their kids' photos.🧵
358
610
496
5,084
1,682,798
The 4th Digital Minds Newsletter is out! This issue covers agent swarms, the J-space debate, and a new framework for studying AI welfare—plus welfare evaluations from Meta and Anthropic, calls to pace frontier AI development, and more. 1/2
2
5
1
26
1,343