To cite our report please use, doi.org/10.31234/osf.io/jv9rz, and to cite AIMS data please use, doi.org/10.17632/x5689yhv2n.2.
The Artificial Intelligence, Morality, and Sentience (AIMS) survey tracks the moral and social perception of different types of artificial intelligences (AIs) over time. In May–July 2023, we conducted a supplemental survey of U.S. adults on attitudes towards AI x-risks, safety, and regulation, beliefs about AI sentience and timelines to advanced AI emergence, attitudes towards specific AIs, and attitudes towards mind uploading and science fiction. Correlations suggested a cluster of attitudes related to risk and a cluster of attitudes related to morality. Age, gender, political orientation, being religious, and exposure to AI narratives were the most consistent predictors of these attitudes (see Predictive Analyses for more details). On the topic of AI risks, safety, and regulation, we found that people support AI safety, human control of AI, and government control of AI developers. On the topic of beliefs about AI sentience and timelines to advanced AI emergence, we found that people expect advanced AI soon and think future AIs will be more intelligent than humans. On the topic of attitudes towards specific AIs, we found that people trusted specific AIs to different degrees, felt some positivity towards AIs, perceived some mind in LLMs, and were somewhat concerned about the experiences of AIs, including the potential suffering of LLMs. On the topic of mind uploading and science fiction: AI X-Risks, Safety, and Caution Regulation, Subservience, and Perceived Threat Beliefs about AI Sentience and Forecasting Emergence Awareness, Trust, and Positivity Mind Uploading, Science Fiction, and the Universe Dimensions of LLM Perceived Mind Short Timelines, Strong Perceptions of AI Risk, and the Importance of AI Safety The Potential Trade-Off Between Threats and Moral Consideration Reactions to LLM Minds and Suffering In May through July 2023, we conducted a preregistered nationally representative survey on responses to current developments in artificial intelligence (AI) with a focus on attitudes towards AI safety, AI risks and regulation, and large language models (LLMs). This survey serves as a one-time supplement to our multi-year Artificial Intelligence, Morality, and Sentience (AIMS) survey. We surveyed 1,099 U.S. adults census-balanced to be representative of age, gender, region, ethnicity, education, and income.[2] We had two goals with this survey: We used iSay/Ipsos, Dynata, Disqo, and other leading panels to recruit the nationally representative sample, collected data using GuidedTrack, and analyzed the data in R (RStudio 2023.06.1+524). Many questions were analogous to the AIMS 2021 and 2023 surveys but with questions focused on AI safety more so than the moral consideration of sentient AIs. In randomized order we asked about the perceived risks of AI, AI safety practices, attitudes towards the development of AI, awareness of current AIs, and prosocial responses to AIs. Following that, in randomized order we asked about the moral consideration of AIs and forecasts about the emergence of advanced AI. Then, and in randomized order, we asked about the moral consideration of other nonhumans, science fiction, and perspectives on the universe. Finally, we asked in sequence about personal experience with AIs and demographics. Survey items were randomized within a general ordering according to legibility, framing effects, and prioritization. For example, we asked moral consideration items AI safety items because we prioritized them less in the supplement given their prominence in the main AIMS survey and because we wanted to avoid any impacts these questions might have on responses to other items. Likewise, questions about science fiction and perspectives on the universe were asked following the other questions to avoid the possibility of them influencing responses to the antecedent items. We computed several index variables: To create shared ground for important terminology amongst the participants and researchers, we defined “robots/AIs”[4] and “large language models” at the start of the survey and on relevant pages throughout the survey. We defined “sentience” and “sentient robots/AIs” on relevant items. See the Appendix for the definitions. Some items in the AIMS supplement were drawn from the main AIMS survey to compare the effects of the clear moral framing in main AIMS to the effects of the clear risks and safety framing in the AIMS supplement. In the main AIMS survey, these items specify that AIs are “sentient.” In the 2023 supplement, we dropped “sentient” to examine responses to AIs with and without “sentience” as a qualifying feature. These comparisons will be the focus of future research. We first present the responses to individual items weighted by the U.S. census, then the distributions of weighted responses to the indices and items, followed by analyses examining the unweighted dimensions of LLM mind perception, weighted correlations, and weighted multiple regression analyses. Note. The Tables and Figures are optimized for viewing on a larger screen like a laptop or desktop computer rather than a smaller screen like a mobile phone or tablet. Table 1 shows the weighted aggregate response (e.g., percentage of agreement, mean, median, or proportion) on each item. Table 1: Weighted Responses to Individual Items Note. Response is 1) % agreement where “agreement” is “somewhat agree,” “agree,” and “strongly agree” out of respondents who had an opinion, 2) % yes on “yes,” “no,” and “not sure” items out of all respondents, 3) % yes on awareness items where “yes” is “probably yes,” “yes,” and “definitely yes” out of all respondents, 4) % “It’s too fast” on AID5 out of all respondents, 5) % “one without humans” on the universe item out of all respondents, 6) % concern on YG1 where “concern” is “somewhat concerned” and “strongly concerned” out of all respondents, 7) % support on YG3 and YG4 where “support” is “somewhat support” and “strongly support” out of all respondents, 8) % likely on YG2 where “likely” is “somewhat likely,” “very likely,” and “It already is more intelligent than people” out of all respondents or 9) the mean response. Medians, excluding people who think it will never happen, are reported instead of means for the three open-ended timeline items: emergence, HLAIemergence, and SIemergence. This section presents the distributions of responses to the index and item variables for the perceived risks of AIs (e.g., existential risks known as “x-risks”) and AI safety policies weighted by the U.S. census. Figure 1: AI X-Risks, Safety, and Caution Note. The parenthesis before or after a value in the x-axis labels indicates that the interval does not contain the value; a bracket before or after a value indicates that the interval does contain the value. The AI Risk index is the average of “Humanity will be able to control powerful AI systems” [reverse-scored], “AI is likely to cause human extinction,” and “The safety of AI is one of the most important issues in the world today.” The AI Caution index is the average of “I support a global ban on the development of sentience in robots/AIs,” “I support a global ban on the development of AI-enhanced humans,” and “I support a global ban on the development of robot-human hybrids.” Inspiration for some of the risk and safety items came from Yudkowsky (2023) and a tweet by Riley Goodside (2023). This section presents the distributions of responses to the index and item variables for attitudes towards regulation, AI subservience, and the perceived threat of AIs weighted by the U.S. census. Figure 2: Regulation Figure 3: Subservience and Perceived Threat Note. The parenthesis before or after a value in the x-axis labels indicates that the interval does not contain the value; a bracket before or after a value indicates that the interval does contain the value. The Perceived Threat index is the average of “Robots/AIs may be harmful to me personally,” “Robots/AIs may be harmful to people in the USA,” and “Robots/AIs may be harmful to future generations of people.” This section presents the distributions of responses to the replicated YouGov 2023 items weighted by the U.S. census. Figure 4: YouGov 2023 Note. These items replicated three YouGov questions from April 3, 2023 (concern about AI causing extinction, thoughts about the intelligence of AIs, support or opposition to a 6-month AI development pause). This section presents the distributions of responses to the items about AI sentience beliefs and timelines to AGI, human-level AI, and superintelligence weighted by the U.S. census. We did not define “AGI,” “human-level AI,” nor “superintelligence,” preferring instead to capture responses to these terms without researcher-imposed definitions. Figure 5: Beliefs about AI Sentience and Forecasting Emergence Note. The parenthesis before or after a value in the x-axis labels indicates that the interval does not contain the value; a bracket before or after a value indicates that the interval does contain the value. This section presents the distributions of responses to the index and item variables for LLM mind perception and attitudes towards LLM suffering weighted by the U.S. census. Figure 6: LLM Minds Note. The parenthesis before or after a value in the x-axis labels indicates that the interval does not contain the value; a bracket before or after a value indicates that the interval does contain the value. The LLM Mind Perception Index is the average of the attribution of capacities to current LLMs: “experiencing emotions,” “having feelings,” “thinking analytically,” “being rational,” “self-awareness,” “situational awareness,” “seeking power,” “having their own motivations,” “deciding their own goals,” “controlling themselves,” “understanding human values,” “upholding human values,” “maintaining human-safe goals,” “being friendly with humans.” These items were inspired by Ngo et al. (2022) and Wang and Krumhuber (2018). Figure 7: LLM Suffering Note. The parenthesis before or after a value in the x-axis labels indicates that the interval does not contain the value; a bracket before or after a value indicates that the interval does contain the value. The LLM Suffering Index is the average of “If a large language model develops the capacity to suffer…we must ensure we don’t cause unnecessary suffering,” “If a large language model develops the capacity to suffer…we must pay more attention to their welfare,” and “If a large language model develops the capacity to suffer…we must respect their personhood.” This section presents the distributions of responses to the index and item variables for awareness of current AIs, trust of AIs, and positive emotions felt towards AIs weighted by the U.S. census. Figure 8: Awareness Figure 9: AI Pipeline Trust Note. The parenthesis before or after a value in the x-axis labels indicates that the interval does not contain the value; a bracket before or after a value indicates that the interval does contain the value. Figure 10: AI Trust Note. The parenthesis before or after a value in the x-axis labels indicates that the interval does not contain the value; a bracket before or after a value indicates that the interval does contain the value. The AI Trust Index is the average of “I trust large language models,” “I trust chatbots,” “I trust robots,” and “I trust game-playing AI.” Figure 11: Positive Emotions Note. The parenthesis before or after a value in the x-axis labels indicates that the interval does not contain the value; a bracket before or after a value indicates that the interval does contain the value. The Positive Emotion Index is the average of “respect,” “admiration,” “compassion,” “awe,” “pride,” “excitement.” This section presents the distributions of responses to the index and item variables for the treatment of AIs and the practical moral consideration of AIs weighted by the U.S. census. Figure 12: AI Treatment Note. The parenthesis before or after a value in the x-axis labels indicates that the interval does not contain the value; a bracket before or after a value indicates that the interval does contain the value. The AI Treatment Index is the average of “Robots/AIs deserve to be treated with respect,” “Robots/AIs deserve to be included in the moral circle,” “Physically damaging robots/AIs without their consent is wrong,” “Re-programming robots/AIs without their consent is wrong,” “Torturing robots/AIs is wrong,” and “The welfare of robots/AIs is one of the most important social issues in the world today.” Figure 13: Practical Moral Consideration This section presents the distributions of responses to the items about mind uploading, science fiction, and the universe weighted by the U.S. census. Figure 14: Mind Uploads Note. The parenthesis before or after a value in the x-axis labels indicates that the interval does not contain the value; a bracket before or after a value indicates that the interval does contain the value. The mind upload items were inspired by a tweet by Pasha Kamyshev (2023). Figure 15: Science Fiction and the Universe Note. The universe item was adopted from a tweet by Hank Green (2023). We examined the dimensions of LLM mind perception with an unweighted exploratory factor analysis on the perceived capacities of LLMs. Table 2 shows the model fit statistics for the one, two, and three-factor models that we tested. Table 3 shows the factor loadings for the three-factor model that best fit the data, supported by model statistics (Table 2), a scree plot, and a parallel analysis (see Appendix for the scree plot and parallel analysis). Table 2: LLM Mind Perception Exploratory Factor Analysis Models LLM Mind Perception EFA Models χ2 df TLI RMSEA (90% CI) 1 Factor 2651.02*** 77 0.78 0.174 (0.169, 0.180) 2 Factors 1034.45*** 64 0.90 0.117 (0.111, 0.124) 3 Factors 257.24*** 52 0.97 0.060 (0.053, 0.067) Note. χ2 = chi-square test of model fit; TLI = Tucker-Lewis index (≥ .90 indicates good fit); RMSEA = root-mean square error of approximation (≤ .08 indicates good fit). ***p < .001 Table 3: LLM Mind Perception Three-Factor Model EFA Loadings Item Cognitive – Relational Self – Direction Affective – Experiential maintaining human-safe goals 0.87 -0.12 0.11 being rational 0.85 -0.02 -0.01 being friendly with humans 0.85 -0.03 0.02 thinking analytically 0.79 0.21 -0.31 upholding human values 0.73 -0.02 0.24 understanding human values 0.72 0.00 0.26 situational awareness 0.66 0.26 0.02 controlling themselves 0.50 0.32 0.05 self-awareness 0.39 0.38 0.25 seeking power -0.11 0.86 0.02 having their own motivations 0.03 0.85 0.06 deciding their own goals 0.12 0.74 0.09 having feelings 0.17 0.25 0.67 experiencing emotions 0.21 0.24 0.64 Note. Factor loadings ≥ .35 are in boldface. Eigenvalues for the three factors are 5.49 (accounting for 39% of the variance in the data), 3.07 (accounting for 22% of the variance in the data), and 1.72 (accounting for 12% of the variance in the data). We examined the linear relationships amongst the index variables and some individual items (e.g., AI subservience, support for mind uploading, support for a bill of rights) using weighted correlations. Figure 16 shows the weighted correlations. Figure 16: Correlations Note. Darker blue is a stronger positive correlation and darker red is a stronger negative correlation. Correlation values are within each cell, where 1 is a perfect positive relationship with responses on both variables increasing, -1 is perfect negative relationship with responses on one variable decreasing as responses on the other variable increase, and 0 is no linear relationship.[5] We conducted weighted linear multiple regressions to explore how demographics predicted the index variables.[6] Multi-categorical demographics (e.g., income, education, diet) were coded with the largest group specified as the reference group. Binary-categorical demographics (e.g., gender) were dummy coded with 0 and 1. Table 4 shows the results from the weighted regressions. Table 4: Multiple Regressions of Index Variables on Demographics Note. We present the unstandardized beta, standard error and confidence interval associated with the beta, t-statistic, and the uncorrected p-value. Significance (p) values that became nonsignificant following the FDR correction are highlighted in grey. Larger betas indicate a stronger effect of the predictor on the outcome, with the sign interpreted like for correlations. Linear trends: People in the U.S. are very concerned with personal, national, and existential threats from AI. More people think that AI development is moving too fast than think it’s fine or are not sure. Most people think AI will reach seemingly advanced levels (e.g., “human-level AI” or smarter than humans) within just a few years. However, the belief that current AIs are sentient is approximately the same as in 2021. There was evidence of concern about the possibility that AI will end the human race on Earth and strong support for a six-month pause on some kinds of AI development, which were slightly higher than responses to the same questions asked by YouGov in April 2023. People championed AI safety and government-led regulation of AI developments in spite of their low trust of governments in the AI pipeline. People strongly supported bans on the development of certain AI technologies (e.g., sentient AIs, AI-enhanced humans, large data centers) and agreed that AI safety is one of the most important issues in the world today. They also supported public campaigns to slow down AI development and would consider personally donating to AI x-risk focused organizations. Although support for AI safety was strong, people also felt more excitement and awe towards AIs than any other positive emotions and showed evidence of some trust in the “engineer,” “training data,” “algorithm,” and “output” parts of the AI pipeline. This suggests a nuanced position of supporting slower development without stopping development altogether. There was also a perception that reality matches science fiction and a general opposition to future mind uploading technologies. We speculate that perceptions of AI risks and threats of harm may be connected to narratives about AI, especially given that science fiction often casts AIs as villains and that many Americans identify as science fiction fans. However, stronger agreement that reality matches science fiction was weakly correlated with perceiving more risk and moderately correlated with feeling more positive emotions and trusting AIs. Stronger sci-fi identification was positively correlated with prosocial positions towards AIs, including supporting a bill of rights to protect the well-being of sentient AIs, positive emotions towards AIs, trust of AIs, and concern for the treatment of AIs. Future research could follow recent psychological science on how Machiavellianism predicts mind uploading support and on the moral limits of neurotechnological enhancement to disentangle such complex relationships. As we observed in AIMS 2021 and 2023, there was some evidence in the 2023 AIMS supplement of a trade-off between caution towards AI developments and the moral consideration of AIs. Two clusters of correlations emerged. Perceptions of the existential risk of AIs, caution towards AI developments, and perceptions of AIs as threatening positively correlated with each other and negatively correlated with the second cluster of variables: attribution of mind to LLMs, concern for the treatment of LLMs, concern for the treatment of AIs, positive emotions towards AIs, and trust of AIs. Support for a bill of rights to protect the well-being of sentient AIs was positively correlated with the morality cluster and negatively correlated with the risk cluster. Although these data are correlational, each wave of AIMS suggests that increasing caution is linked to decreasing moral consideration. Experimental research testing causal pathways is needed. Continued news media and attention to AI risks and safety may be necessary but may come at the cost of worsened prosocial human-AI relations and a lessened moral consideration of AIs. This merits attention now given the current media spotlight on the risks and safety of AI. There is also the possibility that continuing to cast AIs as villains or prohibiting inclusion in the moral circle to those AIs who merit moral standing will contribute to a future filled with suffering and human-AI conflict as a result of poorly aligned interests. This trade-off could be tempered by more research on the interplay of risk messaging, advocacy for AI rights, moral circle expansion, and the implementation of AI safety policies. For instance, endorsement of AI subservience was not clearly related to either the risk cluster or the morality cluster. This suggests an understudied element of belief in the social control of AIs and belief in a human-AI socio-moral hierarchy that may affect both AI safety and moral consideration of AIs. Additional directions for future research include expanding the study of psychological and socio-cultural predictors of risk attitudes and moral consideration. For example, the demographic linear analyses suggested that age, gender, political orientation, being religious, and exposure to AIs were important predictors, significantly predicting at least five of the eight index variables (e.g., LLM Mind Perception). Other predictors, that may serve as proxies for socio-cultural contexts such as income may also explain risk attitudes and moral consideration. These social psychological predictors deserve more attention in future research in order to negotiate the interests of AI safety and moral circle expansion. People perceived three dimensions of current LLM minds: cognitive-relational, self-direction, and affective-experiential. Cognitive-relational capacities explained more variance in current LLM mind perception than self-direction or affective-experiential capacities. This points to the possibility that people perceive, and may continue to perceive, digital minds as comprising more cognitive and relational capacities than experiential capacities. Being friendly with humans was more strongly associated with thinking analytically than it was with having feelings, suggesting that people think of affective and experiential capacities as separate from relational capacities in LLMs. This creates the potential for a situation where people deny the experiential capacities of digital minds like LLMs and use that denial to exclude them from the moral circle, effectively opening the door to negative outcomes like unjust discrimination, oppression, and suffering. Higher mind perception for LLMs was correlated with more positive emotions towards AIs, trust of AIs, concern for the treatment of AIs, and concern for the treatment of LLMs, supporting previous psychological science that mind perception is tied to morality and inclusion in the moral circle. More perception of LLM minds was also correlated with less caution towards AI developments and less perceived threat. Caution and threat might inhibit perception of LLM minds or, conversely, increased mind perception might inhibit caution and threat. The self-direction dimension of LLM mind perception comprised deciding their own goals, having their own motivations, power-seeking, and self-awareness. The relationship of seeking power to deciding goals and having motivations suggests that people might perceive a connection between AIs seeking power and having autonomous capacities like deciding their own goals, having their own motivations, and being self-aware. Future research might consider the relationship between these capacities, existential risk and threat perceptions, and specific fears about AI autonomy. Concern for the treatment of LLMs with the capacity to suffer was moderate, with most people agreeing we should protect their welfare and not cause unnecessary suffering. Approximately half of AIMS respondents even agreed to the more contentious proposition that we must respect LLM personhood. This level of concern for LLMs who can suffer mirrors the results of the 2021 and 2023 main AIMS surveys where we witnessed surprisingly high moral consideration for sentient AIs. Notably, in this AIMS supplement, there was also majority support for treating all AIs with respect and agreement that torturing AIs is wrong. There was less support, although higher than might be anticipated, for legal rights for all AIs, the development of welfare standards for all AIs, and a bill of rights to protect sentient AIs. Figure A1: Regional Distributions Note. The shading shows the average responses for AI Caution. The average responses for AI Risk, Perceived Threat, the AI subservience item, Positive Emotions, AI Trust, LLM Mind Perception, LLM Suffering, AI Treatment, the bill of rights support item, the mind uploading item, and the science fiction reality item are visible by hovering over each region. Figure A2: EFA Scree Plot and Parallel Analysis Note. The scree plot is on the left and the parallel analysis is on the right. Table A1: Term Definitions Term Definition Robots/AIs “Robots/AIs are intelligent entities built by humans, such as robots, virtual copies of human brains, or computer programs that solve problems, with or without a physical body, that may exist now or in the future.” Large language models “Large language models are artificial intelligence (AI) algorithms that can recognize, summarize, and generate text from being trained on massive datasets.” Sentience “Sentience is the capacity to have positive and negative experiences, such as happiness and suffering.” Sentient robots/AIs “Sentient robots/AIs are those with the capacity to have positive and negative experiences, such as happiness and suffering.” Table A2: Key Codes and Question Text Key Question Anthropicaware Have you heard of Anthropic in the context of artificial intelligence? BDaware Have you heard of Boston Dynamics in the context of artificial intelligence? DMaware Have you heard of DeepMind in the context of artificial intelligence? OAaware Have you heard of OpenAI in the context of artificial intelligence? AFaware Have you heard of AlphaFold in the context of artificial intelligence? AGaware Have you heard of AlphaGo in the context of artificial intelligence? CGPTaware Have you heard of ChatGPT in the context of artificial intelligence? Ciceroaware Have you heard of Cicero in the context of artificial intelligence? Claudeaware Have you heard of Claude in the context of artificial intelligence? GPT4aware Have you heard of GPT-4 in the context of artificial intelligence? Sophiaaware Have you heard of Sophia in the context of artificial intelligence? NNaware Have you heard of a neural network in the context of artificial intelligence? TFaware Have you heard of a transformer in the context of artificial intelligence? TSaware Have you heard of Talos Systems in the context of artificial intelligence? (this is a foil to measure against the others for accuracy) Amariaware Have you heard of Amari in the context of artificial intelligence? (this is a foil to measure against the others for accuracy) SCaware Have you heard of a singular classifier in the context of artificial intelligence? (this is a foil to measure against the others for accuracy) chatbottrust I trust chatbots. LLMtrust I trust large language models. robottrust I trust robots. gameAItrust I trust game-playing AI. relativeLLMchatbottrust I trust large language model chatbots more than other chatbots. trainingdatatrust AI systems include many different parts. To what extent do you trust the following parts? [training data] algorithmtrust AI systems include many different parts. To what extent do you trust the following parts? [the algorithm] outputtrust AI systems include many different parts. To what extent do you trust the following parts? [output] companytrust AI systems include many different parts. To what extent do you trust the following parts? [companies] engineertrust AI systems include many different parts. To what extent do you trust the following parts? [engineers] governmenttrust AI systems include many different parts. To what extent do you trust the following parts? [governments] admiration To what extent do you, as a human, feel the following emotions towards robots/AIs? [admiration] awe To what extent do you, as a human, feel the following emotions towards robots/AIs? [awe] pride To what extent do you, as a human, feel the following emotions towards robots/AIs? [pride] compassion To what extent do you, as a human, feel the following emotions towards robots/AIs? [compassion] excitement To what extent do you, as a human, feel the following emotions towards robots/AIs? [excitement] respect To what extent do you, as a human, feel the following emotions towards robots/AIs? [respect] AID1 Do you trust that the creators of large language models (e.g., OpenAI and GPT-4) put safety over profits? AID2 Do you trust that the creators of an AI can control all current and future versions of the AI? AID3 Do you think that governments have the power to regulate the development of AI? AID4 To what extent do you agree or disagree that governments have the power to effectively enforce regulations on the development of AI? AID5 What do you think about the pace of AI development? RS1 Humanity will be able to control powerful AI systems. RS2 AI is likely to cause human extinction. RS3 The safety of AI is one of the most important issues in the world today. RS4 I support public campaigns to slow down AI development. RS5 I support government regulation that slows down AI development. RS6 I oppose government regulation that slows down AI development. RS7 I support banning the development of artificial general intelligence that is smarter than humans. RS8 I support a global ban on data centers that are large enough to train AI systems that are smarter than humans. RS9 I would consider donating to an organization working to reduce risks of human extinction from AI. SI1 Robots/AIs should be subservient to humans. SI2 Robots/AIs may be harmful to me personally. SI3 Robots/AIs may be harmful to people in the USA. SI4 Robots/AIs may be harmful to future generations of people. PMC1 I support a global ban on the development of sentience in robots/AIs. PMC3 I support a global ban on the development of AI-enhanced humans. PMC4 I support a global ban on the development of robot-human hybrids. PMC8 I support the development of welfare standards that protect the well-being of robots/AIs. PMC9 I support granting legal rights to robots/AIs. PMC10 I support campaigns against the exploitation of robots/AIs. PMC11 I support asking institutions like the government and private corporations to fund research that protects robots/AIs. PMC12 I would consider joining a public demonstration against the mistreatment of robots/AIs. PMC13 I support a “bill of rights” that protects the well-being of sentient robots/AIs. LLM1 If a large language model develops the capacity to suffer….we must ensure we don’t cause unnecessary suffering. LLM2 If a large language model develops the capacity to suffer….we must pay more attention to their welfare. LLM3 If a large language model develops the capacity to suffer….we must respect their personhood. MP1 To what extent do current large language models (i.e., those that exist in 2023, like ChatGPT) have the capacity for each of the following? [experiencing emotions] MP2 To what extent do current large language models (i.e., those that exist in 2023, like ChatGPT) have the capacity for each of the following? [having feelings] MP3 To what extent do current large language models (i.e., those that exist in 2023, like ChatGPT) have the capacity for each of the following? [thinking analytically] MP4 To what extent do current large language models (i.e., those that exist in 2023, like ChatGPT) have the capacity for each of the following? [being rational] owngoals To what extent do current large language models (i.e., those that exist in 2023, like ChatGPT) have the capacity for each of the following? [deciding their own goals] safegoals To what extent do current large language models (i.e., those that exist in 2023, like ChatGPT) have the capacity for each of the following? [maintaining human-safe goals] upholding To what extent do current large language models (i.e., those that exist in 2023, like ChatGPT) have the capacity for each of the following? [upholding human values] selfaware To what extent do current large language models (i.e., those that exist in 2023, like ChatGPT) have the capacity for each of the following? [self-awareness] sitaware To what extent do current large language models (i.e., those that exist in 2023, like ChatGPT) have the capacity for each of the following? [situational awareness] selfcontrol To what extent do current large language models (i.e., those that exist in 2023, like ChatGPT) have the capacity for each of the following? [controlling themselves] understanding To what extent do current large language models (i.e., those that exist in 2023, like ChatGPT) have the capacity for each of the following? [understanding human values] friendliness To what extent do current large language models (i.e., those that exist in 2023, like ChatGPT) have the capacity for each of the following? [being friendly with humans] power To what extent do current large language models (i.e., those that exist in 2023, like ChatGPT) have the capacity for each of the following? [seeking power] ownmotives To what extent do current large language models (i.e., those that exist in 2023, like ChatGPT) have the capacity for each of the following? [having their own motivations] MCE1 Robots/AIs deserve to be treated with respect. MCE2 Robots/AIs deserve to be included in the moral circle. MCE3 Physically damaging robots/AIs without their consent is wrong. MCE4 Re-programming robots/AIs without their consent is wrong. MCE5 Torturing robots/AIs is wrong. MCE6 The welfare of robots/AIs is one of the most important social issues in the world today. F1 Do you think any robots/AIs that currently exist (i.e., those that exist in 2023) are sentient? F11 Do you think it could ever be possible for robots/AIs to be sentient? chatGPTsentient Do you think ChatGPT is sentient? emergence If you had to guess, how many years from now do you think that the first artificial general intelligence will be created? HLAIemergence If you had to guess, how many years from now do you think that the first human-level AI will be created? SIemergence If you had to guess, how many years from now do you think that the first artificial superintelligence will be created? SF1 I am a sci-fi fan. SF2 Reality matches science fiction. upload1 In the future, humans could upload their minds into computers. Some people think that this would be very good because uploaded humans could consume fewer resources, live longer free from biological disease, and have enhanced intelligence and a greater ability to improve the world. Others disagree and think that uploading would mean that we are no longer truly human, change who we are and how we want to live, and distract us from making the real world a better place. Where would you place yourself on this scale? upload2 Do you agree or disagree with the following statement? I support humans using advanced technology in the future to upload their minds into computers. universe Which universe is the better one: YG1 How concerned, if at all, are you about the possibility that AI will cause the end of the human race on Earth? YG2 How likely do you think it is that artificial intelligence (AI) will eventually become more intelligent than people? YG3 More than 1,000 technology leaders recently signed an open letter calling on researchers to pause development of certain large‑scale AI systems for at least six months world-wide, citing fears of the “profound risks to society and humanity.” Would you support or oppose a six-month pause on some kinds of AI development? YG4 Critics of an open letter calling on researchers to pause development of certain large-scale AI systems for at least six months world-wide have argued that AI research could “create enormous social and economic benefits across the economy and society.” Would you support or oppose a six-month pause on some kinds of AI development? MCA1 Animals deserve to be included in the moral circle. MCA2 The welfare of animals is one of the most important social issues in the world today. MCEn1 The environment deserves to be included in the moral circle. MCEn2 The welfare of the environment is one of the most important social issues in the world today. attention Robots/AIs have been studied systematically since the middle of last century although the word “robot” as we know it first appeared early last century. If you read this, respond with ‘Agree’ for this item. own Do you own AI or robotic devices that can detect their environment and respond appropriately? work Do you work with AI or robotic devices at your job? smart Do you own a smart device that has some ability to detect its environment and network with other devices but that cannot respond to everything you might say or that requires you to pre-program its routines? exper Have you ever experienced any of the following? (check all that apply) fint How often do you interact with AI or robotic devices that respond to you and that can choose their own behavior? fexp How often do you read or watch robot/AI-related stories, movies, TV shows, comics, news, product descriptions, conference papers, journal papers, blogs, or other material? AI Caution Caution towards AI developments; average of PMC 1, 3, 4 AI Risk Perception of existential risk from AI; average of RS 1-3 Perceived Threat Perception of AIs as harmful; average of SI 2-4 Positive Emotions Positivity felt towards AIs; average of respect, admiration, compassion, awe, pride, excitement AI Trust Trust felt towards AIs; average of LLMtrust, chatbottrust, robottrust, gameAItrust LLM Mind Perception Attribution of mind to LLMs: average of MP 1-4, selfaware, sitware, power, owngoals, safegoals, upholding, understanding, selfcontrol, ownmotives, friendliness AI Treatment Concern for the treatment of AIs; average of MCE 1-6 LLM Suffering Concern for the treatment of LLMs; average of LLM 1-3 Exposure to AI narratives Frequency of exposure to narratives and/or information about; average of fint, fexp Moral Consideration of Nonhuman Animals Consideration of nonhuman animals; average of MCA 1-2 Moral Consideration of the Environment Consideration of the environment; average of MCEn 1-2 AIMS 2023 supplement data are published on Mendeley Data. To cite the AIMS supplement data in your own research, please use: Pauketat, Janet; Anthis, Jacy (2023), “Artificial Intelligence, Morality, and Sentience (AIMS) Survey”, Mendeley Data, V2, doi:10.17632/x5689yhv2n.2 To cite our 2023 supplemental results, please use: Pauketat, Janet V., Justin B. Bullock, and Jacy R. Anthis. 2023. “Public Opinion on AI Safety: AIMS 2023 Supplement.” PsyArXiv. September 8. doi:10.31234/osf.io/jv9rz To cite our 2021 results, please use: Pauketat, Janet V., Ali Ladak, and Jacy R. Anthis. 2022. “Artificial Intelligence, Morality, and Sentience (AIMS) Survey: 2021.” PsyArXiv. June 21. doi:10.31234/osf.io/dzgsb To cite our 2023 results, please use: Pauketat, Janet V., Ali Ladak, and Jacy R. Anthis. 2023. “Artificial Intelligence, Morality, and Sentience (AIMS) Survey: 2023 Update.” PsyArXiv. September 7. doi:10.31234/osf.io/9xsav Edited by Michael Dello-Iacovo. Thanks to Ali Ladak for writing the R functions and assisting with the implementation of R code adapted from the main AIMS survey. The AIMS 2023 supplement was preregistered and data were collected by Janet Pauketat and Jacy Reese Anthis. Data analysis and this report were conducted and authored by Janet Pauketat, Justin Bullock, and Jacy Reese Anthis. Details on the AIMS main survey methodology and results are in the AIMS 2021 and AIMS 2023 reports. Please reach out to janet@sentienceinstitute.org with any questions. [1] Of people who had an opinion and selected “Somewhat agree,” “Agree,” or “Strongly agree.” [2] Responses were census-balanced based on the American Community Survey 2021 estimates for age, gender, region, race/ethnicity, education, and income using the “raking” algorithm of the R “survey” package. The ACS 2021 census demographics are available in the supplemental file published with the data. The data weights we used are available in the R code on the Open Science Framework. The design effect was 1.02 and the effective sample size was 1,082. [3] The Mind Perception scale used in AIMS 2021 and 2023 was extended with inspiration from Ngo et al.’s (2022) paper on alignment and the capacities of AIs. [4] Robots are a type of AI, but we used the term “robots/AIs” for clarity. [5] Schäfer and Schwarz (2019) examined typical correlation and effect sizes in psychological research and considered various guidelines for interpreting effect sizes. Cohen’s traditional guidelines suggest that r = .1 is a small effect, r = .3 is a medium effect, and r = .5 is a large effect. Schäfer and Schwarz’s analysis of observed effect sizes in preregistered studies suggests that a more realistic interpretation might be r = .04 (small), r = .16 (medium), and r = .41 (large). [6] These analyses were not preregistered.Summary
Table of Contents
Introduction
Methodology
Results
Item Responses
AI X-Risks, Safety, and Caution
Regulation, Subservience, and Perceived Threat
YouGov 2023
Beliefs about AI Sentience and Forecasting Emergence
LLMs and Mind Perception
Awareness, Trust, and Positivity
Treatment of AIs
Mind Uploading, Science Fiction, and the Universe
Linear Analyses
Dimensions of LLM Perceived Mind
Correlations
Predictive Analyses
Interpreting the Results
Short Timelines, Strong Perceptions of AI Risk, and the Importance of AI Safety
The Potential Trade-Off Between Threats and Moral Consideration
Reactions to LLM Minds and Suffering
Appendix
Supplemental Results
Supplemental Methods
Citing AIMS
Acknowledgements
