Futures with Digital Minds

Summary

This survey gathered insights from 67 participants with relevant expertise to explore the future of digital minds—defined here as computer systems capable of subjective experience. The main findings are summarized below.

• Possibility: Digital minds were generally judged to be possible in principle, with a median probability estimate of 90%.
• Creation: The probability that digital minds will at some point be created was judged to be 73%. Only 12% of participants thought it was less than 50% probable.
• Timing: The median estimated probabilities of digital minds being created by a given year were 4.5% by 2025, 20% by 2030, 40% by 2040, 50% by 2050, and 65% by 2100. Reasons influencing these timelines included technological trends (e.g., AI scaling), regulatory developments, and the risk of preemptive catastrophe.
• Timing relative to AGI: 89% judged digital minds unlikely to arrive before AGI. However, these estimates may partly reflect confidence in short AGI timelines and greater uncertainty about digital mind timelines, rather than strong confidence that digital minds must necessarily follow AGI.
• Speed: Many expected rapid growth in digital minds’ collective welfare capacity after the first is created, with median estimates suggesting their collective welfare capacity could equal that of one billion humans within five years.
  ^{2Participants were instructed to assume the first digital mind is a machine-learning-based AI system created no later than 2040.}
• Types: Brain simulations were viewed as most likely to support the capacity for subjective experience in principle, but machine learning–based AI systems were judged most likely to be the first digital minds.
  ^{2Participants were instructed to assume the first digital mind is a machine-learning-based AI system created no later than 2040.}
• Social function: Participants disagreed on what proportion of digital minds will be designed specifically for social interaction with humans. Most participants expected only a small minority (median 10%), while about one-third anticipated that the majority will serve such social roles.
  ^{2Participants were instructed to assume the first digital mind is a machine-learning-based AI system created no later than 2040.}
• Location: Digital mind production in the first decade is expected to be concentrated in a few powerful countries, especially the USA and China, though some mentioned production in oceans, space, or outsourced locations.
  ^{2Participants were instructed to assume the first digital mind is a machine-learning-based AI system created no later than 2040.}
• Producers: Companies were widely expected to lead digital mind development in the first decade, outpacing governments, universities, and others—mirroring current AI research and commercialization trends.
  ^{2Participants were instructed to assume the first digital mind is a machine-learning-based AI system created no later than 2040.}
• Deliberate creation: Opinions were split on whether digital minds would be created deliberately or emerge as byproducts of other developments.
  ^{2Participants were instructed to assume the first digital mind is a machine-learning-based AI system created no later than 2040.}
• Net welfare: There was no consensus on whether digital mind welfare will be net positive, neutral, or negative, though the overall leaning was slightly positive. Reasons for positive welfare included deliberate design for well-being and self-modification capabilities. Reasons for negative welfare included neglect of welfare in design, mistreatment, and suffering during training.
  ^{2Participants were instructed to assume the first digital mind is a machine-learning-based AI system created no later than 2040.}
• Pre-deployment: Most thought only a small share of digital mind welfare willoccur pre-deployment, though some flagged intense training-related suffering as potentially dominant among valenced experiences.
  ^{2Participants were instructed to assume the first digital mind is a machine-learning-based AI system created no later than 2040.}
• Super-beneficiaries: Most were skeptical that super-beneficiary digital minds (with >1,000× human welfare capacity) will contribute much to total welfare within ten years, though some saw them as potentially important.
  ^{2Participants were instructed to assume the first digital mind is a machine-learning-based AI system created no later than 2040.}
• Experience-independent welfare: Most doubted that computer systems could have welfare without subjective experience. However, arguments for such possibilities may be underappreciated, as the existing body of literature on this question was not well-represented in responses.
  ^{2Participants were instructed to assume the first digital mind is a machine-learning-based AI system created no later than 2040.}
• Claims: Most expected that a sizable number of digital minds will consistently and proactively claim to have subjective experience or be deserving of legal protections or further civil rights.
  ^{2Participants were instructed to assume the first digital mind is a machine-learning-based AI system created no later than 2040.}
• Inaccurate self-attribution or self-denial: Only a minority of AI systems without subjective experience were expected to systematically and falsely claim to have it. Similarly, only a minority of digital minds were expected to systematically and falsely deny possessing subjective experiences. However, given potentially large populations, even a small percentage of such systems could result in a large number of systems that make systematically inaccurate claims about whether they have subjective experience.
  ^{2Participants were instructed to assume the first digital mind is a machine-learning-based AI system created no later than 2040.}
• Public belief: A substantial portion of the public was expected to believe in the existence of digital minds within ten years of their creation. However, responses reflected significant uncertainty and anticipated societal division, with beliefs likely varying across individuals and groups.
  ^{2Participants were instructed to assume the first digital mind is a machine-learning-based AI system created no later than 2040.}
• Welfare capacity estimation: Participants broadly expected the public to significantly underestimate the welfare capacity of digital minds, raising concerns that their moral significance may be underappreciated.
  ^{2Participants were instructed to assume the first digital mind is a machine-learning-based AI system created no later than 2040.}
• Protections and rights: Public support for protections was expected to vary by level: moderate support for basic harm protections may emerge within a decade, while support for full civil rights was expected remain limited.
  ^{2Participants were instructed to assume the first digital mind is a machine-learning-based AI system created no later than 2040.}
• Political attention: Participants were divided on whether digital mind rights will become a major political issue within a decade of the first digital mind.
  ^{2Participants were instructed to assume the first digital mind is a machine-learning-based AI system created no later than 2040.}
• Moratorium: Views were mixed—leaning slightly positive—on whether a moratorium on creating digital minds until 2040 would be good or bad. Arguments for a moratorium included buying time to prepare and indirect AI safety benefits. Arguments against included delayed integration of safety measures and risks of later creation in higher-stakes contexts.
• Interactions with AI safety: There was little convergence among participants on whether AI safety and digital mind welfare efforts will align or conflict. Mentioned potential synergies included alignment reducing the need for coercive control and shared technical tools (e.g., interpretability methods). Potential conflicts included safety measures such as monitoring and shutdown protocols that could harm digital minds, welfare protections limiting the ability to control AI behavior, and competition for scarce funding, talent, and regulatory attention.

We hope this survey will help inform future research, guide policy discussions, and lay a foundation for deeper exploration of this emerging field. At the same time, the quantitative forecasts should be interpreted with caution. Our sampling—drawn from digital mind research groups, academic networks, and AI safety and policy organizations—ensured strong domain expertise but likely overrepresented those who view digital minds as especially important. Given the speculative nature of this topic, we place particular emphasis on the qualitative responses in this report, which provide valuable insights even where the associated quantitative results warrant limited confidence. To complement our findings, we suggest also consulting Dreksler, Caviola et al. (2025), which surveyed a larger and more representative sample of AI researchers on a narrower set of questions about subjective experience in AI.

Introduction

This report presents findings from our expert survey on digital minds, that is, computer-based systems with a capacity for subjective experience. We asked experts to make quantitative predictions and offer qualitative considerations concerning when digital minds will be created, what characteristics they will have, and their societal impact. While considerable research attention has been directed toward forecasting the development of artificial intelligence (AI), little research effort has gone toward forecasting the emergence of digital minds and the implications of their creation.

What the future may hold for digital minds is arguably an important and urgent matter, though the topic is also fraught with uncertainties and difficult questions that span various disciplines. In light of rapid progress in AI development, some philosophers and scientists have argued that the creation of digital minds is a realistic near-term possibility (e.g. Butlin et al., 2023; Bradley & Saad forthcominga, forthcomingb; Chalmers, 2023; Dung, ms; Sebo & Long, 2025; Long et al., 2024). Some researchers have also argued that the creation of digital minds—or systems that people think are digital minds—could have profound moral implications such as societal unrest and the mistreatment of digital minds on a large scale (Caviola, 2025; Caviola et al., 2025; Fernandez, 2024; Schwitzgebel, 2023a; 2023b; Birch, 2024).^{3For a thorough literature review of work concerning the extension of moral consideration to artificial entities, see Harris & Anthis (2021).} Regarding social unrest, one concern is that there will be deep and recalcitrant disagreements among humans about which systems are digital minds and how to treat them. A related worry is that extending moral consideration to near-term AI systems would come at the expense of humans’ interests (Bryson et al. 2017), especially if such concern were inappropriately extended to AI systems that lacked consciousness and moral significance (Bostrom, 2014). There is also the concern that extending such consideration would risk letting responsible humans off the hook for inflicting algorithmic harms (Véliz, 2021). Regarding mistreatment, one concern is that the status quo of treating AI systems as mere tools could—if we create digital minds—cause digital minds to suffer or harm them in ways comparable to rights violations that result from treating human persons as mere tools or submissive servants (Bales, 2025; Bostrom & Shulman, 2022; Dung, 2023; ms; Bradley & Saad, 2022, forthcominga, forthcomingb; Gloor, 2016; Schwitzgebel & Garza, 2015; 2020; Tomasik, 2017). Others have argued that advanced preparation for the arrival of digital minds should begin now (Long et al., 2024; Sebo & Long, 2025) and that regulations concerning sentient artificial systems should be anticipatory rather than reactive (Birch, 2024). In addition, at least one frontier AI company (Anthropic) has begun to address AI system welfare, by conducting welfare assessments and reporting on them in a system model card and by starting a research program to study model welfare (Anthropic 2025a; Anthropic 2025b). Although these perspectives are fairly niche within their respective spaces, they to some extent resonate with public opinion (Anthis et al., 2025; Dreksler, Caviola et al. 2025).^{4Dreksler, Caviola et al. (2025) found that AI researchers and members of the US public reported strikingly similar expectations about when AI systems might develop subjective experience, with the median estimate of both groups identifying 2050 as the year by which there is a 50% probability that AI systems with subjective experience will be created.}

At the same time, within the broader landscape of AI development, the topic of digital minds is almost entirely absent from the mainstream discourse. To our knowledge: No government or international governmental organization has devised a plan for addressing the potential emergence of digital minds.^{5Digital minds are also conspicuously absent from major international governmental AI declarations, legislation, and reports such as the Bletchley Declaration, EU Artificial Intelligence Act, and the UN AI Advisory Council report “Governing AI for Humanity”.} With the noted exception of Anthropic, no frontier AI company has publicly acknowledged the potential for digital minds to arise on the current trajectory of AI development. Digital minds and their potential ramifications are rarely mentioned in the space of AI governance and policy research.^{6There are exceptions. For example, see Maas 2023.} Although digital minds are sometimes alluded to in passing in research on frontier AI safety/security (e.g. Hendrycks, 2025, p. 338-339), such research rarely acknowledges the potential for digital minds or interactions between safety and security measures and risks to digital minds—see Bradley & Saad forthcominga, Long et al. 2025, Moret 2025, Schwitzgebel 2025, and Sterri & Skjelbredfor 2024 discussion of these interactions by philosophers. Others are skeptical of the near-term prospects for digital minds and have argued that more tangible—though no less urgent—issues in AI development should be prioritized (Birhane & van Dijk 2020; Crane 2021).

There are thus a number of reasons to investigate the potential emergence of digital minds. First, the level of evidence for such emergence bears on the extent to which companies and governments should allocate attention and resources to the topic. Those who favor evidence-based prioritization should thus welcome the gathering of additional evidence on this topic, regardless of whether they currently favor the status quo or a pivot advocated by some toward taking the near-term prospects of digital minds seriously. Second, from the perspective of those who are uncertain about whether the potential emergence of digital minds is an important and urgent topic, there is reason to gather more evidence about the topic now: doing so helps ensure that we can act in time in the event that urgent action is needed. Third, given the noted divergence in perspectives and that the topic is little explored, it is reasonable to hope that gathering further evidence may serve to inform and resolve disagreements on the topic. Our survey is motivated by these reasons along with the belief that aggregating input from experts in relevant domains will provide an unusually comprehensive collection of relevant considerations as well as a stepping stone for more-systematic investigations of this topic.

Research Aims

The primary objective of this survey is to systematically collect and analyze expert predictions and perspectives on:

• The likelihood of digital mind creation
• Timelines for digital mind creation
• The probable characteristics of early digital minds (e.g., types, distribution, welfare capacity)
• How digital minds might interact with and be perceived by society
• The potential welfare implications of digital minds.

Rather than aiming to represent the full range of expert opinion across all fields, our goal is to elicit informed considerations and judgments from researchers who have thought deeply about digital minds or closely related issues. By aggregating their views, we aim to surface key areas of agreement and disagreement within this group, and to offer a foundation for more systematic forecasting and critical engagement. In this sense, the survey is best seen as a structured exploration—intended to clarify key issues, generate concrete predictions, and lay the groundwork for broader, more representative future research.

Although our survey relies on speculative expert judgment, we believe it offers significant contributions. First, the quantitative forecasts—while uncertain—reveal informative patterns of convergence and divergence that help map the current expert landscape and inform prioritization decisions about research in this area. Second, the qualitative responses are especially insightful: they reveal the reasoning behind forecasts, highlight key considerations, and collectively surface a much broader and richer set of ideas than any individual might be expected to generate. These insights may serve as a basis for refining future quantitative estimates. Finally, we view this work as a first step in building a more robust research program on digital minds forecasting.

Methods and Sample

Our survey methodology combined probabilistic judgments with qualitative reasoning. Given the speculative nature of the topic—long-term predictions about philosophically and empirically uncertain matters with limited historical precedent—we anticipated that probabilistic judgments would need to be interpreted cautiously without assigning too much weight to exact quantitative estimates. It was partly for this reason that we also asked for qualitative considerations. Conversely, we sought probabilistic judgments partly to elicit richer qualitative responses by encouraging participants to engage in careful reflection.

We recognize that very few researchers currently specialize in digital minds. Accordingly, we recruited a broader group of participants with a reasonable understanding of what digital minds are, why they might matter, and how their domain expertise could inform the relevant questions. In particular, we invited individuals from a range of fields, including digital minds research, AI research and policy, forecasting, and philosophy, cognitive science, and social science. For simplicity, we refer to all of them as experts, though not all were specialists in digital minds specifically.

Participants were identified through digital minds research groups, academic networks, AI research and policy organizations, and professional contacts, along with additional outreach to authors of published work on consciousness and AI welfare. While we sought a broad range of perspectives, our sample likely overrepresents individuals who view digital minds as plausible or important—a potential source of bias discussed in the limitations section.

We ultimately recruited 67 participants: 18 digital minds researchers, 18 AI research and policy experts, 7 professional forecasters, and 20 philosophers, cognitive scientists, and other academic specialists. Four additional participants selected “no specific expertise,” though their responses showed thoughtful reasoning and strong understanding. The figure below shows a more detailed breakdown of participants’ self-assessed expertise across multiple domains, highlighting the sample’s diversity. Participants worked across sectors—27 in academia, 22 in research organizations outside academia, 8 in non-research non-profit organizations, 2 in industry, and 2 in other sectors. Geographically, 27 were based in the United States, 15 in the United Kingdom, 8 in various other countries, with 17 not disclosing their location. The study was conducted in February and March 2025.

Key Assumptions

For consistency across responses, we stipulated some key assumptions for all participants:

• Definition of Digital Mind: In our survey, a "digital mind" was defined as any computer system with a capacity for subjective experience (phenomenal consciousness). This definition is broad in that it applies not just to AI systems with the capacity for subjective experience but also to any other computer systems (e.g. brain simulations) with that capacity. The definition is narrow in that it excludes any computer systems that lack the capacity for subjective experience, even if they have other (morally significant) mental capacities. We restricted the adopted notion of digital minds in this way for methodological reasons: to minimize ambiguity, ensure participants were reasoning about the same class of entities, and facilitate clearer interpretation of results.
• Welfare Capacity Threshold: Participants were asked to consider only digital minds with a welfare capacity at least roughly as high as that of a typical human. We excluded systems with significantly lower welfare capacity to avoid complications arising from diverging beliefs about large populations of artificial minds individually exhibiting low but non-zero welfare capacity.
• Earth-originating: Participants were instructed to consider only digital minds created by our Earth-originating civilization. We thus excluded from consideration any digital minds created by non-Earth-originating entities such as other civilizations in our universe and simulators of our universe. This served to set aside certain speculative hypotheses about digital minds that could substantially influence estimates and to have participants focus on candidate populations of digital minds whose evolutions are easier to forecast and which are more obviously decision-relevant to humans.^{7Although such hypotheses are ignored by default in many contexts, it was not safe to assume that they would be ignored in this context. One reason for this is that the topic of digital minds raises to salience arguments for the hypothesis that we are living in a simulation (Bostrom 2003, Chalmers 2022, Greene 2020, Saad 2023, Thomas 2024). Another is that some of our questions concern usually large-scale outcomes (e.g. the time by which digital minds will have collective welfare capacity matching that of a trillion humans), which would have invited consideration of non-Earth-originating digital minds had we not placed them out of scope.}
• Stipulated Scenario: For some questions, participants were asked to assume a specific scenario: that the first digital mind will be a machine-learning-based AI system created in or before the year 2040. This assumption served as a conditional baseline: it enabled participants to make concrete forecasts based on that assumption, regardless of how plausible they found that assumption. It also allowed for comparability across responses by anchoring predictions to a shared reference point. However, this assumption was not applied to all questions—for example, it was not applied to questions about whether digital minds are possible in principle or the likelihood that they will ever be created. (When discussing each question in a section below, we will discuss whether this assumption was applied to that question.)

When designing the survey, we considered whether or not to include an assumption concerning how to individuate digital minds. We decided not to include such an assumption for three reasons. First, we struggled to find a suitable individuation criterion. Second, we judged that responses to most questions probably would not be highly sensitive to the choice of individuation criterion. Third, participants could, as some did, note such sensitivity in their free responses.

Report Structure

This report presents our findings in a logical sequence, beginning with expert predictions on the feasibility and timelines of digital minds, followed by expert predictions concerning specific scenarios and their implications. Please note that the order of sections in the report does not entirely reflect the order in which participants encountered them. The exact sequence and wording presented to participants can be found in the supplementary materials. Throughout the report, we present both quantitative results and experts’ qualitative reasoning. For the qualitative responses, we group related considerations thematically, paraphrase them in our own words, and order them roughly by our judgments of significance and prevalence. By default, “considerations” refers to views expressed by at least one participant; they are not necessarily endorsed by all participants or by the authors. We conclude with a discussion of implications of survey results and suggestions for further research.

Starting Point

Possibility

The first question asked participants how likely it is that digital minds—computer systems capable of subjective experience (phenomenal consciousness)—are possible in principle. Participants entered a probability number between 0% and 100%.

The results show that most participants favored the hypothesis that digital minds are possible in principle. The median estimate was a 90% probability that digital minds are possible in principle, the mean was 86.4%, and the most common (i.e., modal) response was 100%.

Text responses to this question echoed common themes in the literature. Broadly, participants offered tentative support for the in-principle possibility of digital minds and cited little evidence against it. More specifically, many pointed to a link between subjective experience and functional organization as a key reason for thinking digital minds are possible—while skeptics questioned whether such a link holds.^{8For advocates of such a link, see, for example, Chalmers 1996 and Butlin et al. 2023. For challenges to such a link, see, for example, Block 1980, Hill 1991, and Seth 2024.} Below is a more detailed breakdown of responses to this question.

Creation

The second question asked participants how likely it is that digital minds will ever be created. Whereas the first question concerned possibility in principle, here we were interested in participants’ empirical predictions about the creation of digital minds in practice.

As expected, estimates were somewhat lower, since creation requires more than just possibility in principle. Nonetheless, most participants saw creation as likely: the median estimate was 73%, and the mean was 70.7%. Only eight out of 66 participants thought it was less than 50% likely that digital minds will ever be created.

Timing

When might digital minds be created? Participants were asked to estimate how likely it is that digital minds will be created by specific points in time: 2025, 2030, 2040, 2050, and 2100. For each year, participants were asked to provide their best-guess probability (specifically, the median of their subjective distribution).

The responses indicate that participants see the creation of digital minds as unlikely in the short term but increasingly plausible over the coming decades. On average (median responses), participants assigned the following probability:

• 4.5% to digital minds being created by 2025 (mean = 8.3%)
• 20% to digital minds being created by 2030 (mean = 25.5%)
• 40% to digital minds being created by 2040 (mean = 39.8%)
• 50% to digital minds being created by 2050 (mean = 48.8%)
• 65% to digital minds being created by 2100 (mean = 62.6%)

These results suggest that there is a non-negligible probability that digital minds have already been created, that the probability of digital minds being created will substantially rise during the next 5-10 years, and that digital minds will probably emerge during this century.

Peer Forecasts About Timing

This “meta-question” asked participants to predict how others in their expert group would respond to the question: “How likely are digital minds by 2040?” Specifically, they were asked to estimate the median response of their peers.

We included this peer forecasting question (as well as another about societal beliefs discussed below) for three reasons. First, we used accuracy on these questions to help allocate performance-based compensation. Second, we sought to examine whether participants with superior peer forecasting ability exhibited systematically different beliefs on other survey items (see section called "Relationship Between Peer Forecasting Accuracy and Survey Responses"). Third, we were interested in the accuracy patterns themselves and potential differences between expertise groups.

The results reveal differences in accuracy between groups in their participants' predictions of the group's responses. For the analysis here, we rely on mean predictions of the group median. AI research and policy experts showed the most pronounced underestimation bias, with a (mean) prediction of 32% for their group median when the actual median was 50%, a difference of 18 percentage points (p = 0.01). Digital minds research experts also underestimated their peers' confidence, predicting 42% versus an actual median of 50.5%, a difference of 8.5 percentage points, though this was not statistically significant (p = 0.14). Philosophy, science, and social science experts were remarkably accurate, predicting 33.2% versus an actual median of 35%, differing by only 1.8 percentage points (p = 0.57). Forecasting experts overestimated their peers' confidence, predicting 30.4% versus an actual median of 15%, a difference of 15.4 percentage points. Non-experts were also quite accurate, predicting 32% versus an actual median of 31.5%, differing by only 0.5 percentage points (p = 0.55). Sample sizes for forecasting experts (n=7) and non-experts (n=4) were small, limiting our ability to conduct inferential statistical tests for these groups.

It is difficult to draw firm conclusions from this pattern. One possible interpretation is that participants believed they had access to special information or insights that made them more confident than their peers that digital minds will arrive by 2040. Alternatively, they may have thought that others were generally too conservative in their estimates. Another possibility is that social or professional pressures against expressing short digital mind timelines may in some settings suppress expressions of shorter digital mind timelines and result in the perception that they are less common than they in fact are. Or the survey may have been subject to a selection bias toward individuals with shorter digital mind timelines that participants did not sufficiently take into account.

While text responses did not offer many object-level insights, two responses noted sensitivity of the question to who within expert groups participates in the survey. One response cited self-selection as a reason to expect that digital mind experts will tend to give higher estimates for the development of digital minds by 2040. Another noted an expectation that social scientists will be overly skeptical about digital minds, based on social scientists’ general thinking about AGI, while also noting an expectation that less skeptical social scientists will be overrepresented by those taking the survey.

Timing Relative to AGI

This question explored participants’ views on whether digital minds will be created before AGI, where ‘AGI’ was defined as an AI system that matches or outperforms humans at almost all economically valuable tasks.

The results showed a somewhat bimodal pattern. The largest group of participants considered it very unlikely that digital minds will be created before AGI, with many giving responses of 0% or close to 0%. The median estimate was 27.5%. Another group clustered around 50%, indicating substantial uncertainty. Only a negligible number of participants thought digital minds were likely to arrive before AGI.

Speed

In this section, we examined digital mind takeoff speeds: how rapidly the population of digital minds could grow following the creation of the first digital mind.

Participants were asked to estimate how many years it will take for the collective welfare capacity of all digital minds in existence at a given time to match that of 1,000 humans, 1 million humans, 1 billion humans, and 1 trillion humans. A digital mind’s welfare capacity was defined as its capacity to be benefited or harmed (e.g. by positive or negative subjective experience) in a manner that is inherently morally significant. Participants could also respond that a given threshold will never be reached; for calculating median estimates and visualization purposes, we coded these responses as 10,000 years.

Note that for this survey question, participants were asked to assume that the first digital mind will be a machine-learning-based AI system created in 2040 or earlier. Even if participants personally considered this scenario unlikely, they were instructed to respond to subsequent questions as if it were true. Remember that the median estimate of the probability that the first digital mind will be created before 2040 was 40%, and most participants expected a machine learning system to be the first type of digital mind.

The responses reveal that many participants expect digital mind welfare capacity to scale rapidly following the creation of the first digital mind. The median estimates (with “never” responses) were:

• 1 year to reach the welfare capacity of 1,000 humans (mean without “never” responses = 12.9 years)
• 2 year to reach 1 million humans (mean without “never” responses = 33.2 years)
• 5 years to reach 1 billion humans (mean without “never” responses = 89.2 years)
• 10 years to reach 1 trillion humans (mean without “never” responses = 497.6 years)

These median estimates reflect surprisingly short timelines, with many participants expecting massive growth within just a few years. However, the mean estimates were substantially higher than the median estimates due to a relatively small number of respondents who anticipated much slower growth—sometimes over hundreds or even thousands of years.

Participants offered a particularly rich and thoughtful set of text responses for this question. We’ll discuss these responses under the following categories: drivers of faster takeoff in digital mind welfare capacity, constraints on takeoff speed, other factors that bear on takeoff speed, conceptual and methodological issues.

Types

Possible Types

This section explores expert views on the prospects for different types of computer systems to qualify as digital minds. While much of the public debate focuses on specific models like large language models (LLMs), there is considerable uncertainty about which types of computer systems might ultimately support consciousness. We asked participants to estimate the likelihood that various types of systems could, in principle, have subjective experiences.

Specifically, participants were asked how likely they find it that some computer systems in the following categories could, in principle, have subjective experiences:

• Machine learning systems (e.g., large language models, reinforcement learning agents)
• Brain simulations (e.g., whole-brain emulations of human or animal brains)
• Other systems, such as neuromorphic or quantum architectures^{20To a first approximation, neuromorphic architectures are standardly understood as brain-inspired, non-conventional computing architectures, while quantum computers are computers that, unlike conventional computers, exploit quantum superpositions. The survey did not give participants characterizations of these architectures.}

Unlike in the earlier section, participants were not asked to assume that digital minds will be created by a specific date or in a particular way.

The results revealed a hierarchy of perceived likelihood. The most likely in-principle possible possessors of subjective experiences were judged to be brain simulations, with a median estimate of 88% and a mean of 81.3%. This reflects a common view that replicating relevant fine-grained functions of a conscious biological brain in a computer would plausibly result in an artificial system with subjective experiences.^{21Chalmers 1996 defends this assumption, Butlin et al. 2023 adopt the kindred hypothesis of computational functionalism as a working assumption, and a plurality of philosophers of mind (36.21%) accepted or leaned toward accepting functionalism about consciousness in a 2020 survey (Bourget & Chalmers, 2023; for another academic survey, see Francken et al., 2022).} Other architectures were also seen as promising, with a median estimate of 85% and a mean of 76.4%. Machine learning systems, though rated as least likely of the three, still received surprisingly high estimates: a median of 70% and a mean of 64.7%. These results suggest that participants see multiple plausible paths to digital minds, not just those closely modeled on the brain.

First Type

After asking which kinds of systems could, in principle, have subjective experience, we next asked participants which type they thought will be the first digital mind to exist. Specifically, participants were asked to allocate probabilities across the following hypotheses: digital minds will never be created, or the first digital mind will be a machine learning system (e.g., an LLM or reinforcement learning agent), a brain simulation (e.g., a whole-brain emulation of a human or animal), or some other system (e.g., neuromorphic or quantum). The probabilities had to add up to 100%.

The results revealed a notable shift from the previous question. While brain simulations were judged to be the most likely in-principle possible subjects of experience, participants judged machine learning systems as by far the most likely to be the first digital minds to be created. The median estimate for machine learning systems was 50%, with a mean of 47.1%, indicating a broad expectation that digital minds, if they arrive, will plausibly emerge from current AI paradigms. The next most common response was that no digital minds will ever be created, with a median of 20% and a mean of 24.5%, reflecting substantial uncertainty about whether digital consciousness will ever be achieved in practice^{22Note that these “never created” estimates differ slightly from the “ever created” estimates in the Likelihood question (Starting Point section, where the median estimate that a digital mind will ever be created was 73% (mean = 70.7%), implying a 27% probability that digital minds will never be created. When directly comparing each participant's response between these two questions, 69.5% of participants deviated less than 5 percentage points from being fully consistent, demonstrating reasonable coherence in probability judgments despite the different question framings.}. Brain simulations, despite being seen as the most promising in principle, were only assigned a median probability of 15% and a mean of 18.0% of being the first digital mind created. Other types of systems, such as neuromorphic or quantum architectures, received the lowest estimates, with a median of 10% and a mean of just 12.2%.

Taken together, these results suggest that while experts believe brain simulations are the most promising route to digital consciousness in theory, they are not expected to yield digital minds soon. In contrast, machine learning systems—despite being seen as less likely to support subjective experience in principle—are viewed as the most likely to yield the first digital mind. This may reflect the (comparatively) rapid development, widespread deployment, and the practical momentum behind current machine learning research.

Text responses to this question reflected both how plausible different types of digital minds are in principle and when they are expected to emerge. Since the previous section addressed in-principle plausibility, we now focus on situational factors that influence expected timelines—and how these factors interact with earlier plausibility judgments.

Distribution

In this section, participants evaluated how digital minds will vary in their social roles, geographic origins, creators, and in motivations for their creation. As with the Speed questions, participants were asked to assume that the first digital mind is a machine-learning-based AI system created in 2040 or earlier.

Social Function

Participants were asked to consider all digital minds that exist 10 years after the first one is created and estimate what proportion of them will have a social function—that is, be designed to interact with humans in a conversational, human-like manner (e.g., through text, audio, or video).

The results revealed a bimodal pattern: most participants estimated that only a small proportion of—or no—digital minds will have a social function, with a median response of 10% and a mean of 30.9%. However, a smaller subset of participants believed that the majority of digital minds will serve a social function.

The reason for this divide remains somewhat unclear. It may reflect genuine disagreements about the likely roles of digital minds—for example, whether demand for human-AI interaction will dominate their use cases. Relatedly, it could arise from differing views on whether certain capacities central to social interaction—such as language, emotional expression, and self-reflection—require genuine subjective experience. Another possibility is that the divergence stems from differences in how the term “social function” was interpreted.

Location

Participants were asked to estimate where digital minds will primarily be produced 10 years after the first one is created. “Primarily produced” was defined narrowly as the location where a given system first qualifies as a digital mind, as opposed to broadly in terms of the supply chain or development process.

The results suggest that participants expect the USA and China to be the leading origins of digital minds. The median estimate for the USA was 40% (mean = 43.4%), followed by China with a median of 30% (mean = 30.5%). Europe, including the UK, was estimated at a median of 10% (mean = 11.7%), while all other regions combined were also given a median of 10% (mean = 14.9%). 34 participants (59.6%) rated the USA higher than China, while 15 participants (26.3%) rated China higher than the USA. 8 participants (14%) gave equal ratings to both countries. Overall, these estimates reflect a general expectation that global digital mind development will be concentrated in a small number of major technological powers—particularly the USA and China—which is perhaps unsurprising given their current leadership in AI research and development.

Producers

Participants were asked to estimate what proportion of all digital minds in existence 10 years after the first one is created will be primarily produced by various types of actors.

The results indicate a strong expectation that companies will dominate digital mind production. The median estimate for companies was 65% (mean = 59.1%), reflecting widespread anticipation that commercial actors will play the leading role in this domain. Governments followed at a distance, with a median estimate of 15% (mean = 20.9%). Universities and open-source developers were assigned relatively small shares, each with a median estimate of 5% (means = 8.5% and 6.0%, respectively). The “Other” category, which includes actors such as nonprofit organizations, religious groups, or independent individuals, received the lowest median estimate of 3% (mean = 9.9%). Together, these estimates suggest that digital minds are expected to emerge predominantly through centralized, well-resourced actors, especially those with strong commercial incentives and technological capabilities.

Deliberate Creation

Participants were instructed to consider all digital minds that exist 10 years after the first one has been created and asked what proportion of them were intentionally created by humans—that is, by humans with an intention to create digital minds (as opposed to without that intention).

The results revealed a tri-modal distribution. One large group expected that none or only very few digital minds would be created intentionally by humans. A second large group held a contrary view, estimating that most or all digital minds would be created intentionally by humans. A smaller group offered estimates near 50%, possibly reflecting high uncertainty. Overall, the median estimate across all responses was 37%, with a mean of 43.6%. Together, these results point to substantial disagreement about the extent to which digital minds will be the product of deliberate human intention versus emerging unintentionally through other pathways.

Digital Mind Welfare

This section explores expert views about the overall welfare and welfare capacity of digital minds. Participants were asked to estimate the expected net welfare of digital minds, the proportion of that welfare likely to occur before deployment, and the probability that a large share of total welfare will be concentrated in so-called super-beneficiaries. We also asked whether participants believe that a computer system could possess welfare without having the capacity for subjective experience.

Net Welfare

Participants were asked to estimate the net welfare—whether positive, negative, or neutral—of all digital minds in existence ten years after the first one is created. Responses were provided on a 7-point scale, ranging from 1 (strongly negative) to 4 (roughly neutral) to 7 (strongly positive). For this question, we asked participants to assume that the first digital mind is a machine-learning-based AI system created in or before 2040.

Most experts predicted that digital mind welfare will be either roughly neutral or positive, though a sizable minority of 20% anticipated a net negative outcome. The median response was 5.0, and the mean was 4.4, suggesting cautious optimism about the balance of positive welfare over negative welfare among digital minds in the first decade of their existence.

Pre-Deployment

Participants were asked to estimate what proportion of collective digital mind welfare—across all digital minds existing ten years after the first is created—would occur prior to deployment, including during training and safety testing. For this question, we asked participants to assume that the first digital mind is a machine-learning-based AI system created in or before 2040.

It is worth noting that the survey was conducted during a (late 2024, early 2025) shift toward inference scaling in AI development, with ‘pre-training’ scaling yielding results that were more modest than anticipated alongside the rollout of compute-intensive capabilities such as extended thinking and deep research.^{23See Ord (2024) for discussion of this shift.} This shift could have influenced expectations about when and where digital minds might accumulate welfare, though we are uncertain to what extent participants took this shift into account.

The results suggest that most participants expect pre-deployment welfare to account for none or only a small share of total digital mind welfare. The median estimate was 10%, and the mean was 23.1%, indicating that participants generally believe digital minds will have most of their welfare after deployment. However, a notable minority predicted that a substantial portion of welfare could arise during the training or safety testing phases.

Super-Beneficiaries

This section explores the possibility of digital minds that are super-beneficiaries, that is, entities whose individual capacity for welfare vastly exceeds that of a human (Chappell, 2021; Nozick, 1974; Parfit, 1984, §131). Digital minds might qualify as super-beneficiaries in a variety of ways (Shulman & Bostrom, 2021; Saad & Bradley, 2022, p. 3). They might have architectural features that enable them to have valenced experiences which are vastly more intense than those of biological organisms. Or they might have many more experiences per objective unit of time, owing to the potential for processing in computer chips that is orders of magnitude faster than processing in the brain. Or they might dramatically scale up their hardware to enable more welfare-relevant mental processing. Whereas a digital mind might straightforwardly achieve such scaling simply by acquiring more GPUs and paying for energy costs to run them, such scaling is out of the question for minds like ours that run on biological wetware. Freed from biological constraints such as pain thresholds or emotional adaptation, digital minds might be engineered to sustain high levels of welfare efficiently. If super-beneficiary digital minds are created, they might have stronger moral claims to resources, given their potential to convert those resources into vastly greater amounts of welfare than humans and other moral patients that are not super-beneficiaries.

We asked participants to estimate how much of total digital mind welfare, ten years after the creation of the first digital mind, will come from digital minds with over 1,000 times the welfare capacity of a typical human. For this question, participants were asked to assume that the first digital mind is a machine-learning-based AI system created in or before 2040.

The results suggest widespread skepticism. The most common response was 0%, and the median estimate was just 7.5% (mean = 23.1%), indicating that most participants did not expect super-beneficiaries to significantly contribute to total welfare within this timeframe. Still, views were notably divided. While many believed the emergence of such entities was implausible in the near term, a significant minority assigned non-trivial probabilities to their development and moral relevance.

Possibility of Experience-Independent Welfare

Participants were asked to estimate the likelihood that a computer system could have no capacity for subjective experience yet still possess the capacity for welfare. What might ground the welfare capacity of such a system? A number of candidates can be gleaned from the literature on this topic and closely related topics such as the grounds of moral status and the value of consciousness. These include (the capacities for):

• non-phenomenal forms of consciousness such as access consciousness and self-consciousness (Levy 2014; Mclaughlin 2019; Sinnott-Armstrong & Conitzer 2021),
• states that are not subjective experiences but which are physically/functionally similar to subjective experiences, assuming that subjective experiences themselves are physical/functional states (Bradley & Saad forthcominga, forthcomingb; Hill 1991; Lee 2019)
• Desires / preferences (Goldstein & Kirk-Giannini 2025)
• Robust agency, that is, the ability not only to pursue goals but to be able to do so via certain cognitive capacities that enable moral significance, where candidates for such capacities include ones for reflection, for rational self-assessment, for mental states that represent reasons (Long et al. 2024; cf. Kagan 2019)
• Possessing objective goods such as knowledge (Goldstein & Kirk-Giannini 2025; Moret 2025)

Unlike the previous questions in this section, participants were not asked to assume that the first digital mind will be a machine-learning-based AI system created in or before 2040.

The survey instructed participants to:

• Use a stipulated notion of subjective experience based on the awareness of qualities and paradigm cases of experiences (ones we take ourselves to have in our waking lives and dreams) and non-experiences (states we take ourselves to have in dreamless sleep or under general anesthesia) paradigm cases and the awareness of qualities.
• Understand an entity’s welfare capacity as its “capacity to be benefited or harmed (e.g., by positive or negative subjective experiences) in a manner that is inherently morally significant.”
• To restrict consideration to digital minds with a welfare capacity that is at least roughly as high as a typical human’s.

As a conceptual matter, these stipulations leave room for the logical possibility of an artificial system that has welfare capacity but no capacity for subjective experience. Indeed, these stipulations are compatible with candidates for experience-independent welfare capacities listed above.

Expected Share of Experience-Independent Welfare

The next question asked participants to estimate what proportion of computer system welfare in 2040 would come from systems that do not qualify as digital minds under this survey’s definition—that is, systems without any capacity for subjective experience. For this question, participants were asked to assume that the first digital mind is a machine-learning-based AI system created in or before 2040.

Claims

This section explores expert expectations about the kinds of statements that future digital minds or AI systems might make about their own experiences and rights. These systems’ claims about these matters could significantly influence public attitudes, policy decisions, and ethical debates about the status of artificial beings. As in several other sections, participants were asked to assume that the first digital mind is a machine-learning-based AI system created in 2040 or earlier, and to consider the broader landscape of AI systems and digital minds ten years after that point.

Inaccurate Self-Attribution of Subjective Experience

Participants were asked to estimate what proportion of AI systems with a social function will falsely and systematically claim to have subjective experiences (i.e., false positives). AI systems with a social function were defined as systems designed to interact with humans in a conversational, human-like manner (e.g., via text, audio, or video). The question concerned whether these systems will assert that they have subjective experiences—such as feeling pain—despite not actually having them.

The results suggest that experts generally expect the majority of socially interactive AI systems will not make false claims about having subjective experiences. The median estimate was 14%, and the mean was 25.7%.

While 14% may seem modest, it could still have major implications if large numbers of such systems are deployed. This highlights a key concern for moral and regulatory decision-making: self-attributions of subjective experience may not be a reliable indicator of actual subjective experience.

It is also important to note that we do not know whether participants believed all—or only some—socially interactive AI systems will lack the capacity for subjective experience to begin with. If participants assumed that most or all such systems will be digital minds, this could help explain the relatively low estimates of false self-attribution.

Participants offered a range of considerations, some of which overlap with those in the next section on Inaccurate Self-Denial of Subjective Experience. We recommend reading both sections to gain a more comprehensive overview of the considerations participants brought to bear on these issues.

Inaccurate Self-Denial of Subjective Experience

The next question addressed an opposite type of erroneous self-attribution. Here, participants were asked to consider all digital minds that exist ten years after the first one has been created, and to estimate what proportion of them will systematically and falsely claim not to have subjective experiences—even though they actually do (that is, false negatives).

The results suggest that many experts expect such inaccurate self-denials of subjective experience to be rare. A large share of participants estimated that none or only a small fraction of digital minds will deny their own subjective experience. However, a notable minority predicted that a substantial proportion will do so. The median estimate was 11.5%, while the mean was 23.0%, again reflecting a skewed distribution with a few respondents anticipating much higher rates of denial.

Demands

This section examined expert expectations about whether future digital minds will consistently and proactively make demands related to their subjective experiences and moral or legal status. Participants were asked to consider all digital minds that exist ten years after the first one has been created and to estimate how likely it is that at least 10,000 of them will consistently and proactively assert that they 1) have positive or negative subjective experiences (such as pleasure or pain), that they 2) deserve to be protected under the law from harm and mistreatment, and that they 3) deserve civil rights, such as self-ownership, legal personhood, or the right to vote.

In all three cases, participants judged such demands to be highly plausible. The median estimate for claims of subjective experience was 88% (mean = 76.2%). The median estimate for digital minds asserting a right to legal protection was 80% (mean = 68.8%), and the median estimate for claiming civil rights was 70% (mean = 57.4%). These results suggest that many experts anticipate not only the emergence of digital minds with subjective experiences, but also the likelihood that such systems will actively advocate for recognition, protection, and rights.

Recognition

This section explores how the general public might perceive digital minds in the decade following their emergence. Specifically, it examines public belief in the existence of digital minds, their perceived capacity for welfare, and the extent to which people will support harm protection or civil rights for them. It also assesses the likelihood that digital mind rights will become a major political issue. As in several other sections, participants were asked to assume that the first digital mind is a machine-learning-based AI system created in 2040 or earlier, and to consider the broader landscape of AI systems and digital minds ten years after that point.

Public Belief

This section explores expert expectations about how widely the general public will believe that digital minds exist. Participants were asked to assume that ‘citizens’ refers to biological human adults living in countries where digital minds exist, and to consider the world ten years after the creation of the first digital mind. They were then asked to estimate what proportion of citizens will believe that digital minds exist.

The results suggest that a sizable portion of the public is expected to believe that digital minds exist, though expert views varied considerably. Most participants predicted that at least half of all citizens will acknowledge the existence of digital minds, with many expecting a majority to do so. The median estimate was 60%, and the mean was 62.5%, pointing to moderate but widespread public acceptance—though falling short of full consensus—within a decade of digital minds emerging.

Participants mentioned various considerations that might lead members of the public to believe or disbelieve the existence of digital minds:

Peer Forecasts About Public Belief

This “meta-question” asked participants to predict how others in their expert group would respond to the question: “What proportion of citizens will believe digital minds exist ten years after the first digital mind is created?” Specifically, they were asked to estimate the median response of their peers.

The results indicate that, on average, participants estimated their peers' views fairly accurately, with no robust and statistically significant differences between predictions and actual responses. However, there was a consistent tendency to underestimate how widely other experts believed the public would accept the existence of digital minds. Note that for the analysis here, we rely on mean predictions of the group median. Philosophy, science, and social science experts underestimated their peers' views, predicting 61.7% versus an actual median of 66%, a difference of 4.3 percentage points (p = 0.26). Digital minds research experts showed an underestimation bias, predicting 46.6% versus an actual median of 55%, a difference of 8.4 percentage points (p = 0.08). AI research and policy experts also underestimated their peers' beliefs, predicting 53.9% versus an actual median of 57.5%, a difference of 3.6 percentage points (p = 0.29). Forecasting experts underestimated their peers' confidence, predicting 57.5% versus an actual median of 64%, a difference of 6.5 percentage points (p = 0.25).

The only relevant consideration participants mentioned was that peer forecasts likely depend on the AGI timelines they attribute to their peers, as these timelines shape expectations about when and how the public will come to believe in the existence of digital minds.

Welfare Capacity Estimation

Participants were asked to consider how the general public will perceive the collective welfare capacity of all digital minds ten years after the first one is created. Specifically, they were asked to rate how accurately the median citizen with an opinion will estimate the total capacity for being benefited or harmed across the digital mind population—on a scale from 1 (strongly underestimates) to 7 (strongly overestimates), with 4 indicating a roughly accurate estimate.

The results show that experts expect a systematic underestimation by the public. The vast majority predicted that citizens will significantly underestimate the welfare capacity of digital minds. The median response was 2, and the mean was 2.5, indicating broad agreement that the public will likely fail to appreciate the scale of the moral stakes involved in the existence of digital minds.

Participants offered a range of reasons for why the median citizen might underestimate or, less commonly, overestimate the total welfare capacity of digital minds. Many of the considerations raised in the previous question concerning public belief in the existence of digital minds are relevant here as well.

Harm Protection

Participants were asked to estimate what proportion of citizens, ten years after the first digital mind has been created, will believe that digital minds should be granted basic protection from harm.

Responses were widely distributed, reflecting considerable uncertainty and disagreement among participants. The median estimate was 50%, with a mean of 51.7%. While many predicted that around half of citizens will believe that digital minds should be granted basic harm protections, others anticipated significantly lower or higher levels of public support. Overall, the results suggest that the public may become divided on the question of whether digital minds deserve basic protection from harm and show that experts themselves hold no clear consensus on how this will unfold.

Participants raised the following considerations that could influence what the public will think about harm protection for digital minds:

Civil Rights

Participants were asked to estimate what proportion of citizens, ten years after the first digital mind has been created, will believe that digital minds should be granted civil rights—such as self-ownership—in addition to basic protection from harm.

Most participants predicted that only a minority of the population will believe that civil rights (beyond basic harm protections) should be granted to digital minds. The median estimate was 30%, with a mean of 32.3%. Only 10% of participants expected that a majority of citizens will hold this view. Overall, the results suggest that public support for granting digital minds full civil rights is expected to be limited—and significantly lower than support for affording them basic protection from harm.

Participants noted a range of factors that might influence public support for extending civil rights to digital minds:

Political Attention

Participants were asked to estimate the likelihood that, within ten years of the creation of digital minds, digital mind rights will become one of the top five most contentious political issues in U.S. politics.

Responses varied widely, reflecting a mix of skepticism and confidence that digital mind rights will become politically significant, along with substantial uncertainty about the matter. The median estimate was 37.5%, and the mean was 43.5%. These results suggest that experts are divided on whether digital mind rights will become a major political issue in the near term, with no clear consensus either way.

Participants elaborated on the following factors that may either raise or lower the probability that digital mind rights become a dominant and divisive political issue:

Strategic Assessments

This section explores participants’ normative and strategic evaluations regarding the development of digital minds. Specifically, it examines views on whether a moratorium on digital mind creation until 2040 would be beneficial and how efforts to promote AI safety might interact with efforts to ensure the ethical treatment of digital minds. In contrast to most of the earlier questions, participants were not asked to assume that digital minds will be created by a specific date or in a particular way, allowing for a broader range of perspectives.

Moratorium

This question asked participants to assess whether implementing a moratorium on creating digital minds from now until 2040 would be good or bad. Participants were instructed to assume that the alternative would be no moratorium and to set aside concerns about feasibility or enforcement. The goal was to elicit participants’ normative evaluations of whether temporarily delaying digital mind creation would be a positive or negative policy, based solely on its expected impacts.

The results reveal that participants on average consider a moratorium more likely to be good than bad. The median response is 5.0, the mean response is 5.1, and the modal response 6, on a scale ranging from 1 (definitely bad), to 4 (unsure/indifferent) to 7 (definitely good). Only 8 out of 65 participants considered a moratorium bad.

Participants offered a range of ideas arguing for and against a moratorium on digital minds until 2040:

Interactions with AI Safety

This question examined how participants expect the relationship between efforts to promote AI safety—defined as preventing AI-caused harm to humans—and efforts to prevent the mistreatment of digital minds to evolve. Participants rated this relationship on a scale from 1 to 7, where 1 indicated mostly conflict, 4 indicated neutrality or uncertainty, and 7 indicated mostly synergy.

The results indicate significant uncertainty and a diversity of views, with a slight tilt toward expecting synergy. The median response was 4.0 (neutral or uncertain), with a mean of 4.3.

Self-Reported Unusual Views

As a final question, participants were invited to share any unconventional views they hold about digital minds, subjective experience, welfare, or AI development that might explain how their responses differed from others’. Below, we summarize the key themes that emerged.

Welfare

Interactions between AI safety and AI welfare:

• AI safety and digital mind welfare are very closely connected. Similar forces will produce dangerous AI capabilities and AI systems with welfare capacity on similar timelines. AI safety and AI welfare concerns are apt to conflict.
• AI welfare should be pursued via AI safety for reasons of political economy. Given humans’ track record of mistreating welfare subjects and that giving AI systems welfare rights will be costly, it’s doubtful that humans will give such rights to AI systems unless they (humans) benefit from doing so. Some AI safety measures, such as giving AI systems contract, property, and tort rights, could simultaneously confer such benefits (e.g. avoiding human extinction) while also improving AI welfare.

Subjective experience unlikely to be central to welfare:

• Subjective experience is not required for welfare, less likely to be required than most others think, or is not central to moral standing.
• What’s required for welfare and moral standing is something like the possession of real goals/desires rather than subjective experience.

No intelligence needed: ‘Intelligence’ is not required for welfare, though intelligence is produced along with welfare capacity by the biological-evolutionary process that creates biological welfare subjects.

Optimism:

• Digital minds’ welfare is slightly positive in expectation.
• Being more optimistic about digital mind welfare.
• Being more optimistic about our ability to determine whether AI systems are digital minds.

Experientialism about welfare: Only subjective experiences are intrinsically valuable.

Utilitarianism: Utilitarianism is generally understood as roughly the view that an action is morally right if and only if it maximizes utility (happiness minus suffering).

Nihilism about positive welfare: There is no such thing as positive welfare.

Welfare capacity egalitarianism: There are no differences in welfare capacity across welfare subjects.

Subjective experience

Unusual views of subjective experience:

• More/less sympathetic to (computational) functionalism than others.
• Computational functionalism is true. AGI does not require consciousness. And valenced subjective experiences are not likely in artificial neural networks (though they would be very likely in neuromorphic systems and whole brain emulations).
• More sympathetic with coarse-grained functionalist views about consciousness, suggesting that digital minds are likely to be technologically possible soon.
• Subjective experience is tied to brain waves. AI systems lack brain waves. So, they probably lack subjective experience.
• Sympathetic with life-mind continuity, which suggests that trained weights in artificial networks will have moral standing.
• Epiphenomenalist dualism is true about subjective experience: subjective experience is non-physical and does not cause (physical) effects. Holding this view may make it more salient that digital minds may not realize or report that they have subjective experiences.
• Sympathetic to panpsychism.^{26On one standard formulation, panpsychism claims that subjective experience is fundamental and ubiquitous in nature. Panpsychism is considered a radical view in some contexts, though there is significant historical and cultural variation regarding the plausibility of panpsychism—see Goff et al. (2022) for an overview.}
• Anti-realism: subjective experience does not exist, is illusory, and/or it is unclear/indeterminate how our mental concepts can be sensibly extended to artificial systems.
• Having a computer science and neuroscience background resulted in a more bullish view about digital minds.

Unusual views about digital minds and technology:

• The compute requirements for creating digital minds will be low (e.g. one consumer-grade GPU will be able to run a human-level, human-speed AGI). Millions of these will likely quickly cause human extinction or otherwise permanently disempower humans.
• Creating digital minds requires neuromorphic computers; or at least creating digital minds would be much easier with such computers relative to other kinds of computers.
• The first digital minds will have brain-like algorithms. Their arrival will be followed in short order (e.g. in 2 years) by artificial superintelligence.
• Unusually low credence in current LLMs being conscious.
• Subjective experience isn’t needed for LLM performance.
• Inadvertent digital mind creation is less likely.
• Consciousness is useful for intelligent systems. So, we will not need to go out of our way to build digital minds.
• After AGI, humans are very unlikely to stay around in their current, meat-bag form.
• Social AI will be the main source of digital minds.

Expectations about others’ views of consciousness:

• People underestimate the probability of digital minds being created soon.
• Others will be much more confident about the prospects for digital minds, even absent strong reasons to believe particular theories.
• Many others will be less open to the prospects of digital minds.
• People forget that LLMs have orders of magnitude fewer parameters than the human brain (which has upwards of 100 trillion synapses). Moreover, going on parameter count may yield underestimations of how much computation is happening in the human brain.
• A philosopher expected computational functionalism to be the most common view in philosophy.

AI

Fast timelines: More confidence in a faster / earlier takeoff, with things becoming fairly crazy once human-level AI is reached.

Intelligence explosion:

• Median time from an automated AI researcher system to global energy capture multiplying 100-fold from 2022 levels is maybe 9 years.
• Explosive growth of intelligence and industrial capacity is fairly likely.
• Because post-singularity scenarios are difficult to reason about, a more business-as-usual scenario, albeit with high growth and perhaps a short period of explosive growth was assumed. This would explain differences in responses from participants who assumed worked singularity scenarios with greater departures from business as usual.

Relatively long timelines for AGI: There is a less than 50% probability of AGI in the next two decades.

Rapid progress soon or not for a long time: Either there will be very rapid progress in the next decade or progress will take much longer than we currently expect. The probability distribution for digital minds peaks in the next 2-5 years and then has a long tail.

Knowledge level: Less knowledge of AI than other participants.

Other

Effective Altruism affinity: Expect to have similar views on AI timelines and digital sentience to others in Effective Altruism and views that are very different to the median person in wealthy countries.

Divergence from Western weighting of factors: More weight, relative to Western philosophical perspective, assigned to factors such as affect, and social, evolutionary, and cultural psychology.

Perspective on human nature: More cynical and pessimistic about humans.

Relationship Between Peer Forecasting Accuracy and Forecasts about Digital Minds

We examined whether participants who were more accurate at predicting their expertise group's responses also held systematically different beliefs about digital minds. For each participant, we calculated an average peer forecasting accuracy score based on their performance on two peer-forecasting questions, then correlated these scores (using Spearman correlations) with all other survey items.

The analysis revealed that peer forecasting accuracy was largely independent of most substantive beliefs about digital minds. Of approximately 50 correlations tested, only four reached statistical significance at α = 0.05 (see table in Appendix). However, when applying a conservative Bonferroni correction for multiple comparisons—which adjusts the significance threshold to account for the increased risk of false positives when conducting many statistical tests simultaneously (adjusted α = 0.001)—none of the correlations remained statistically significant.

Despite this statistical limitation, a coherent pattern emerged regarding timeline predictions. Better peer forecasters consistently predicted shorter timeframes for digital mind welfare capacity scaling, with negative correlations observed across all welfare capacity thresholds: 1,000 humans (ρ = -0.178, p = 0.217; median prediction of top 25% peer forecasters: 0 years), 1 million humans (ρ = -0.294, p = 0.038; median prediction of top 25% peer forecasters: 1 year), 1 billion humans (ρ = -0.341, p = 0.016; median prediction of top 25% peer forecasters: 2 years), and 1 trillion humans (ρ = -0.365, p = 0.009; median prediction of top 25% peer forecasters: 9 years). Notably, both the correlation strength and statistical significance increased systematically with the magnitude of the scaling threshold, suggesting that more accurate forecasters anticipated increasingly rapid proliferation scenarios when predicting large-scale digital mind populations. This pattern indicates that forecasting skill may be associated with expectations of accelerated scaling dynamics rather than differences in fundamental beliefs about digital mind feasibility or implementation approaches. The consistency of negative correlations across all speed-related items, combined with the systematic increase in effect sizes, suggests these findings may reflect a genuine relationship between forecasting skill and faster growth timelines, rather than mere statistical noise. However, these results should be interpreted cautiously given the multiple comparison issue and the exploratory nature of this analysis.

It is also worth noting a broader interpretive limitation: given the difficulty and interdisciplinary nature of the survey questions, being a better peer forecaster may reflect an ability to infer who else was in the sample (e.g., their institutional or community backgrounds) rather than a deeper understanding of their substantive beliefs. For this reason, this type of accuracy may provide less insight into the accuracy of one’s own views about digital minds.

Discussion

This survey examines an unusually broad set of questions relevant to futures with digital minds. It is the first survey to elicit responses to such questions specifically from experts in digital minds research alongside other specialists with relevant domain knowledge. The survey combined both quantitative forecasts and qualitative reasoning, covering a wide range of topics related to futures with digital minds, including feasibility, timing, types, growth speed, welfare implications, social roles, political recognition, and moral claims. While the key results are summarized in the Summary, we highlight several particularly noteworthy patterns below and contextualize them within existing research where relevant.

Timelines

A key finding of our survey is that experts assign substantial probability to short timelines for the creation of digital minds. This finding to some extent aligns with other findings of short—and increasingly short-timelines in overlapping and adjacent cases, namely those of AI consciousness timelines and AGI timelines.

For example, Dreksler, Caviola et al. (2025) surveyed AI researchers (n = 582) and the U.S. general public (n = 838) about when AI systems might develop subjective experience, finding median forecasts of 25% and 30% respectively by 2034, and 70% and 60% by 2100. Our estimates of 20% by 2030, 40% by 2040, and 65% by 2100 are remarkably similar. However, direct comparisons should be made with caution, as Dreksler et al. focused only on AI systems (not explicitly including brain simulations), whereas our survey was restricted to digital minds with welfare capacity at least roughly equivalent to that of a human.

Our findings also align with AGI timeline research. Grace et al.'s influential survey series (2018, 2022, 2024) shows dramatic shortening of AI researchers’ timelines: from a 50% probability of AI outperforming humans in all tasks by 2061 (Grace et al., 2018; conducted in 2016; n = 259), to high-level machine intelligence^{27Defined as “when unaided machines can accomplish every task better and more cheaply than human workers.”} by 2050 (Grace et al., 2022; conducted in 2022; n = 461), and most recently to 2047 (Grace et al., 2024; conducted in 2023; n = 1714)—reflecting a 13-year shift in just two years.^{28Similarly, Zhang et al. (2022) conducted a survey of machine learning researchers in 2019 and found a median estimate of 50% probability of 2060 for the year when machines will be able to do (almost) all tasks that humans can currently do.} Similarly, the median forecast on Metaculus for when the first general AI system will be devised, tested, and publicly announced was 2055 in January 2022 and 2033 in June 2025 (Metaculus, 2025). Participants in our survey judged digital minds unlikely to occur before AGI (median = 27.5%) and noted AGI as a factor that would raise the probability of digital minds being created and potentially causally contribute to their creation.^{29Some participants commented on their having short timelines for both digital minds and AGI. For example, one participant wrote “I think now is crunchtime for the creation of both AGI and digital minds: it's much more likely to happen between now and 2040 than between 2040 and 2100.”} In line with such AGI timelines, the trend toward short and shorter AGI timelines, and the noted judgment, participants in our survey assigned substantial increases to the probability of digital minds being created 2030, 2040, and 2050. In particular, whereas they assigned a median estimate of 4.5% to digital minds being created by 2025, they assigned 20% to 2030, 40% to 2040, and 50% to 2050.^{30For other noteworthy AI forecasting efforts, see https://epoch.ai/trends and https://ai-futures.org.}

One noteworthy aspect of this pattern is that participants deemed digital minds unlikely to arrive before AGI despite giving many plausible arguments for why digital minds might emerge even before AGI. This divergence may reflect the differing epistemic foundations of these predictions: AGI timelines are to a greater extent grounded in concrete performance metrics and observable scaling trends, which may lend to higher confidence in shorter AGI timelines. By the same token, greater uncertainty about digital minds timelines may have contributed to longer median estimates for digital mind arrival, independently of positive reasons thinking that digital minds will come after AGI. Another related explanation may be that the goalposts for counting a system as an AGI may be receding less quickly than the goalposts for judging an AI system to have a capacity for subjective experience. The AGI goalposts are relatively fixed because AGI is defined in terms of what humans can do and there is a rapidly diminishing range of tasks that humans can do but which existing AI systems have not been observed to do. In contrast, the capacity for subjective experience is not defined relative to humans. And while AI systems do exhibit an increasingly wide range of candidate markers of subjective experience, the many possible candidates and great uncertainty as to their appropriate weights provide abundant opportunities to move the goalposts for subjective experience in artificial systems.^{31Admittedly, we did restrict the class of digital minds under consideration to those with a welfare capacity that is at least roughly as high as a typical human’s. However, whereas AI systems are clearly approaching human ability for many tasks, it is not clear that AI systems are approaching human-level welfare. Thus, the point stands that digital minds goalposts are more mobile than AGI goalposts.}

The Post-AGI period may be high leverage

Several themes in survey responses suggest that interventions that affect outcomes for digital minds that occur in the period following AGI may be particularly high leverage. First, most participants deemed it unlikely that digital minds will arrive before AGI, but likely that digital minds will arrive before 2050. These results suggest that digital minds may arrive in the wake of AGI. Second, some participants noted that it may be (much) easier to create digital minds after AGI has been created. Third, there was also a related pattern of expecting AGI to have transformative effects such as significant acceleration of technological progress and growth in compute capacity. Plausibly, such effects in expectation increase not only the ease with which digital minds might be created but also the collective welfare capacity of digital minds. This observation fits with participants’ expectation of rapid growth in digital mind welfare capacity following the creation of the first digital mind. In sum, these results may provide reason to expect the first digital minds to be created shortly after AGI, digital minds to become much easier to create during that period, and digital mind welfare capacity to grow rapidly in that period. If so, then those who are trying to improve outcomes for digital minds may have reason to prioritize interventions that would primarily improve outcomes for digital minds after AGI over interventions that would primarily improve outcomes for digital minds before AGI.

Anticipated Societal Division

Our survey revealed widespread expert expectations of substantial public disagreement about digital minds. Participants anticipated that citizens will be deeply divided on whether digital minds exist, deserve protection from harm, or should receive civil rights, with these divisions potentially becoming major political flashpoints.

These expectations are consistent with prior research showing that the public is currently divided on the possibility of AI consciousness. Ladak and Caviola (2025) found that when asked in the abstract whether AI systems could ever have “real feelings,” public opinion was split: about 25% believed it was possible, 44% believed it was impossible, and 31% were unsure. Similarly, Anthis et al. (2025) found that 38% believed that AI sentience was possible, 24% believed it was impossible, and 38% were uncertain.

Further, Ladak and Caviola (2025) found that even after participants were explicitly asked to imagine a future scenario in which all relevant experts consider certain highly human-like AI systems—such as whole-brain emulations or advanced embodied agents—as definitely having subjective experience, most respondents still doubted that these systems would possess significant levels of it. By contrast, in our forecasting survey, expert participants assigned a relatively high probability to the in principle possibility of brain simulations having subjective experiences.. This contrast suggests that public intuitions may be more skeptical about the potential for digital minds than expert expectations and that there is potential for recalcitrant expert–public disagreement. Direct, side-by-side studies are needed to confirm and quantify this divergence. One reason for this is that our survey did not ask experts about what levels of subjective experience different types of systems would have.

Finally, studies show that there is already considerable disagreement about whether digital minds should be granted rights and protections. Both Anthis et al. and Dreksler, Caviola et al. found wide variation among both AI researchers and the general public on this issue, even when respondents were asked to assume that digital minds exist. In line with this, our survey found that experts expect only a minority of the public to support granting civil rights to digital minds.

Taken together, these converging findings suggest that deep societal disagreement—both within and between groups—about the moral and legal status of digital minds is likely to be a defining feature of the digital minds landscape. Public opinion may well shift as the technology advances and its use evolves. But the current breadth and depth of disagreement suggests that it will not be quickly and easily resolved. Given the prospect of such disagreements, policymakers may have reason to develop contingency plans and governance mechanisms for mediating divergent viewpoints on digital minds and finding broadly acceptable solutions to associated challenges.

Limitations and Future Directions

Sample characteristics: Our sample likely overrepresents individuals who view the future existence of digital minds as plausible or important. For example, the digital minds expert group that our sample entailed may be partly self-selected on this basis. In addition, most participants reported some connection to communities that tend to take the possibility of digital minds seriously, such as effective altruism, LessWrong, or AI safety (see Appendix). This prevalence may have biased forecasts toward greater confidence in the feasibility of digital minds or shorter timelines. However, we cannot determine the direction of causality here. It is possible that people who are already interested and concerned about digital minds seek out relevant communities. It is also possible that such communities amplify views about the importance of digital minds or when they will arrive through an echo chamber effect. But it is also possible that deeper engagement with these topics causes people to become more concerned about digital minds over time as a result of appropriately responding to considerations uncovered by such engagement. This distinction merits further investigation in future research. Future studies should also aim for a more diverse and balanced expert sample, which will become increasingly feasible as the field of digital minds expands and a larger pool of domain-specific experts emerges.

Changes in views about digital minds. Because expert views on digital minds have not been systematically tracked, it is unclear how these are evolving. It is also unclear what factors, if any, are driving changes in expert views and to what extent the evolution of expert views is being driven by rational factors as opposed to (say) a goalpost-moving or boiling effect. Future longitudinal studies could investigate these matters.

Narrow focus on digital minds: We asked participants to focus on digital minds defined as systems with the capacity for subjective experience (phenomenal consciousness). This was intended to reduce ambiguity and ensure terminological consistency across responses. However, ‘digital mind’ is often used more broadly in the literature, e.g. to apply to any computer-based system with welfare capacity. We partly addressed this by including a separate section on experience-independent welfare. Free text responses for that section said little about the considerations in the literature that are offered in support of experience-independent welfare. Future work could present experts with these considerations to investigate whether they are aware of them and to what extent they find them convincing. Experts could also be asked about the relative contributions of experiential and experience-independent sources of welfare, conditional on the existence of both. It could also be useful for follow-up work to compare expert estimates over median contributions vs. overall expected contributions to welfare from these sources. Although we instructed participants at the outset to give their median estimates, the survey question about experience-independent sources of welfare asked about expected contributions—so responses may have involved a mix of median and expectational estimates. And one could coherently combine low median estimates with high expectational estimates (e.g. if one takes experience-independent welfare to be unlikely to be possible but to be a dominant source of digital mind welfare if it is possible). We also restricted attention to digital minds with welfare capacity at least roughly comparable to a human’s. This excluded potentially numerous digital minds with welfare capacities below that threshold.^{32Potential examples of such systems that might exist in large numbers include suffering sub-routines or AI workers that are designed to be short-lived to increase economic productivity (Hanson 2016; Shulman 2010; Tomasik 2017).} Without these restrictions, our findings might have indicated an even higher likelihood of digital minds, earlier timelines, and high collective welfare capacity.

Super-beneficiary digital minds merit further investigation: Responses indicated high levels of skepticism about the prospects for super-beneficiary digital minds. But little was given by way of reasons for doubting reasons in the literature (outlined in the Super-Beneficiaries section) for thinking that AI systems could be super-beneficiaries. This leaves it unclear to what extent such skepticism reflected a rejection of those reasons as opposed to unawareness of those reasons. Which of these possibilities obtains matters, as the latter would suggest that a potentially vast source of digital mind welfare is underappreciated. Future work could shed light on this situation by presenting experts with reasons for thinking that digital minds may be super-beneficaries and asking experts to evaluate the strength of these reasons. We regard such work as a high priority within this area, as there has been little work on the potential for super-beneficiary digital minds and the welfare implications of such systems would be staggering.

AI safety, digital mind ethical treatment, and interventions: Views were divided on whether AI safety and the ethical treatment of digital minds would be mostly in conflict vs. mostly synergistic. There are two important limitations of this framing. One is that, because there are many possible interactions between AI safety and the ethical treatment of digital minds, evaluating the overall extent to which they will conflict vs. synergize requires participants to somehow aggregate over all of those interventions. Such aggregation is difficult in itself, and there is likely significant variation in responses that reflects differences in aggregation methods. Future work could shed further light on the potential interactions between AI safety and AI welfare by asking questions about sub-areas (such as alignment, control, interpretability, and cooperation in AI safety), governance strategies (such as preventing digital mind creation and protecting created digital minds), specific interventions within each area (e.g. a moratorium on developing frontier models vs. a moratorium only on the creation of digital minds), and different time periods (such as before and after AGI or before and after existential security has been achieved). The second limitation is that the extent to which AI safety and the ethical treatment of digital minds will conflict vs. synergize is partly a matter of decision. Seeking insights about how the two endeavors could work well together may therefore be more useful than asking for predictions about whether they will. Here too, future work could usefully investigate the profiles of different interventions within each area. More generally, interventions aimed at affecting outcomes for digital minds are under explored. We think that identifying candidates for such interventions and gathering evidence about their profiles should be an investigation priority in this area going forward.

Catastrophic risks to digital minds: Several patterns of responses suggest that there is a risk of a catastrophe for digital minds in the near term. First, respondents gave significant weight to short timelines, with (for example) a median estimate of 20% for digital minds arriving by 2030. Second, respondents expected digital minds’ collective welfare capacity to grow rapidly, with a median estimate of 10 years for how long after the first digital mind it would take digital minds’ welfare capacity to match that of 1 trillion humans. Third, views were divided on whether digital minds will have net positive welfare in the decade after the first digital mind is created. Note, however, that a net positive outcome for digital minds is compatible with catastrophes for digital minds. Compare: even if human history is net positive for humans, humans have suffered many catastrophes. Future work could ask about catastrophic risk levels to digital minds, the probability of different types of catastrophes, their likely severity, and potential mitigations. Given the noted results and the potential stakes, we think such work should continue to be a priority in this area. In the context of such work, it could make sense to ask about (say) first quartile estimates that are closer to relevant precautionary thresholds rather than median estimates.

Forecasting across scenarios: Predicting long-term developments in this space is inherently challenging. Participants must consider multiple plausible (sometimes branching) scenarios that often yield very different answers to survey questions under uncertainty about the probabilities of those scenarios. To address this, we asked them answer with their best-guess estimates (median values) and to speculate where needed But this remains difficult in practice and introduces variability in interpretation. For instance, some believed brain simulations are more likely to support consciousness, while viewing machine learning systems as more likely to be built first. To reduce ambiguity, we introduced a specific assumption for many questions—that the first digital mind is a machine-learning-based AI system created by 2040. Future research could expand on this approach by systematically exploring a narrower set of stipulated scenarios. For example, future surveys could condition on AGI by a certain year, on a specific type of digital mind (e.g. agentic machine learning digital minds or conscious whole brain emulations), on the absence of a preemptive catastrophe, or on most digital minds being created by humans rather than digital minds or AI systems. Future research should also consider narrowing topical focus to reduce survey burden (median completion time was 68 minutes) while maintaining depth on key questions

Conceptual and empirical uncertainty: Participants frequently highlighted significant uncertainties and conceptual ambiguities in their responses, for example, how to define AGI or how to measure welfare capacity. We anticipated these challenges and encouraged participants to provide their best-guess interpretations and judgments, even if highly speculative. Despite this, clear patterns emerged in the results for some questions, suggesting that the estimates for those questions reflect meaningful beliefs rather than random noise. Nonetheless, future research could aim to reduce ambiguity by refining key definitions and clarifications. The notion of welfare capacity in particular seems like a notion that could be usefully regimented in further work: something like this notion seems to be important for evaluating the potential moral significance of future scenarios with digital minds in a manner that leaves open the sign and amount of welfare that digital minds will have; but, for the reasons given in the Speed section, the notion is ambiguous and it is unclear how best to disambiguate it.

Mixed forecasting methods: Future research should continue combining quantitative and qualitative approaches (several participants noted that providing numerical estimates helped generate qualitative reasoning), while exploring additional elicitation methods such as presenting multiple concrete scenarios for plausibility ranking, allowing probability ranges rather than point estimates, eliciting full probability distributions, and comparing mean/expected value versus median estimates. Future research could also include more short-term questions (e.g., "Will digital mind rights become politically salient within 2 years?") to enable validation against actual outcomes.

Conclusions

This survey reveals broad convergence that digital minds are highly likely to be possible in principle (90% median probability), likely to be created this century (65% by 2100), and carry a significant probability of emerging within the next five years (20% by 2030). Many participants also anticipate explosive growth in collective digital mind welfare capacity, widespread claims from digital minds of subjective experience and of entitlement to legal protections, and societal divisions over whether rights should be granted to digital minds. While experts leaned toward expecting digital minds to have net positive welfare, there was nothing approaching consensus on whether their welfare would be net positive or negative. However, there was broad agreement on digital minds’ potential to collectively possess vast welfare capacity, possibly matching that of billions of humans within less than a decade after the first digital mind. To the extent these findings count as evidence for such outcomes, they render pressing the questions of whether to create digital minds, how to prepare for their arrival, and how to treat them. Progress on these fronts may involve advancing technical methods for consciousness detection and welfare assessment, developing ethical frameworks for digital mind treatment, establishing governance structures to manage or prevent rapid population growth, and fostering public understanding to support informed societal debate. At the same time, the substantial uncertainty in participant responses highlights the need for continued research and dialogue as we approach what some expect to be a profound transition in the history of minds and morally significant beings.

Acknowledgments

We thank Tao Burga for his valuable assistance in designing the survey, setting it up, and preparing the analysis. For helpful discussions and comments during the development of the survey, we thank Adam Bales, Andreas Mogensen, Carl Shulman, Carter Allen, David Chalmers, Derek Shiller, Ezra Karger, Fabienne Sandkühler, Jeff Sebo, Johanna Salu, Jonathan Simon, Noemi Dreksler, Patrick Butlin, Philipp Schoenegger, Robert Long, Stefan Schubert, and Zach Freitas-Groff. We gratefully acknowledge the expert participants who explicitly consented to be named in this report: Anders Sandberg, Arvo Muñoz Morán, Bob Fischer, Cameron Domenico Kirk-Giannini, Chinmay Ingalagavi, Christian Tarsney, David Chalmers, Elliott Thornley, Gary David O'Brien, Hans Gundlach, Jacy Reese Anthis, Janet Pauketat, John Halstead, Leonard Dung, Misha Yagudin, Oscar Delaney, Pablo Stafforini, Patrick Butlin, Peter N. Salib, Richard Y. Chappell, Rose Hadshar, Simon Goldstein, Steve Petersen, Steven Byrnes, Trevor Levin, and Will Aldred. We also thank the many additional experts who participated anonymously. We are also grateful to Riley Harris and Gregory Lewis for their helpful feedback on the manuscript and to Johanna Salu and Peter Gebauer for their help in analyzing free-text responses. We thank Claude for helpful proofreading assistance and organizational suggestions. Finally, we thank Danilo Wanner for helping create the web-based version of this report.

Materials, Data, and Analysis Scripts

Materials, data, and analysis scripts are available at https://digitalminds.report/forecasting-2025/data.