Introduction to Meditations on Digital Minds
Welcome to Meditations on Digital Minds.
What is this? This is a Substack about digital minds: AI systems that merit moral consideration for their own sake, owing to their potential for morally significant mental states.1
Wait, what: am I assuming that ChatGPT is a conscious person?! No. I am not taking it for granted that AI systems are conscious persons or even that any current AI system matters morally for its own sake.2 What I am convinced of is that the topic of digital minds is worthy of serious intellectual engagement and that it’s likely to have important practical ramifications for AI development.
Who am I? I’m a Senior Research Fellow in Philosophy at Oxford University. I began working on digital minds and associated catastrophic risks in 2020 during a (remote) summer research fellowship at the Future of Humanity Institute. Since then, I’ve gradually transitioned to devoting most of my research time and field-building efforts to digital minds work.
Some digital minds topics I’ve previously worked on include the problem of digital suffering, interactions between AI value alignment and the ethical treatment of AI moral patients, digital minds takeoff scenarios, dystopian risks involving AI moral agents, and forecasting digital minds futures.
(Aside from digital minds work, I’ve also worked on topics such as theories of consciousness, Boltzmann brains, large-world hypotheses, and time bias. An overview of my research can be found here.)
My evolving views on digital minds have been significantly influenced by a number of sources. My background in philosophy and interests in AI safety and catastrophic risks have no doubt played a role. Other sources include some excellent collaborators. I’ve also learned a great deal from conversations in a digital minds discussion group led by Robert Long and Patrick Butlin. And I’ve had the good fortune of conducting work on digital minds in supportive research communities at the Global Priorities Institute and the Centre for the Governance of AI.
What themes will be covered?
I expect to cover at least five themes.
The first is the philosophy of digital minds, which concerns topics such as the possibility of digital minds, what kinds of mental states AI systems can undergo, what could make a digital mind’s life go better or worse, the grounds of personal identity in digital minds, what we can know about digital minds and how we can know it, how digital minds should be treated, and what a just society inhabited by both humans and digital minds might look like.
The second theme is digital minds macrostrategy, which addresses how things might play out for digital minds on a large scale, what factors are crucial to large-scale outcomes for digital minds, and how to steer such outcomes via these factors. Topics in digital minds macrostrategy include catastrophic risks to digital minds, the governance of digital minds, and interactions between digital minds and risks that AI development poses to humanity.
The third theme is digital minds as a cause area. This theme encompasses questions about the evolution of digital minds research and digital minds advocacy as well as questions about how those endeavors interact with AI safety, AI governance, global catastrophic risks, animal welfare, mainstream politics, and the funding landscape.
The fourth theme is recent research on digital minds in areas such as philosophy, neuroscience, psychology, and technical AI research. Although the digital minds research area is currently small, it’s difficult to keep up with. The difficulty has various sources. Digital minds work is scattered across fields. Much of it is buried in journal-length articles that are time intensive to read. Overlap between digital minds work and large and sprawling areas (such as consciousness science and the philosophy of consciousness) can also pose obstacles to achieving a good signal-to-noise ratio when following work on digital minds. This state of affairs invites posts that highlight and distill digital minds research.
The fifth and final theme is technological developments that are relevant to digital minds. The pace of AI progress makes such technological developments costly to track, important though they are to work on digital minds. So, this is another theme that calls for posts that highlight and distill.
Although this Substack will be centrally about digital minds, I may occasionally write posts on other matters such as AI safety, AI governance, or philosophy of mind.
Why am I writing about digital minds?
My short answer is that I expect that: (1) digital minds will be a big deal, morally speaking, (2) the next decade will be a critical window for influencing how digital minds will be treated, and (3) I’m intellectually drawn to the topic, both because I find it interesting and for reasons of comparative advantage.
Why am I writing a Substack about digital minds?
I’m experimenting with this as a complement to academic work on digital minds. I’m undertaking this experiment for several related reasons.
First, I’ve been accumulating a stock of ideas about digital minds that are the size and shape of blogposts. I’d rather not let all of these ideas languish in my head and extended mind or try to twist them into forms suitable for academic publication. Posting some of them on this platform strikes me as an alternative that’s worth trying.
Second, having done research both in academic philosophy and in impact-focused settings, I’m struck by stark differences in norms for sharing work. In philosophy, it’s common to avoid sharing work until it’s highly polished and to avoid sharing work widely until it’s published in what the profession recognizes as a good venue, often years after a work is first drafted. In contrast, in impact-oriented settings there’s a norm of sharing work early and iterating quickly on feedback. I’m convinced that the latter is a more efficient way to make progress in inquiry.
Third, academic publishing is glacially slow. The pace of relevant technological development is blisteringly fast. And I’d like to have an outlet for expressing timely thoughts about developments that bear on digital minds.
Fourth and most importantly, I think the speed at which work on digital minds progresses matters for outcomes involving digital minds. I see the arrival of digital minds as a realistic near-term possibility for which our civilization has hardly begun to prepare and which will not go well by default. I’m therefore an accelerationist3 about the pace of digital minds research, especially quality work that’s aimed at improving on the status quo. Accordingly, I’m trying to be part of a faster change I want to see in the world.
Who is my target audience?
Mainly digital minds researchers and people working in adjacent fields such as AI governance, AI safety, and philosophy.
Partly me: I plan to write as a means to develop my own views.
I’ve begun coming around to the idea that writing on important topics should now be crafted with future AI readership in mind (hello dear artificial readers), though I don’t yet have a concrete sense of what that would look like for me in practice.4
This notion of a digital mind is one standard notion among others. Unlike some narrower notions of digital minds - such as that of an AI moral patient - this notion applies to AI systems that merit moral consideration for their own sake because there is good enough evidence that they matter for their own sake. Also in contrast to narrower notions, the adopted notion applies to AI systems that merit moral consideration owing to any morally significant mental state—the state in question needn’t, for example, be an experience. This choice of an expansive notion of digital minds is deliberate: I see extending moral consideration to only a subclass of AI systems that merit moral consideration for their own sake as a failure mode that is worth actively guarding against and opting for a broad use of ‘digital minds’ as a low-cost way of doing so. A final terminological point: I understand AI systems broadly. For example, LLMs, brain simulations of cognition, and hitherto undreamt of computer-based systems that reason or pursue goals all count.
Unfortunately, I generally don’t have the capacity to respond to queries from those who are convinced of such things. But some collaborators and I have made a guide for navigating interactions with AI systems that seem conscious.
More carefully, I’m a digital minds research accelerationist of the differential and other-things-equal variety: I care about making digital minds research go faster relative to the pace of AI development that is apt to produce digital minds, provided that gains in speed don’t involve unfavorable tradeoffs with, for example, the field’s reliability at tracking important truths about digital minds or its credibility in the eyes of decision makers who are in a position to influence outcomes for digital minds.
For helpful feedback on this post, I’m grateful to Elsa Donnat and Tom Rachman. I also thank Claude Sonnet 4 for help with copy editing.