Perceptions of Sentient AI and Other Digital Minds: Evidence from the AI, Morality, and Sentience (AIMS) Survey
Jacy Reese Anthis, Janet V. T. Pauketat, Ali Ladak, Aikaterina Manoli
CHI 2025
cs.AI, cs.CY, cs.ET, cs.HC
2024-07-12
Nationally representative U.S. survey (N=3,500): 18.8% of adults think some AI is already sentient, 69.5% back a ban on sentient AI development, median forecast: five years.
In February 2022, OpenAI's chief scientist Ilya Sutskever tweeted that today's large neural networks might be "slightly conscious." Months later a Google engineer went public claiming LaMDA was sentient and needed legal protection, and was fired for it. Sentient AI had moved from science fiction into the public agenda, yet the research literature had almost nothing solid on what the general public thinks. Existing studies of machine mind perception and AI moral status relied on small, one-shot Mechanical Turk samples. Basic questions had no answers: do people support protecting sentient AI from harm? Do they want it banned? Is that share rising?
The AIMS survey targets exactly this gap with three waves of nationally representative U.S. adult samples, N = 3,500 total. Wave one landed in November-December 2021, seven months before the LaMDA controversy; wave two ran April-May 2023, four months after ChatGPT's public release; a supplemental wave followed in May-July 2023.
The core design is "same questions, different years, no overlapping participants." All waves were recruited through major survey panels, balanced against U.S. census data on age, gender, race, income, and education, then weighted with raking (an iterative procedure that adjusts sample weights to match census benchmarks), for a margin of error of roughly ±3%. Anyone who took an earlier wave was excluded from later ones, so year-over-year comparisons are independent cross-sections.
Two choices stand out. First, terms were defined before being asked: sentience appeared on-screen as "the capacity to have positive and negative experiences, such as happiness and suffering," repeated on every page that used the word. Words like "suffering," by contrast, were deliberately left undefined, because the survey aims to capture how the public actually understands them. Second, every wave was preregistered, and professional forecasters predicted the results before data collection. This paid off: for the 2021 question on banning sentient AI development, the median forecaster predicted 24.4% support; the actual figure was 57.7%, more than double. Across the study, both the team and the forecasters systematically underestimated public moral concern for sentient AI, while their estimates of perceived threat were about right.
Changes from 2021 to 2023 were tested with GLMs and survived FDR correction for multiple comparisons, with one exception the authors flag themselves.
The most informative comparison is the adjective effect: adding "sentient" lifts agreement across the board. "Torturing sentient robots/AIs is wrong" gets 76.4% agreement; the identical sentence without "sentient" gets 60.7%, a 15.7-point gap.
| Item | 2021 | 2023 |
| Some existing AI is sentient | n/a | 18.8% |
| Torturing sentient AI is wrong | n/a | 76.4% |
| Support legal rights for sentient AI | n/a | 37.7% |
| Support banning sentient AI development | 57.7% | 61.5% (69.5% in supplement) |
| Support banning smarter-than-human AGI | n/a | 62.9% |
| AI may be harmful to "me personally" | 50.7% | 58.7% |
Mind perception rose across the board on 0-100 sliders: "thinking analytically" went from 62.7 to 67.1 and "experiencing emotions" from 34.3 to 36.8, all significant. Moral concern varies sharply by system type: exact digital copies of human brains rank highest (3.43 of 5) and video game NPCs lowest (2.46), with human-likeness and physical embodiment as the strongest sorting factors.
Policy attitudes lean hard one way: 48.9% say AI development is "too fast" and only 2.5% say "too slow"; 71.0% support government regulation that slows AI down; 69.1% supported the six-month pause from the March 2023 open letter, almost exactly matching YouGov's 69%, which serves as an external check on the sample.
Timelines are the most aggressive numbers: excluding those who said "never," median forecasts were two years for AGI and five years each for sentient AI, human-level AI, and superintelligence. The generational gap is wide: 25.6% of 18-35 year-olds say some existing AI is sentient, versus 10.7% of those 55 and older.
For product teams, user mind-attribution is far higher than technical communities tend to assume, and it rose within a single year. The paper organizes design risks into a 2x2 of over- and under-attribution across agency and experience. Companion chatbots (Replika and similar) face over-attribution of experience and the emotional dependence that follows, especially among vulnerable users like children. Autonomous vehicles face over-attribution of agency: extending the "moral crumple zone" concept, the authors warn that the AI itself can become the scapegoat that deflects blame from a company with weak safety practices.
On the policy side, the gap between "accelerate innovation" narratives and U.S. public opinion is wider than usually assumed. In a country generally averse to government regulation, seven in ten support regulation that slows AI down.
This is baseline data, not causal claims. Its value is that every future wave can be checked against it.
The authors list their own: the topic is so new that validated scales mostly don't exist, and many measures are original; the "years until sentient AI" question changed wording between 2021 and 2023, and they advise caution there; the rise in support for banning sentient AI had an unadjusted p of 0.046 but an FDR-adjusted value of 0.1005, just over their 0.1 cutoff, which they report plainly. The sample is U.S.-only.
A skeptical reading: the median forecast of "AGI in two years" (May 2023) has not materialized under any researcher definition, which suggests public timelines track sentiment more than technical judgment. And while sentience was defined on-screen, "positive and negative experiences" is itself contested philosophy; what the one in five who say AI is already sentient have in mind cannot be verified by a survey. Attitudes also diverge from behavior (the authors cite the privacy paradox), so these support rates do not translate directly into votes or purchases.