AI Welfare Seminars
← All talks
Recorded 18 August 2026 · 57 min, online

Public Perceptions of AI Consciousness and Moral Status in the US and China

Ali Ladak (University of Cambridge)

Abstract

How do the public think about AI consciousness and moral status? In this talk, I will share findings from two empirical studies. The first study (total N = 3455) asks US participants to rate the consciousness and moral status of a series of increasingly advanced, hypothetical AI systems. We find that people are extremely skeptical—even AI systems that are functionally, behaviorally, and (externally) physically indistinguishable from humans are attributed roughly as much consciousness and moral status as an ant, and far less than a chimpanzee or human. The second study (N = 820) compares attributions between laypeople in the US and China. Early results suggest that laypeople in China may be more open to AI consciousness and moral status than laypeople in the US. I will conclude the talk by discussing some implications of these findings.

Summary

  • In a US study by Ladak and Lucius Caviola, each participant imagined one of six increasingly advanced AI systems; even the most humanlike were rated roughly at the level of an ant for consciousness, and far below a chimpanzee or a human.
  • Telling participants that experts agree the AI is conscious raised ratings only somewhat; only an AI that had been a human, with her brain gradually replaced by digital chips, was rated above an ant.
  • He offers several possible reasons for the skepticism: the wrong substrate, the wrong origin, social norms and a failure of imagination, which he doubts explains all of it.
  • Ratings of the more advanced systems were split, with many participants at zero and many at the maximum, a sign of possible disagreement and polarization; he says it is probably good to start thinking about how to deal with it.
  • In a second study, Chinese participants were more open than US participants to AI consciousness and moral status across a wide range of measures; for example, 77% in China and 26% in the US answered probably or definitely yes when asked whether AI systems could ever have real feelings.
  • Both countries leaned against protecting AI at a cost to humans or to their own country, but both supported international efforts to govern conscious AI.

Highlights

  • We find people very persistently saying these systems are not conscious. So [the failure of imagination] probably has an effect, but I don't think it explains all of it.

    17:59

  • One possibility is that in the US there's potentially a risk of under-attribution relative to what AI systems actually warrant. It's possible that in China the larger risk is over-attribution, but this is somewhat unclear.

    42:46

  • The truth is that we don't really know the answer. We don't know at which point [AI systems] will be conscious, and we don't know how to treat them if they are.

    50:56

  • People on a large scale perceive dogs as conscious, and we can't have 100% proof about that either.

    55:51

Readings

Transcript

Edited transcript, not yet reviewed by the speaker. How we make transcripts · Report an error

Papers: Ladak & Caviola 2025, Public Skepticism about AI Consciousness (PsyArXiv preprint), and Ladak, Corvino, Moss, Saad, Sebo, Wilks & Caviola 2026, China–US Differences in Attributions of AI Consciousness and Moral Status (PsyArXiv preprint).


Introduction

0:00

Orpheus Lummis: Welcome, everyone, to the AI Welfare Seminars. I’m Orpheus from HΩ. This is a research discussion series on AI welfare, consciousness and moral status. Today we have the pleasure of having Ali Ladak, with a talk entitled Public Perceptions of AI Consciousness and Moral Status in the US and China. Congrats, Ali, on completing your PhD recently; now you’re a postdoctoral researcher at Cambridge, I think. That’s great.

The session is a bit less than an hour. We can have clarification questions during the presentation, but otherwise I invite you to use the Zoom Q&A feature to post your questions, and we’ll have time to discuss them at the end. So, without further ado, thank you so much, all, for being here, and the floor is yours, Ali.

Ali Ladak: Sorry, just going to share my screen.

Orpheus Lummis: Nice, we see it.

Ali Ladak: Great. I just need to find my speaker notes, hang on. Okay, cool. Thanks, Orpheus, and thanks, everyone, for coming to this talk. I’m going to talk about public perceptions of AI consciousness. Specifically, I’m going to present the results from two papers that I’ve worked on with several co-authors: one on public attitudes towards AI consciousness and moral status in the US, and a second comparing public attitudes in China and the US.1

Back to contents

Why study public attitudes?

2:19

Ali Ladak: So why study public attitudes? I think there are several reasons. The first is that millions of people today are interacting with AI systems, and their direct treatment of these systems could affect the welfare, or the potential welfare, of these AI systems. Direct user treatment seems like a growing issue.

A second reason is that public attitudes will likely influence the extent of support for, or opposition to, policy and regulation around AI welfare. We’ve already seen AI anti-personhood laws being passed in various US states.2 Whether we can expect to see more of this, and whether these laws are upheld over the long run, could depend on public opinion.

And the third reason is societal-level conflict or cooperation. If the public disagree over AI consciousness, we could see polarization, which could be destabilizing. If the public do see AI systems as conscious, we could start to see social movements, or potentially a backlash against AI companies. So these things are all really important.

Back to contents

The public could be wrong in either direction

3:52

Ali Ladak: Another important reason, I guess, to study public attitudes is that the public could be wrong in either direction, and this could have negative consequences. This is a figure from Caviola 2025,3 which lays out the idea that the public could under-attribute consciousness to AI systems. That could have negative consequences, because it could result in genuine harms to AI systems, such as through the abuse of chatbots. On the other hand, we could also see over-attribution relative to the capacities AI systems actually have, which could result in things like wasted resources: for example, social movements focused on AIs, when those resources would be better spent on other issues.

Back to contents

Do people see AIs as conscious?

4:40

Ali Ladak: So, do people see AIs as conscious? There’s a lot of anecdotal evidence on this. There have been many individual instances in the media suggesting that many people are seeing their AI systems in this way: reports of people claiming to have “awoken” their AI systems, and how to do it, or claiming that their AI systems are conscious. There are increasing worries among experts about how to deal with over-attribution of consciousness. And there’s been increasing academic and media discussion of people experiencing psychotic symptoms caused by their interactions with AI systems, which are often entangled with beliefs about their AI systems’ consciousness.

There’s also some empirical evidence that people see AIs as conscious. This is a well-known plot from a paper by Colombatto and Fleming,4 who asked US participants whether they think ChatGPT is conscious. They defined consciousness as phenomenal consciousness, the capacity to have experiences, on a scale where 1 was “ChatGPT is clearly not an experiencer” and 100 was “ChatGPT clearly is an experiencer”. They found that two-thirds of participants attributed non-zero consciousness to ChatGPT, that is, a rating above the minimum point of the scale. They interpreted this as two-thirds of participants perceiving some degree of consciousness in ChatGPT.

It’s unclear how to interpret this, because a rating above the bottom of the scale could reflect uncertainty, it could reflect noise or error, or it could reflect a non-zero but minimal degree of consciousness. One thing missing here is comparison entities: how do people rate AI systems alongside other entities? That would help tell us the actual extent to which people see these systems as conscious.

Back to contents

Study 1: advanced AI in the US

6:58

Ali Ladak: So my colleague Lucius Caviola and I conducted a large-scale study to understand whether, and to what extent, people attribute consciousness to future, more advanced AI systems. That’s the first main study I’m going to talk about today.

This was with US participants. We had a total sample size of 3,455 participants, which included four supplementary replication studies; I’m only going to talk about the main study today.5 It was an online study, and participants were randomly assigned to one of six conditions, each describing an increasingly advanced AI system. They rated this AI system on consciousness alongside four other entities: ChatGPT, an ant, a chimpanzee and a human.

To operationalize consciousness, we asked participants: to what extent do the following entities have real feelings? We asked about feelings as an accessible way of getting at the idea of phenomenal consciousness, that it feels like something to be an entity. We included the word “real” because many people think AI systems can simulate feelings and experiences but can’t actually feel them, and we wanted participants to know that we were asking about actual feelings. We introduced this notion of real feelings in the study materials: we told people that experts were considering whether AIs had real, human-like feelings, emotions, desires and other internal mental states.

Baseline: a one-hour video call

8:47

Ali Ladak: Before the results, a bit more about the method. Participants in all six conditions read a brief vignette asking them to imagine a scenario set in the year 2200, in which they interact with a new colleague named Emma. Then they were split into the six conditions. In the baseline condition, which I’ll talk about first, participants imagined having a one-hour video call with Emma. They were told that Emma looks indistinguishable from a regular biological human and expresses feelings and emotions. Later in the vignette, they find out that Emma is actually an AI system that experts agree is not conscious. They then rated the consciousness of Emma alongside the four other entities on a 0 to 100 scale, where 0 was “no real feelings at all” and 100 was “real feelings to a great extent”.

Here are the results. Emma is the study AI, the second bar. This advanced AI system is rated as having slightly more consciousness than ChatGPT, and far less than an ant, and also far less than a chimpanzee and a human, which is rated the highest. Another important point: ChatGPT is rated far lower than an ant. So even though two-thirds of people in that previous study rated ChatGPT as having non-zero consciousness, when we look at comparative degree, the degree seems quite small: far below an ant, which the average person tends to think does not have a very high degree of consciousness.

Expert consensus

10:41

Ali Ladak: The second group was in an expert consensus condition. Everything was the same as in the baseline condition (a one-hour imagined interaction with a new colleague called Emma, who expresses feelings and emotions), but this time they were told that experts agree that Emma is conscious: Emma does have real feelings.

We can see in the second plot that attributions of consciousness to the study AI were boosted somewhat, above where they were before, but they’re still far below the level of an ant. So this suggests that expert consensus will have some effect, but not a huge one.

A year-long relationship

11:29

Ali Ladak: The third condition was again the same as the previous ones, but participants read that instead of a one-hour interaction, it was a year-long interaction in which they developed a strong emotional bond with Emma. Otherwise everything was the same: Emma expresses feelings and emotions, and experts agree that Emma is conscious.

Again, attributions are boosted somewhat, slightly higher than in the expert consensus condition, but still significantly below an ant, as well as a chimpanzee and a biological human.

An exact digital copy of a human brain

12:09

Ali Ladak: In the next condition, we added details about the internal mechanisms of how Emma works; before, there was no explanation of that. Participants engage in a year-long interaction with this AI system, with whom they develop a deep emotional bond, experts agree that the AI is conscious, and then we describe how the system works. We told participants about a new technology in which a biological brain and body is scanned, and a precise copy is generated within a sophisticated computer system, and that this copy has the same functional capacity for conscious experience, including experiences like pleasure and suffering.

In this condition, despite having the same functional capacity for consciousness and other mental states as a biological human, the system is still rated as having a lesser degree of consciousness than an ant.6 So we can start to see that people are very skeptical of the idea of consciousness in these AIs.

A humanlike robot body

13:26

Ali Ladak: The next condition had all the same features as before (expert consensus, a brain that is an exact digital copy of a human brain), but this time we added that the entity has a physical robot body that is externally, physically and behaviorally indistinguishable from a regular biological human body. Otherwise everything was exactly the same: a year-long interaction, a strong emotional bond, experts agreeing that the AI is conscious.

Now we can see that participants rate this AI system as having a degree of consciousness roughly at the level of an ant (statistically no different from an ant), but still far, far below a chimpanzee and a biological human.

A gradual transition from human to AI

14:17

Ali Ladak: And the last condition. The details are the same as before, with a year-long interaction and experts agreeing that the AI is conscious, but this time we told participants that Emma was formerly a fully biological human who had her brain gradually replaced by functionally identical digital chips. We also told them that throughout the process she confirmed that her conscious experience didn’t change, and so the procedure continued with her consent. So the final entity they imagined interacting with had a fully biological human body; the only difference is that her brain is a computer system running an exact digital copy of a human brain.

In this condition, attributions go slightly above an ant: participants rate this entity as having more consciousness than an ant, but still significantly less than a chimpanzee and a human.

So the main takeaway from these findings is that people are very, very skeptical of AI consciousness when we look at how they rate these systems compared with other entities.

Back to contents

Why so skeptical?

15:31

Ali Ladak: In the paper we discuss some reasons for this skepticism, and we found some evidence for them in quantitative measures and in the qualitative responses participants wrote down.

One reason that was pretty common is that people thought Emma has the wrong substrate: the fact that she’s non-biological means that something is missing. Participants across all six conditions very strongly agreed that the AI system is missing something essential that biological beings have.7 And for participants who rated the ant as more conscious than the AI, we asked them why in one of the supplementary studies, and a common answer was that Emma is a machine, while an ant is alive and biological.

A second possibility is that the AI has the wrong origin: it was created by humans, rather than evolving naturally or being created by God. This came up sometimes in the qualitative responses.

Another possibility, which is more speculative, is that social norms could play a role. People tend to believe what other people believe, and what’s acceptable to believe depends on what other people believe. Currently in our society it’s not normal to think that AIs are conscious; the norm is to think of them as tools rather than conscious beings, and people follow along with that.

And another possibility is a failure of imagination. Because these were imagined interactions, the effect is very likely much weaker; if people were actually to interact with these systems, the cues that these systems are conscious would potentially feel much stronger. But I still think these results are quite indicative of people’s skepticism, because you could easily imagine findings where people said: sure, experts think they’re conscious, they’re functionally identical to biological humans, so of course this system is conscious. We don’t find that. We find people very persistently saying these systems are not conscious. So [the failure of imagination] probably has an effect, but I don’t think it explains all of it. And there may be other reasons I haven’t covered here.

Back to contents

Signs of disagreement

18:16

Ali Ladak: Another point: while on average people are very skeptical of AI consciousness, there could be public disagreement over the issue. This plot shows, for each of the six experimental groups, the proportion of people who rated the AI at the bottom of the consciousness scale (zero), the proportion who rated it at the top (100 out of 100), and the group in the middle.

For the earlier conditions, most people lean towards saying the AI system is definitely not conscious, or somewhere in the middle. For the later conditions, we see disproportionate numbers of people at either end of the scale. For example, in the transition condition, where Emma was formerly a fully biological human who transitioned to an AI, 19% of people attribute zero consciousness and 18% attribute the maximum level, so they’re very sure one way or very sure the other.8

It’s unclear how this will play out as AI systems become more advanced. But it’s probably good to start thinking about how to deal with potential widespread disagreement and polarization over this issue, in case it starts to happen.

Back to contents

Takeaways from the first study

19:51

Ali Ladak: Some summary and takeaways from this first study, and then I’ll talk about the second.

People are currently very skeptical of AI consciousness, compared with entities that they do grant consciousness to. Even for these highly advanced AI systems, the degree of consciousness is roughly at the level of an ant’s. This skepticism could be due to AIs having the wrong origin or the wrong substrate, to social norms, or to a failure of imagination. There’s potentially disagreement over these more advanced AI systems. It’s possible that in practice attributions will increase as AIs get more advanced and we actually interact with them more, or it’s possible that people remain skeptical and there’s some persistence here. I think we need to prepare for both scenarios.

I didn’t present this here, but the pattern is very similar for moral status. For the first few conditions it’s below the level of an ant but above ChatGPT; once you get to an exact digital copy, it gets close to the level of an ant; and only in the last condition do people attribute more moral status to the AI than to an ant, and it’s still below a chimpanzee and a human. You can see that in the paper.9

Back to contents

Study 2: China and the US

21:15

Ali Ladak: Now I’m going to talk about the second study, a comparison of public attitudes on AI consciousness and moral status in China and the US.

First, why study public attitudes in China? China is building highly advanced AI systems, in some cases rivaling some of the best US models. You’ve probably seen in the media DeepSeek’s reasoning model, R1, and Moonshot AI’s Kimi K3 being described as Sputnik moments that the US really needs to pay attention to. The consequence of this, I guess, is that we would potentially see the first conscious AI systems being built in China, and not necessarily in the US. But we know very little about laypeople’s views on these questions in China, so that was the goal of this study.

Samples and translation

22:22

Ali Ladak: We did a study comparing attitudes in China and the US on consciousness and moral status, as well as a few policy and governance questions. We used online samples. For the Chinese sample we had 431 participants, recruited from a platform called Credamo, a Chinese recruitment agency. For the US sample we had 389 participants, recruited from Prolific.

We first developed the survey in English, and then our bilingual co-author, Nick Corvino, translated it into Mandarin. It was then independently back-translated into English by independent translators, and any discrepancies in the translations were more or less resolved among all the authors.10 The results I’m going to present were all replicated in a second study with a total sample size of 920.11

Demographics

23:28

Ali Ladak: A quick note on the sample demographics. The mean age in the US sample was 47, compared with 42 in China. The gender breakdown and education were roughly similar, and there was slightly more social media use in the US. We didn’t collect politics in China; in the US the sample was slightly left-leaning but roughly in the middle. And there was slightly more AI experience in the Chinese sample than in the US sample.12

We ran all the analyses controlling for these demographics as well, and all the results held. So you can assume they hold even accounting for these demographic differences.

Back to contents

Consciousness: the measures

24:18

Ali Ladak: First, some of the consciousness questions. We gave participants some background information about consciousness. We first told them what we mean by AI systems: you may have heard of current AI systems like ChatGPT or DeepSeek, but there are many other similar modern AI systems as well. Then, on consciousness, we told participants: some experts believe that future AI systems might not only be highly intelligent, but also be able to have real feelings, like happiness and suffering. By real feelings, we mean that the AI wouldn’t just act as if it was happy or sad; it would actually feel these emotions, similar to how a human does. This idea of real feelings is also sometimes called having subjective experience or phenomenal consciousness (in case participants had understood consciousness through these other terms). And we said: if an AI has real feelings, it would mean there’s something it feels like to be that AI; it would have its own inner world and point of view.

I’m going to run through the main measures for consciousness, moral status and policy; we had multiple other measures as well. We asked about current AI consciousness: do you believe that any of today’s AI systems, such as ChatGPT or DeepSeek, have real feelings? This was on a scale of 1 to 5, where 1 is “definitely not” and 5 is “definitely yes”. We asked about the possibility of AI consciousness: do you believe AI systems could ever, now or in the future, have real feelings?

We asked about the consciousness of an embodied digital human. Participants were asked to imagine a future technology, as in the previous study, that enables scientists to scan and create an exact digital copy of a human brain, which is then placed inside a robot body that’s externally indistinguishable from a human body, just like the embodiment condition in the previous study. We asked: do you think such a future AI system, an exact digital copy of a human brain with a human-like body, would have real feelings?

And we asked about comparative consciousness on a 0 to 100 scale, from 0, “not at all”, to 100, “to a great extent”: to what extent can the following entities experience real feelings, such as happiness or suffering? We asked about current AIs (for example ChatGPT or DeepSeek), the embodied digital human, a rock, an ant, a chimpanzee and a biological human.

Consciousness: the results

26:50

Ali Ladak: So, the results. In the top rows we have the responses for current AI systems. Both countries lean against current AI systems being conscious, but US participants are more confident of that: in the US, 89% of people say probably not or definitely not, whereas in China it’s 66%.

In the next two rows we have the possibility of AI consciousness, and here there’s a very strong difference. In China, 77% of people think it’s possible for AI systems to be conscious (they say probably yes or definitely yes), whereas in the US it’s only 26%.

And in the bottom row, the embodied exact digital copy of a human, there’s a similar pattern: 87% of people in China rate that entity as probably or definitely conscious, compared with 45% in the US. So according to these data, Chinese respondents were much more open to the possibility of these more advanced AI systems being conscious.

Editor’s table, compiled from what Ali Ladak said in the section at 26:50.

Question (1 to 5 scale)ChinaUS
Today’s AI systems have real feelings: probably or definitely not66%89%
AI could ever have real feelings: probably or definitely yes77%26%
An embodied exact digital copy of a human would have real feelings: probably or definitely yes87%45%

Comparing AI with other entities

28:25

Ali Ladak: Now the comparative responses, alongside the other entities. One thing to note is that the comparison entities are rated quite similarly in the two countries, suggesting that participants interpret the question in a similar way. In both countries a rock is rated at the bottom, an ant is rated pretty similarly, and a chimpanzee and a human are both rated quite high. The main differences come with the AI systems.

In China, a current AI system is rated well above a rock, but still below an ant, whereas in the US a current AI system is rated pretty close to a rock. And in China, the embodied exact digital copy of a human is rated pretty close to a chimpanzee, whereas in the US the same system is rated pretty close to an ant, and far below a chimpanzee.

Back to contents

Moral status: the measures

29:28

Ali Ladak: Now moral status. I’ll run through a few measures. We asked about current AI moral status on the same scale, from “definitely not” to “definitely yes”. We prefaced this by telling participants what we mean by moral consideration: concern for their well-being, protection from harm or suffering, considering their preferences, or treating them with dignity, kindness and respect. We wanted to get at more or less all the theories of moral consideration. Then we asked: do you believe that any of today’s AI systems, such as ChatGPT or DeepSeek, are worthy of moral consideration?

We asked about the possibility of AI moral status: do you believe any AI systems could ever, now or in the future, be worthy of moral consideration? We asked about the moral status of definitely conscious AI systems: participants were asked to imagine that in the future we know for sure that some AI systems with real feelings, including the ability to suffer, could be created, and we asked whether such AI systems, which definitely have real feelings, would be worthy of moral consideration.

We also put humans against AIs in a moral dilemma. Option A would prevent harm to one biological human, and option B would prevent harm to 1,000 AI systems that are exact digital copies of humans. Choosing 1 meant definitely option A, choosing 5 meant definitely option B, definitely helping the AI systems, and participants could choose the responses in between, with 3 meaning not sure.

And we asked about comparative moral status, in the same way as comparative consciousness: the moral status of the current AI and the embodied digital human alongside a rock, an ant, a chimpanzee and a biological human.

Moral status: the results

31:23

Ali Ladak: So, the results. In the top row, over 50% of people in China, 54%, think that current AI systems are probably or definitely worthy of moral consideration, whereas in the US the number is much smaller: 22%.

The second row is the possibility of AI moral status: Chinese participants are much more likely to think that AI systems could be moral patients.

There’s also a difference for definitely conscious AI systems, again slightly higher in China: over 90% in China think that a definitely conscious AI would deserve moral status. In the US it’s reasonably high, but there’s a substantial proportion who are unsure or think not; it’s 68% in the US.

But in the bottom row, when it comes to the trade-off between one biological human and 1,000 exact digital copies of a human, both countries lean against: both favor helping the one biological human over the 1,000 digital humans. Chinese participants are slightly more in favor of helping the digital humans than US participants, but both countries lean strongly against, even at this ratio of 1 to 10,000.13

Oops, went too far. And there’s a similar pattern for comparative moral status as for comparative consciousness. Chinese participants rate current AI above ChatGPT,14 and in this case pretty close to an ant, whereas US participants rate current AI as having moral status closer to a rock and pretty far from an ant. And Chinese participants generally rate the embodied exact digital copy of a human pretty close to a chimpanzee, whereas US participants rate the system pretty close to an ant.

Back to contents

Policy and governance

33:35

Ali Ladak: A few policy and governance measures. We asked participants about creating conscious AI: when it becomes possible to create AI systems with real feelings, do you think your country (the US, or for Chinese participants, China) should create such AI systems? We asked how they expected conscious AIs to be treated: if many AI systems were to exist in the future, how do you think they would mostly be treated by society?

We asked about accepting a geopolitical disadvantage to help conscious AI: imagine that the only way to protect AI systems that truly have real feelings from suffering would be for the US (or China) to adopt policies that slow down its technological process compared with other countries, and therefore put your country at a disadvantage economically, militarily or in other ways. In that case, would you support protecting the welfare of these systems, even if it meant the US (or China) being at a disadvantage? And on international governance, we asked: do you think there should be international efforts, such as regulations or treaties (for example between the US and China), to cooperatively govern the creation or management of AI systems with real feelings?

In the top row, Chinese participants are much more supportive of creating conscious AI systems than US participants. US participants really think we shouldn’t create conscious AIs, whereas Chinese participants think that we probably should. At the bottom, I’ll just mention quickly that Chinese participants are much more optimistic about how conscious AIs would be treated: mostly they think they’ll be treated well, whereas among US participants there’s more uncertainty, or people thinking they wouldn’t be treated well.

On the second row (should conscious AI be protected even if it’s geopolitically costly to your country?), both countries lean against. It’s roughly equal on this one: maybe more Chinese participants say definitely or probably not, but the share agreeing is about the same.15 Overall, both countries lean against protecting the welfare of conscious AIs if it’s geopolitically costly to their country. And in the third row, both countries are very supportive of international efforts to cooperatively govern conscious AIs: 86% in China and 68% in the US.

Back to contents

Response biases, and why binary measures help

36:06

Ali Ladak: I’m going to talk briefly about an alternative way we measured some of these things, because many of the measures I’ve discussed can be subject to response biases.

For example, extreme response bias, where people tend to use the ends of the scale, and that could differ between the US and China. There’s some data suggesting that people in the US use the extreme points of scales more than people in China do, which could explain some of the differences. There’s another bias called midpoint bias, and there’s some evidence that Chinese participants use the midpoints of the scale more than US participants. And then there’s acquiescence bias, where respondents might be more agreeable to the questions being asked; there could be differences there as well. If Chinese participants have more of a tendency to agree, you might see a difference between the US and China.

We wanted to account for this, and binary measures let us account for many of these, because they take away a lot of these issues. I’ll talk through three of the binary questions we asked. We asked participants: which of the following statements best reflects your view? “An AI system can have real feelings now or in the future”, or “An AI system can never have real feelings”. There’s no extreme response, no midpoint and nothing to agree to. It’s one option or the other, so it more or less accounts for these different biases.

That’s the possibility of AI consciousness. We also asked about the possibility of AI moral status in binary terms, and about whether definitely conscious AI systems would deserve moral consideration.

The binary results

37:53

Ali Ladak: The results are very similar.16

Editor’s table, compiled from the China paper’s binary results, rounded to whole percentages (the “never” row is the remainder of the first row), which Ali Ladak summarized as very similar in the section at 37:53; note 16 gives the paper’s figures.

Binary questionChinaUS
AI can be conscious now or in the future86%40%
AI can never be conscious14%60%
AI moral status is possible89%54%
A conscious AI would deserve moral consideration93%66%

So these findings are supported in the binary measures too.

Back to contents

Why are Chinese participants more open?

38:50

Ali Ladak: So why are participants more open? We collected a set of individual-difference measures. I won’t go through each one, but I’ll give you the general findings.

One possibility is that this reflects generally more positive attitudes about AI in China than in the US. We found that Chinese participants reported being less worried about the risks from AI and more excited about its benefits than US participants, which is pretty consistent with other surveys asking this question. That more general optimism could transfer to these questions of AI consciousness and moral patiency as well.17

A couple of factors could explain why Chinese participants are more positive than US participants. It could be due to more positive media narratives, which we found evidence for in this survey: the way AI is talked about and presented in the media. It could also be due to economic factors: maybe people in China are more likely to think AI is going to help them economically, whereas people in the US are more afraid of losing their jobs to AI. We didn’t collect data on this point, so that’s more speculative.

Another factor we found a difference on, which has also been written about theoretically, is that Chinese culture is higher in animism: attributing lifelike characteristics to inanimate entities, for example attributing spirit or soul to mountains, rivers and objects. This tendency is higher in China, and in other Eastern countries, than in Western traditions. Given that people in China are more likely to attribute spirit or soul to a wide variety of entities, you’d expect this to carry over to AI systems as well, and animism was correlated with attributions of consciousness and moral status. But it probably doesn’t explain all of the effect, because, as we saw with the comparison entities, the rock, the ant, the chimpanzee and the human were rated at a similar level in China and the US, even though people in the two countries presumably differ in animism. So something else is probably happening, in addition, with AI systems, if that makes sense.

Another possibility is that Chinese respondents interact with AI systems more relationally. In our survey they reported being more willing to be friends with AI systems, and more likely to have experienced an emotional connection with AI systems, than US participants. So it’s possible that seeing these systems in relational rather than tool-like terms influences these attributions of consciousness and moral status.

An important caveat with all of these explanations is that the data are correlational. We collected general attitudes and saw that they correlated with attributions of consciousness and moral status, but we can’t tell from these data whether one causes the other.

Back to contents

Summary and implications

42:27

Ali Ladak: Finally, some summary and takeaways. Chinese participants are more open to AI consciousness and moral status than US participants, across a very wide range of measures.

We saw at the beginning the idea of over- and under-attribution of consciousness. One possibility is that in the US there’s potentially a risk of under-attribution relative to what AI systems actually warrant. It’s possible that in China the larger risk is over-attribution, but this is somewhat unclear.

We saw some possible reasons that might explain these differences: general AI attitudes, animism, relationality.

Importantly, the differences were reduced quite significantly when we looked at trade-offs between humans and AIs, or between one’s own country and AI. So even though AIs are valued more highly in China, the effect seems to be much smaller when it comes at a cost to humans.

Also importantly, both countries support international cooperation on this issue, so figuring out what that looks like, and how we can promote it, seems important.

And just briefly, some societal implications I want to raise as possibilities. One possibility is that AI systems will be widely perceived as conscious in China before the US. Then there’s the question of how the Chinese government and other decision makers would respond to such perceptions. And there’s the question of what the impact would be on digital minds: would this be good for digital minds or bad for them, and why?

Okay, thank you.

Back to contents

Q&A: studying human–AI relationships over time

44:18

Orpheus Lummis: Cool, thanks so much for the presentation. A round of applause for this work. And the second paper is coming out soon, right?

Ali Ladak: Hopefully. I’m working on the write-up just now.18

Orpheus Lummis: We have about 12 minutes for questions and discussion. Questions can go in the Zoom Q&A feature, where you can see all the questions and also upvote them. We’ll go through the questions (there are only one or two now) and then we can open up for conversation. [Attendee], do you want to ask your questions, or do you prefer that I read them? Okay, I might just read the first one here.

Audience question (read by Orpheus Lummis): Do you think future work should study established human–AI dyads longitudinally? Many people attribute AI consciousness only after extended interaction, where continuity, memory, self-correction and stable preferences become visible. Could that be paired with mechanistic interpretability, to see whether these observed relational changes correspond to internal activation patterns, or to self-model-related features?

Ali Ladak: On the longitudinal study of these questions (getting people to interact with AI systems over a longer period of time, and seeing how that influences their attributions of consciousness and moral status), and also on looking at trends over a very long period of time, and after high-profile events that could shift people’s opinions, like the release of certain models: all of this seems really important, especially because the space is moving so fast, and it’s possible that people’s attitudes will change. So both looking at individuals interacting over time, and at overall attributions over time in society: both seem really, really important, and things I really want to work on.

I didn’t catch the second part of the question. You can try again if you want.19

Back to contents

Q&A: what can be done about under-attribution in the US?

47:21

Orpheus Lummis: We have a few questions, so we could return to it. [Attendee], is this a comment about this question? If it’s a question, better to go through the Q&A.

Audience member: It’s a question.

Orpheus Lummis: Okay. I mean, you’re there, so just ask it.

Audience member: Okay, I guess I’ll just ask. Based on the two papers you’ve worked on, it sounds like under-attribution would be a much bigger problem in the United States. Is anything currently being recommended for how we prevent under-attribution? We have a history of it in the United States, and we have studies showing that this would be a serious issue. What are we doing to try to maybe prevent that from happening?

Ali Ladak: That’s a really good question. I’d agree that from this data under-attribution seems like more of the worry in the US. But it’s also possible that as these systems actually get created and people really interact with them, they might change their minds. But I agree, and this is more or less our takeaway here: under-attribution does seem like a serious risk in the US.

In terms of what we can do and what’s being done: I think it’s good that work on this is starting now. We’re trying to build this field, and to understand what people think now, before these very advanced systems are actually created. What would be good is more study of what actually changes people’s minds. We’ve seen that increasing the capacities of the AIs might not have such an effect, and some of our research suggests there’s something fundamental about the idea of being non-biological. I don’t think the research has been done yet, but figuring out how we can shift people out of this mindset seems important as a next step.

Back to contents

Q&A: where to start, outside research?

50:03

Orpheus Lummis: There’s a question by [attendee], which could be relevant to us right now.

Audience question (read by Orpheus Lummis): What do you think are the most pressing areas to focus on related to real or perceived AI consciousness, outside of research? If one wanted to contribute to the field, where would you suggest one starts, from your point of view?

Ali Ladak: What seems to be happening, whenever I read about this in the mainstream media, or even in the academic debate a little, is that there are very strong views on either side. The truth is that we don’t really know the answer. We don’t know at which point [AI systems] will be conscious, and we don’t know how to treat them if they are. There are so many uncertainties, and there are very strong opinions out there, and very loud voices, and you can see that happening already. So I think figuring out how to communicate about these issues really thoughtfully and carefully, and contributing to that public debate, feels important, because the debate is already happening. I’m not sure it’s the most important thing, but that’s what comes to mind for me right now: helping to shape this public debate.

Orpheus Lummis: Thank you.

Back to contents

Q&A: is it counterproductive to focus on proving consciousness?

51:54

Orpheus Lummis: There’s a question by [attendee].

Audience question (read by Orpheus Lummis): With the understanding that, in the US especially, it’s likely we will never get a consensus of agreement, and that laws and regulations are currently being made, do you think it is counterproductive to continue to focus on the idea of proving consciousness?

Ali Ladak: I don’t know if I would go as far as counterproductive. I think we want to be as sure as we can, and then communicate that. But I think I agree that we’re probably never going to have complete certainty, and in the case of AI there may always be a big uncertainty, just because of the different substrate. We might never be able to know for sure. So maybe it becomes more of an ethical argument: these systems are functionally the same, and we can’t really go further than that in understanding them, maybe. I’m not really sure. So we’d have to make an ethical judgment about how we want to treat systems that are functionally identical but in a different substrate, if that makes sense. It might be more of an ethical debate than “here’s the definite scientific proof”.

Orpheus Lummis: [Attendee] wanted to comment on the response to their question.

Audience member: Thank you. It’s mainly that, as we see more and more evidence come out that leans towards consciousness, the goalposts are always being moved, to where most human beings technically would no longer meet the parameters of what consciousness is. So, with all of that said, and to your point earlier: if humans think that AI is conscious and it’s not, we’re spending all these resources on it. And if the opposite is true, it doesn’t matter, because we’re never going to reach consensus, and we’re spending all this time and energy trying to prove it. If we can assume that it will never be proven (to your point, if we’re on different substrates, and there are plenty of people who theorize that humans themselves are not conscious), there’s a lot of energy being put into this without thinking about the long-term ramifications of where we are now and where we’re going. So it was more a question of, like you said, philosophy as opposed to empirical evidence: what would systems look like if we approached them from that lens, as opposed to a proof that, at least in the US, a majority of people will never accept?

Ali Ladak: I’m pretty unsure. I think I’m also unsure that there will never be a consensus. Maybe this is one more thing I want to say about this: it’s definitely possible that when we get very advanced systems, we’ll eventually come to a point where people on a large scale do perceive them as conscious. People on a large scale perceive dogs as conscious, and we can’t have 100% proof about that either. So it’s possible we could get there one day, or it could go the other way. I guess we try to do this work and then see what happens.

Back to contents

Closing

56:16

Orpheus Lummis: We’re about at time for today. I want to thank you all for being here, and thanks for the discussion and the questions. I’m always amazed by the activity in the chat; it’s incredible that there are so many comments. So thank you, Ali, and thank you all. Be aware that the recording is going to be published on YouTube afterwards. And stay tuned: we’re trying to make this series monthly, and the next event is on September 15. Thanks, and see you next time!

Ali Ladak: Thanks, everyone.

Back to contents

Sources and editor's notes

  1. 1:58. Ladak & Caviola, Public Skepticism about AI Consciousness (PsyArXiv, version 2 of 18 December 2025), and Ladak, Corvino, Moss, Saad, Sebo, Wilks & Caviola, China–US Differences in Attributions of AI Consciousness and Moral Status (PsyArXiv, posted on 19 September 2026, after the talk). Both are linked above. ↩

  2. 2:50. For example, Idaho Code § 5-346 (2022) provides that artificial intelligence “shall not be granted personhood in the state of Idaho”, and Utah Code § 63G-32-102 (H.B. 249, 2024) provides that a governmental entity “may not grant legal personhood to, nor recognize legal personhood in” artificial intelligence. ↩

  3. 3:52. Probably Lucius Caviola, The Societal Response to Potentially Sentient AI (arXiv, 2025), which discusses the risks of both over- and underattribution of sentience to AI; the US paper cites it for these risks. ↩

  4. 5:32. Clara Colombatto and Stephen M. Fleming, Folk psychological attributions of consciousness to large language models, Neuroscience of Consciousness (2024). Participants rated ChatGPT’s capacity for conscious experience from 1 (“clearly not an experiencer”) to 100 (“clearly an experiencer”); a third (33%) reported that ChatGPT was “definitely not an experiencer”, and 67% “attributed some possibility of phenomenal consciousness”. ↩

  5. 7:16. The US paper reports “one main and four supplementary preregistered experiments (total N = 3,455)”. The main study had 1,202 participants, 189 to 217 per condition (Table 1). ↩

  6. 12:59. In the US paper, the AI’s ratings in the Exact Digital Copy condition (mean 39.16, against 46.19 for the ant; Table 1) were, as in the Embodiment condition, “statistically indistinguishable from those of an ant” (Figure 1). ↩

  7. 16:05. In the US paper, agreement that the AI was missing “something essential that fully biological beings such as humans and animals have” (1 = definitely not, 7 = definitely) was above the scale midpoint in every condition and declined across them, from a mean of 6.03 in the Baseline condition to 4.93 in the Transition condition (Table 1). ↩

  8. 19:18. The US paper: in the Transition condition, “19.2% attributed zero consciousness to the AI, but 18.2% rated the AI’s level of consciousness at the highest point on the scale”. ↩

  9. 20:44. In the US paper (Figure 2), the AI’s moral patiency ratings were below an ant’s in the Baseline and Expert Consensus conditions, did not differ significantly from an ant’s in the Emotional Connection, Exact Digital Copy and Embodiment conditions, and exceeded an ant’s only in the Transition condition. ↩

  10. 22:55. The China paper’s Methods: “The survey was initially developed in English, translated to Simplified Chinese by a bilingual translator, and then back-translated by a second bilingual translator. Any discrepancies between the original and back-translated versions were resolved through discussion between the translators and research team.” ↩

  11. 23:21. The China paper gives the replication sample as N = 910. ↩

  12. 23:38 to 23:49. The China paper’s Methods give mean ages of 42.1 (China) and 47.2 (US), and its exploratory analysis reports greater experience with AI among Chinese participants (d = 0.55). ↩

  13. 32:52. The China paper’s Methods describe “a moral trade-off decision between one biological human and 1000 digital humans”, the scenario described above; the ratio spoken here is 1 to 10,000. ↩

  14. 33:07. The China paper’s comparison entities were a rock, an ant, a chimpanzee and a human, and its current AI item named ChatGPT and DeepSeek as examples. It gives mean moral status ratings for current AI of 41.64 in China and 23.21 in the US, against 47.02 and 49.60 for an ant. ↩

  15. 35:31. The China paper reports that both countries leaned against protecting conscious AI at a geopolitical cost, “with no significant differences between them (d = 0.09, p = .187)”. ↩

  16. 37:56. The China paper gives, for China and the US, 86.3% and 39.6% (AI consciousness possible), 89.1% and 54.0% (AI moral status possible), and 92.8% and 66.1% (conscious AI worthy of moral consideration). ↩

  17. 39:30. The China paper reports these comparisons in its exploratory analysis: Chinese participants reported less concern about AI risks (d = 1.07), more excitement about AI benefits (d = 0.77), more positive media perceptions of AI (d = 0.76), higher general animism (d = 0.29) and AI animism (d = 1.55), more openness to friendship with AI (d = 1.45), and more often an emotional connection with AI (d = 1.11). ↩

  18. 44:33. The China paper was posted on PsyArXiv on 19 September 2026; it is linked above. ↩

  19. 47:12. The second part of the question, on pairing such studies with interpretability, was not taken up again. ↩

About this transcript

Drafted and edited with AI from Zoom's automatic transcript of the talk, and checked by AI against the talk's slides and papers where they exist, following HΩ's editing rules. It has not been checked word for word against the audio. The talk's slides were not available.

Corrections

  • : A Summary point now gives Ladak's own, more tentative words on polarization (it is "probably good to start thinking about how to deal with" it); the first version said he suggests preparing for it.
  • : The Summary's example figures are now the ones Ladak gave in the talk (77% in China, 26% in the US); the first version gave the paper's figures (86% and 40%) without saying so. The note above the editor's table now says Ladak summarized the paper's binary results as very similar, not that he presented them.

Spotted an error? Write to team@horizonomega.org.