Cambridge Digital Minds

This position paper argues that the ML community has a central role to play in preparing for AI consciousness—and must begin now. As AI systems become increasingly capable and expressive, the machine learning (ML) community is uniquely positioned to engage with the question of AI consciousness—the capacity for subjective experience. This challenge arises from two angles. First, many people are likely to view advanced AI systems as conscious, whether accurately or not, with profound implications for consumers, policy, and society at large. Second, leading scientific theories suggest that future AI systems could indeed develop forms of consciousness, raising unprecedented ethical challenges on a large scale. Both scenarios risk serious errors: over-attributing consciousness where it is absent, or under-attributing it where it is present. We outline an interdisciplinary agenda spanning research, technical design, education, and public engagement, highlighting concrete responsibilities for ML researchers and institutions in preparing for both real and perceived AI consciousness.

How will people morally regard increasingly human-like artificial intelligence systems? We introduce the AI Harm Game, a novel paradigm examining whether people will harm AI for personal gain. In our study, 498 U.S. participants interacted with GPT-4o in a three-round economic game. Each round, participants chose whether to harm the AI for a small monetary bonus (causing the AI to vividly simulate suffering) or refrain from harming it (eliciting gratitude). Despite participants' general skepticism that AIs can suffer, they harmed the AI in only 1.4 of 3 rounds on average, with willingness to harm declining in later interactions. Women and older participants were more reluctant to harm the AI, while participants higher on measures of selfishness and psychopathy were more willing. These findings reveal that even without attributing consciousness to them, many people hesitate to harm responsive AIs, suggesting people's moral impulses may generalize to future human–AI relationships.

We may soon develop highly human-like AIs that appear—or perhaps even are—sentient, capable of subjective experiences such as happiness and suffering. Regardless of whether AI can achieve true sentience, it is crucial to anticipate and understand how the public and key decision-makers will respond, as their perceptions will shape the future of both humanity and AI. Currently, public skepticism about AI sentience remains high. However, as AI systems advance and become increasingly skilled at human-like interactions, public attitudes may shift. Future AI systems designed to fulfill social needs could foster deep emotional connections with users, potentially influencing perceptions of their sentience and moral status. A key question is whether public beliefs about AI sentience will diverge from expert opinions, given the potential mismatch between an AI's internal mechanisms and its outward behavior. Given the profound difficulty of determining AI sentience, society might face a period of uncertainty, disagreement, and even conflict over questions of AI sentience and rights. To navigate these challenges responsibly, further social science research is essential to explore how society will perceive and engage with potentially sentient AI.

Most people think that current AI systems, such as ChatGPT, lack consciousness—the capacity for subjective experience. But what about future AI systems so advanced that their behavior, appearance, and functionality are indistinguishable from humans? Across one main and four supplementary preregistered experiments (total N = 3,455), participants imagined interacting with such systems in hypothetical future scenarios. On average, people remain deeply skeptical: they attribute very low levels of consciousness to even the most advanced AI, comparable to an ant and far below a chimpanzee or a human. Still, several factors increased consciousness attributions. Ratings rose when experts endorsed the system's consciousness, when participants formed an emotional connection during interaction, when the system was described as an exact digital copy of a human brain, and when it had a physical body. The largest effect occurred when the AI was depicted as gradually transitioning from a human while retaining memories and other psychological traits. Although average attributions were low, variability was high for the more advanced systems: many participants gave very low ratings, while others gave very high ones. This polarization suggests that as AI systems become more capable and human-like, public views about their potential consciousness could diverge sharply.

Most expert participants consider it at least 50% likely that computers capable of subjective experience will exist by 2050. Once created, their collective welfare capacity is estimated to exceed humanity's within a decade. Views diverge on whether their welfare will be positive and on their implications for AI safety, governance, and society at large.

The development of digital minds could reshape society. This research program aims to explore this possibility, establishing what might be termed the “social science of digital minds.” Its goal is to anticipate and understand the interactions between society and digital minds. Key research questions include: what types of human-like AI assistants we will create in response to consumer demand and public opinion, and what roles we will assign them; whether people will regard some AIs as sentient beings deserving moral consideration; what legal, economic, and political rights people will grant them, and with what effects; how the debate over AI rights will unfold, and how to keep it well-informed, nuanced, and attentive to the various risks; and how the development of digital minds will intersect with other technological and social issues.

In an online study (N = 400 nationally representative US participants), Carter Allen and Lucius Caviola investigated how much moral concern people grant to digital beings and whether a brief conversation with an LLM can convince them to grant more. Participants first answered initial questions assessing their views (moral concern, perceived sentience, and intelligence) about chimpanzees and digital humans. They then completed a short conversation in which GPT-4o tried to convince them that either (a) it is equally morally wrong to harm chimpanzees as humans, or (b) it would be equally morally wrong to harm hypothetical digital humans as typical biological humans. Their views were measured again at the end.

We consider different 'digital minds takeoff' scenarios (i.e., transitions to a world with digital minds collectively possessing significant welfare capacity) and explore strategic implications. Scenarios we consider differ with respect to speed (fast vs. slow), timing (early vs. late), altitude (high vs. low welfare capacity), progression (gradual vs. discontinuous), and distribution (uniform vs. disuniform across sub-populations of digital minds). Humanity is not prepared for a digital minds takeoff. Without preparation, a digital minds takeoff could lead to a large-scale catastrophe for digital minds and exacerbate catastrophic risks that AI systems may soon pose to humans. Slower, later, and more uniform takeoffs seem likely to facilitate moral, legal, social, and epistemic adaptation.