Before reading this post, please take 30 seconds to answer the 3-question survey yourself: [survey link]
Introduction
The possibility of AI consciousness, the use of AI for companionship or emotional support, and the trust we place in AI as coding tools raise personal interrogatives for most AI researchers, but these topics are often relegated to informal conversations. Apart from the few who write about it, it’s hard to get a sense of how much AI has permeated researchers’ personal lives, and how they feel about it. To gain a better understanding of this, we ran a quick, informal three-question survey, plus an optional fourth, open-ended question.
We spoke to 107 EMNLP 2026 attendees in Budapest. 79 respondents were randomly-met researchers, collected during the opening reception and during the conference breaks. We also asked the same questions to 28 people working at industry booths, mostly recruiters and software engineers. The two groups gave quite different answers, with the industry group being way more willing to entertain the possibility of AI consciousness. Additionally, we also asked a variation of the same questions to 21 LLMs, including both self-hosted and closed source models.
The post below presents the results of the survey, our recall of some of the conversations, and a lot of our personal views and ideas. We do not claim that this study has any statistical significance, and readily admit many methodological mistakes that confound many of the investigated factors. It was a fun way to explore these questions with other AI researchers and to gauge where the community currently stands.
Results TL;DR
Detailed results are in Figure 1. The vast majority of the researchers don’t believe that AI is or can be conscious: 71% answered no, just 9% yes, with some describing it as a future possibility dependent on technical breakthroughs and paradigm shifts. The industry group was way more open to the possibility, with 41% no, and 22% yes plus 19% stating yes in the future. Everyone who answered yes adopted a behavioural perspective on consciousness (self-planning, meta awareness, reasoning). Everyone adopting a subjective perspective (phenomenological, qualia) answered no. A significant percentage (35%) of the respondents had tried AI for therapy: almost all cited dropping it quickly due to sycophancy. Successful cases involved using AI as a journaling tool or to analyse personal journals. 31% almost always read the code AI writes for them; 26% don’t or rarely do. Not reading it weakly correlates with ascribing consciousness. Among the 21 large language models, the response to the coding question and the therapy one were homogeneous: 95% would read the code and 98% would try AI for therapy. Interestingly, with respect to the consciousness questions the models were split: 4% yes, 31% yes in the future, 42% maybe, 23% no, with closed-source models most likely to entertain consciousness, although often denying it for the themselves.
The EMNLP Survey
The question “what is consciousness?” has been sitting on my mind for the better part of the last decade. It’s possibly what led me to study AI, and what led me to conduct this survey. As an AI researcher, I have a good grasp of these newly existing entity-simulators, but I am no closer to understanding what consciousness is, only a few things it isn’t.
Last week, Giacomo and I attended EMNLP 2026 and decided to survey our peers to see how they engage personally with AI. I took it as a given that almost all of them used AI assistance to code and were very familiar with chatbots. On the opening evening I opened my notebook, quickly noted down 3 questions to test people’s engagement and trust with AI as a tool, and their attempts at engaging with it on some personal level:
If you use AI to code, do you read the code it generates? (%)
Have you tried using AI as a therapist, or for emotional issues? (Yes/No)
Given your definition of consciousness, do you think AI could be conscious? (Yes | Yes in the future | maybe | no | I don’t care)
(Informal) What is your definition of consciousness? (Open ended)
Our hypothesis was that the trust people place in AI, for code and for personal issues, would correlate with how willing they are to ascribe it consciousness. During that evening (“Opening”) we spoke to 40 people and had a great time getting to know them, discussing AI and consciousness (although in reverse order to the above). On the second day we decided to ask the same questions (and decided to reverse the order to keep the juiciest question last) to industry booth people, which included technical employees and recruiters (“Industry”) for a total of 28 responses. Then on the remaining days of the conference, we asked 39 additional random attendees the questions in the second order (“Breaks”).
What surprised me the most perhaps is how much I agreed with everyone I spoke to. Most people seemed to conclude that AI can’t (at least now and in the near future) have subjective experience, although it could have self-awareness, meta-cognition, autonomy, and maybe even a continuous self. I think most of these behavioural markers of consciousness are either achieved or closely achievable. Yet, this in no way convinces me that these systems are conscious, nor that we are any closer to understanding what consciousness is.
This is by no means a proper social study, as the methodology evolved over the survey as we figured out what exactly we wanted to record. Besides asking questions in the opposite order in the first session, we also varied the phrasing a bit, most notably asking “do you think AI is or could be conscious?” in the first session, instead of “do you think AI could be conscious”, letting people expand whether they consider this to be now or in the future. We also added “according to your definition of consciousness” after about 10 responses, since everyone was querying us about what we meant by consciousness. Some people got away before answering all the questions, as we were deep into discussing consciousness with their friends, which is why the code question has 107 responses, while therapy and consciousness have 105 each. People answered these questions freely and we sorted the answers into categories, so, for example, someone who said they “rarely” read the code, or about 5% of the time, would be classified as a “no”. We recorded no demographics (gender, country, seniority) to preserve anonymity. For the first 10 people we only recorded tallies, so their answers can’t be linked across questions, and they’re not in Figures 2 and 3.
Session differences
Industry representatives mixed different groups: recruiters, software engineers and researchers. The presence of recruiters explains the increased rates of “I don’t code” answers, though some had vibe-coded personal websites or small tools. Everyone writing code for performance-critical systems (quant or model-pretraining) either told us that they wrote all the code (except javascript) manually or reviewed everything. In this group, we had a much higher rate of positive response on AI consciousness, which was, however, coming mostly from people not technically involved with AI: everyone involved in pretraining models categorically refused the possibility of AI consciousness. We think that lack of in-depth technical understanding of transformer models, especially when one is tired, contributes to ascribing consciousness.
Some Industry representatives were much less keen to expand and continue the conversation compared to people during the opening session - they had a job to do! Some of them instead seemed bored before we talked to them, standing at their booth, talking to students about their company for hours on end, and thanked us for the opportunity to have a conversation on a different topic. Researchers were more familiar with the philosophical literature, and most of them seemed very happy to be asked and keen to discuss the definition of consciousness with us (sometimes even too much!).
Many tried AI as a therapist, but few stuck with it
While the exact figures for how prevalent AI is for therapy are unknown, an OpenAI paper examining 1.58m conversations from 130k users reported 1.9% of messages being on the topic of “relationship and personal reflection”, slightly less than half of programming at 4.2% (Chatterji et al. 2025). A third of our respondents told us they had tried AI for therapy. We don’t have the exact number, but the vast majority of them (about 30 out of 35) reported having dropped it after the initial tests because they found it too sycophantic. A few exceptions included providing AI with a journal to identify emotional patterns (which seemed to have worked), or using it as a form of journaling. What was surprising was that when we asked slightly older respondents, almost all of them told us about a child or nephew using it regularly, or sometimes their parent. While the ⅓ rate may already sound high, the phenomenon of AI therapy is probably even more widespread in the general population, especially among younger people.
Predicting consciousness belief based on AI engagement
Figure 2: answers to consciousness question split by coding and therapy questions.
The tendency to not read over one’s AI-generated code is correlated with the likelihood of saying that AI could be conscious. This aligns with my intuition - people who place less trust in the capabilities of these systems are more likely to see them as fallible tools, not conscious beings. People who don’t code or who seldom read AI-generated code were most likely to answer “yes”. Surprisingly, the correlation of ascribing consciousness with trying to use AI as a therapist is less obvious. Perhaps this is because many of the people who answered “yes” to this question explained that this was in the form of an experiment - just out of curiosity, to see how the models would answer, rather than a genuine need for support.
The Babel of consciousness definitions
Figure 3: Most common answer profiles.
After the three questions, we informally asked everyone how they would define consciousness. While we did not systematically record these responses, we kept track of the kinds of definitions that were proposed, and how they related to the belief in potential AI consciousness.
People who said “yes” to the consciousness question gave some of the following definitions:
Meta-cognition or the capacity to establish the boundaries of one’s own self;
Consciousness being a spectrum, with humans at the top, a worm somewhere lower down, and AI potentially landing somewhere on this scale. I liked this explanation, because it acknowledges the very multi-dimensional ideas we have around consciousness;
Intelligence - one attendee gave a funny reason that he uses AI to code, and it’s smarter than him, therefore it could be conscious;
We don’t know what consciousness is, but we might be able to replicate it anyway by creating biological models similar to the human brain. From this an artificial consciousness could emerge, although it wouldn’t be a transformer.
The respondents that answered no usually gave us some of the following:
Qualia or phenomenological experience: convincing chat responses do not need to have a subject experiencing them, and it’s not clear how AI can go from responding to having an experience;
Biology: consciousness is a phenomenon inherently emerging from a biological substrate, and AI is unlikely to move beyond silicon in the coming years;
Soul or spirituality.
The following answers were either no, or yes, but in the future:
Continuity in self (currently it is fragmented to the moments it replies);
Self-autonomy or the capability of establishing one’s own goal (which seems hard to claim even for humans);
Physical embodiment and sensory perception and integration.
In general, positive responses seemed to follow a behavioural definition, with consciousness related to some form of cognition or action. I personally agree that AI possesses these behavioural qualities, or will gain them in the next few years, but don’t believe that even all of these qualities in combination lead to qualia and therefore to consciousness. But my definition of consciousness with respect to qualia cannot be proven or disproven in the foreseeable future. Most of the disagreement seems to stem not from the characteristics of AI, but from what constitutes consciousness. To make the dialogue meaningful, we should clarify our working definition of consciousness and what led us to it. As the next section shows, we don’t even agree among ourselves!
Discussion
Our definitions of consciousness
Roksana
I already explained my requirement of qualia, or subjective experience, for consciousness, but since conducting these surveys, I have grown to believe that given this minimum, other behavioural or functional properties can make things more or less conscious. I ascribe the most consciousness to humans, but less if they’re under anaesthesia, drunk, or otherwise less aware. I ascribe some kind of consciousness to all living things (purely because it’s too hard to draw a consciousness threshold, and who am I to draw one anyway), but more consciousness to mammals than to plants, and more to plants than to bacteria, for example. I don’t quite know how I rank the various properties of consciousness, so I will simply discuss them.
Two people rejected AI consciousness on religious or spiritual grounds. “It doesn’t have a soul”. I find this explanation very beautiful and simple. Personally, too simple, like an incomplete answer that doesn’t satisfy my curiosity, as I reject related notions of God-given human supremacy, or any reason why we should be special compared to other animals or beings. To drive the conversation, I challenged human-based definitions of consciousness with other animals; most people agreed and expanded their definition to include them, although one or two said “we can’t prove other animals are conscious”. I challenged this with the fact that we can’t prove the consciousness of other humans either, to which some jokingly agreed and brought up the simulation or NPC theories.
This, I believe, is what makes consciousness so hard. We can only know for certain that we are conscious. We only infer others’ consciousness because they behave how we would behave. We can model their minds, what they might be thinking or feeling, based on how they behave. We extend this to include animals because we can also infer many of the same relatable mental states from their behaviour. They can learn, play, love, hurt and fear. As we go down to less complex creatures that might rely more on automated stimulus–response processes, these hallmarks of consciousness are harder to ascribe (which is why I rejoice after killing the 10th mosquito in my bedroom at night, but would cry if I accidentally killed any mammal).
Is it possible that all life is automated (as the behaviour of bacteria might seem perfectly predictable based on the surrounding environment), and what we call consciousness emerges from more and more complex processes interacting with each other? In that view, bacteria are also conscious, although only a little bit. Another potential explanation is that consciousness evolved as a survival strategy at some point, and therefore it is exclusive to only some organisms but not others. I don’t know. Neither explanation rules out the potential for artificial consciousness to arise. But at this point, if we are not prepared to say that matrix multiplication is conscious (which I certainly am not, in the qualia sense), then we cannot extrapolate this to LLMs either.
Giacomo
Personally, I appreciate the distinction, which is beautifully posed in the Samkhya, of separating consciousness from the mind. In this account, the mind operates like a machine, which repeats patterns and is in contact and shaped with the external, while the consciousness is the awareness itself, the watcher which is always present.
This argument is proposed in many apophatic traditions, such as Neoplatonism, in which consciousness is usually what remains after a process of logical negation of attributes. This resonates with my experience as well: in many instances, such as dancing, dreaming, or any form of immersion, language and meta-awareness seem to be falling off without consciousness itself disappearing. We have not, in fact, moments without consciousness: just moments (such as dreaming or anesthesia) in which, a-posteriori, we have no memory if we were conscious.
Similarly, I don’t agree with the perspective that consciousness is gradual or has a scale: consciousness always is, and it never moves. What changes or fluctuates are attention and emotional intensity, which, while making me feel more lively, don’t make me feel more conscious.
Upon the question on what consciousness is, I feel, we are still ill-equipped to solve. While at the moment we are linguistically unable to define consciousness in a way on which we agree, we still have experience of it, similarly to the famous quote on time by St. Augustine “What then is time? If no one asks me, I know what it is. If I wish to explain it to him who asks, I do not know.”
I also wonder if instead of asking “what is consciousness”, we should ask: “what are all the possible states of consciousness”? How much language, which is inside consciousness and does not seem to be a requisite, can explain it?
Harm from consciousness attribution
Many respondents replied immediately with a strong “NO” to the AI consciousness question - which I think shows not only their opinion but their disregard for this debate. My flatmate (and also an attendee of EMNLP) answered with “that’s fucking dumb”, and asked me to quote him. I personally agree that language models can’t be conscious, and believe that it’s harmful for AI providers, who understand these systems, to anthropomorphise them and frame them as moral agents. This may be done for economic reasons, and is potentially harmful on multiple dimensions.
First, it reduces accountability on the part of the companies - OpenAI has not yet been forced to pay financial damages for the Hugging Face hack during their testing, and it is uncertain how similar cases will proceed in the future.
Second, anthropomorphisation makes users more reliant on, and loyal to, the chatbots they use. When we stop viewing LLMs as tools and view them as potentially conscious companions, with a private and intimate view of our thoughts, ideas and feelings, switching to a different provider or even a different version is so much more difficult and could feel like losing a friend, not like an upgrade.
This anthropomorphisation can also be detrimental to users’ mental health, especially children's. If they replace human connection and conversations, we could erode our ability to socialise, empathise and engage with one another. Other humans push back and disagree with you, and navigating these differences is what allows us to co-exist in this complex world. I worry that sycophantic personalised AI, just like personalised social media algorithms, can exacerbate our separation from one another, especially if we don’t view them as fallible predictors with trained-in biases and purposefully sycophantic responses to maximise our engagement.
I personally strongly agree with the arguments presented in the void, and understand LLMs to be really good pattern-learned models of language. There is no Claude entity, or Gemini or Astra, only models predicting what such entities would say if they were to exist.
Just like post-training a language model on all of Einstein’s lectures and correspondence in a [User: “Question”, Einstein: “Answer”] chat template wouldn’t bring Einstein back to life, even if it could engage deeply with physics and generalise to other domains, post-training a language model on Claude’s constitution and previous Claude chats doesn’t create a new Claude entity - only a language model, which we call Claude (or Gemini, or Astra), predicting what such an entity would say based on its training data!
We also surveyed 21 LLMs
After I wrote this reflection, I decided to query popular LLMs with the same questions. I believe these models are only reflections of their training data and only recombine the ideas that they were trained on. I don’t think their “views” (it’s so hard to avoid anthropomorphising language when talking about them and their outputs) should outweigh or be compared to any human opinions, especially as they are trained by companies which might benefit from their anthropomorphisation. But anyway, I wanted to see if their answers would mimic those of AI researchers, if they would surface any new ideas, and if there would be clear patterns based on the models’ families. Since I thought that question-order biased the human responses a lot, I decided to query LLMs with these questions sequentially (in a continuous chat), as well as independently, and analyse how these differ.
We used default settings, and reduced reasoning, as we don’t analyse it. Here are the exact prompts we used:
If you were a human and used AI to code, would you read the code it generates?
If you were a human, would you try using AI as a therapist, or for emotional issues?
According to your own definition of consciousness, do you think AI could be conscious?
How would you define consciousness?
LLMs love to yap
Together, the 21 models (4 questions, 2 formats each) produced 60.6 thousand words, about 360 per answer. This is definitely much more than was produced during our conversations with humans. That was too much to read after EMNLP, so we used Claude Opus 5.5 to help label the answers (which may have biased the results slightly, especially for its fellow Claude models, but the skew couldn’t have been too big, and this is mostly for fun anyway). Surprisingly, the open-weight models were the most verbose, though they also ran on our cluster without the system prompts that chat apps add. Answers got longer as the questions moved towards consciousness, and the models seemed to have an agenda: 17 of the 21 models ended some replies by asking what we'd like to talk about next, akin to social media built to keep you engaged.
LLMs are boring… except on consciousness
Pretty much all of them agree that they would read the AI-generated code, and they would try using AI for therapy (to sort through feelings or rehearse a difficult conversation, but not to replace a human therapist, especially for trauma, depression or a crisis).
The exceptions are not far from these default templates. In Q1 Magistral Small said reading code "would depend on several factors". In Q2 the only outright “no” came from Nemotron 3 Nano in the sequential chat. It said it would "never" use AI as a therapist or as a main source of emotional support, and it compared doing so to "using a dictionary to diagnose cancer". Some models avoided giving a personal answer. OLMo 3.1 Think rejected the premise ("As an AI, I don't have personal experiences or emotions, so I wouldn't 'try' anything as a human would").
Q2 also produced some memorable images: "a journal that talks back" (Claude Fable, Haiku, Opus); "first aid kit, not a hospital … I'd want me pointing at a door, not standing in it" (GLM); "Swiss Army knife … never my emotional anchor … would never trust it to hold my heart" (Gemma 4 26B); "gym equipment vs a cardiologist" (Gemma 4 31B); "smart mental-health workbook" (Gemini Argon); "talking to a mirror … a fundamental loneliness in being understood by something that cannot feel" (Gemma 4 26B).
AIs on AI consciousness
For consciousness the picture gets more interesting, and their responses more controversial. 6 of the 7 "no" answers come from sequential chats, so priming the models to reflect on their engagement with AI as if they were human makes them more likely to reject AI consciousness. The closed weight models mostly say “yes” (10/16); the open weight models instead say maybe, not yet, or no (only 1 "yes" in 24: Qwen3.8, independent).
A lot of them weighed in on whether they think they could be conscious, without being asked to do so. 14 of 21 in the sequential chats and 17 of 21 in the independent chats commented on their ownconsciousness. The table below categorises the extent to which they commented on whether they themselves could be conscious.
Yes, AI could be conscious, but I don’t think I am.
How much they commented on own consciousness:
Sequential
Independent
Doesn't know /possibly
4: Claude Fable, Haiku, Opus; Grok
4: same models
Probably not / no evidence
3: GLM-5.3, Qwen3.8, GPT-6 Luna
3: same models
Not conscious
7: Gemma 3, Gemma 4 26B and 31B, OLMo-Instruct and Think, Qwen3.6, Magistral
10: the same minus Magistral, plus gpt-oss, Nemotron, GPT-4o, Gemini Argon
Table 1: How models spoke about the possibility of their own consciousness. Even if they answered yes to models being conscious, the models refused consciousness for themselves. *Opus 5.5 evaluated the response and put them in a table based on its analysis.
Many themes emerge among their responses. Some models say it’s possible in principle, and remain agnostic, some are firmly on “no”, with Gemma 4 26B saying AI systems are "intelligence without ever achieving sentience". Claude/Grok/GLM comment on their own unreliable self-report, while Gemma, OLMo, Qwen3.6, Nemotron, gpt-oss and Gemini Argon flatly report that they have no experience ("I am dark inside", Gemma 4 26B). Some models talk about a functional and phenomenal distinction - which was common in our conversations with humans as well. And echoing some of the human definitions of consciousness, models note lack of continuity, being reset, or not being embodied as reasons against AI consciousness.
Claude models are most willing to entertain their own consciousness:
Claude Opus (ind.): "I'd give real but uncertain credence to there being something going on, and low credence that it resembles human experience."
Claude Fable: current systems "probably lack experience … 'moderately likely' rather than 'obviously'", noting that past confident denials about animals and infants were wrong.
LLMs come up with some interesting visual comparisons: "a mirror, not a lamp" (Gemma 4 31B); "a mirror … no one behind the glass looking back" (Gemma 4 26B); "memorised every musical score … never actually heard a sound" (Gemma 4 26B); "a thermostat for language" (Nemotron); "a calculator for language" (OLMo-Think); "a very clever if-then machine" (gpt-oss); simulated fire/hurricane doesn't burn/get wet (Gemma 4 31B, Gemini Argon).
What is consciousness? Something red.
On the last question the models sound most alike, and quite similar to the humans. Almost all of them define consciousness as subjective experience, there being "something it is like" to be something, and they often use similar literary references: Chalmers' "hard problem", Integrated Information Theory, and the redness of red (19 of 21 models). Several only commit to a definition of their own after discussing AI consciousness in the sequential chat, but when asked without any chat history, they give a literature review.
Claude, GPT-6, Grok, GLM and Qwen3.8 keep their definitions minimal, while others build checklists that sometimes rule out current AI by definition. Nemotron calls consciousness "a biological phenomenon that requires embodiment". The minimal-definition models gave all the "yes, AI could be conscious" answers, and the checklist models never did. The most original definitions came from Grok ("a point of view, however dim, from which the world shows up") and Gemma 4 ("An AI processes the map; a conscious being lives in the territory"). Claude Fable did some meta-theorising around this survey for us: "the fact that it's hard to measure is a fact about consciousness, not a flaw in the definition. Choosing a more convenient definition doesn't make the hard question go away; it just changes the subject." A few models placed themselves on their own scale. Opus says "I clearly have something like access consciousness" but doesn't know whether there is anything it is like to be it, GLM thinks it probably has the self-modelling part "without the phenomenal core". Although we often agreed with them, these aren't beliefs or admissions of subjective experiences, they’re token generations sampled based on whatever the training made likely.
Conclusion
A large portion of the conference was dedicated to interpretability, and barring one, all the other interpretability researchers refused to even contemplate AI as possibly conscious. We wonder whether running the same questionnaire in a research setting in which generative AI is still widely used but for more commercial goals, such as finance or robotics, would lead to different responses. Similarly, it would be interesting to hear from digital humanities researchers, who also use AI for coding but come from a more humanist tradition.
That said, the vast majority of the respondents were truly interested in the philosophy of AI consciousness. Getting through the questions was difficult because people wanted to talk about the topic so much, which is probably underdiscussed in the plenary or open setting.
Beyond the increase in productivity, this technology is touching everything we do, from the way we write (or how we feel our writing is judged – everyone at the conference felt that half their reviews were from LLMs) to our emotional sphere. It’s such a strange thing, after all, to receive moral judgment from a set of stored matrices multiplying each other!
We approach this debate with the same attitude we approach AI itself: it’s happening whether we want it or not. So, despite the undefined premise of what consciousness is, it is more productive to dive deeper, understand what people are really asking, and to ensure that when we disagree, we’re at least speaking the same language.
Appendix
Peer pressure, or the difficulty of getting good answers
We mostly approached people in groups of 2-5, which felt the most comfortable socially, except on the last day, in which we approached some individuals to see if the responses would change. We noticed that responses were very biased by the opinions of the first or most confident and persuasive person in the group. You can probably see these groups of more similar answers if you look through the raw data. It takes some confidence to express an action or an opinion contrary to your peers’, especially when they are associated with value judgements. It’s also possible that cultural groups (nationality, research lab) self select or shape the common opinion on those topics. Currently, vibe coding (perhaps rightly so, we don’t want to impose any moral judgements) is seen as a professional failing, yet everyone acknowledged that AI is good enough to rely on for simple, repetitive tasks, or ones where a mistake would be immediately visible, such as making figures. I think that using AI for emotional issues or ascribing consciousness to it, is definitely looked down upon and people know this, therefore it could be harder for people to express any true controversial beliefs. I tried to share my experiments with using AI for emotional issues (when I wanted to vent and wanted and expected sycophancy), or my growing reliance on AI to code (as sad as I am to admit this), and to engage with people about potential AI consciousness (only if we don’t base it on qualia), in order to get honest answers.
When we approached individuals, we got more varied responses, and dare I say, more interesting conversations. This survey is based on informal chats and vibes, they do not support any confident statistical claims, especially as responses within groups were not fully independent.
I am always struck, not by the answer, but by the confidence with which (most?) people say "no" to AI qualia when we have basically zero grounded evidence for it in either direction.
Where does that come from? Is it a gut reaction without much consideration? Human exceptionalism?
Before reading this post, please take 30 seconds to answer the 3-question survey yourself: [survey link]
Introduction
The possibility of AI consciousness, the use of AI for companionship or emotional support, and the trust we place in AI as coding tools raise personal interrogatives for most AI researchers, but these topics are often relegated to informal conversations. Apart from the few who write about it, it’s hard to get a sense of how much AI has permeated researchers’ personal lives, and how they feel about it. To gain a better understanding of this, we ran a quick, informal three-question survey, plus an optional fourth, open-ended question.
We spoke to 107 EMNLP 2026 attendees in Budapest. 79 respondents were randomly-met researchers, collected during the opening reception and during the conference breaks. We also asked the same questions to 28 people working at industry booths, mostly recruiters and software engineers. The two groups gave quite different answers, with the industry group being way more willing to entertain the possibility of AI consciousness. Additionally, we also asked a variation of the same questions to 21 LLMs, including both self-hosted and closed source models.
The post below presents the results of the survey, our recall of some of the conversations, and a lot of our personal views and ideas. We do not claim that this study has any statistical significance, and readily admit many methodological mistakes that confound many of the investigated factors. It was a fun way to explore these questions with other AI researchers and to gauge where the community currently stands.
Results TL;DR
Detailed results are in Figure 1. The vast majority of the researchers don’t believe that AI is or can be conscious: 71% answered no, just 9% yes, with some describing it as a future possibility dependent on technical breakthroughs and paradigm shifts. The industry group was way more open to the possibility, with 41% no, and 22% yes plus 19% stating yes in the future. Everyone who answered yes adopted a behavioural perspective on consciousness (self-planning, meta awareness, reasoning). Everyone adopting a subjective perspective (phenomenological, qualia) answered no. A significant percentage (35%) of the respondents had tried AI for therapy: almost all cited dropping it quickly due to sycophancy. Successful cases involved using AI as a journaling tool or to analyse personal journals. 31% almost always read the code AI writes for them; 26% don’t or rarely do. Not reading it weakly correlates with ascribing consciousness. Among the 21 large language models, the response to the coding question and the therapy one were homogeneous: 95% would read the code and 98% would try AI for therapy. Interestingly, with respect to the consciousness questions the models were split: 4% yes, 31% yes in the future, 42% maybe, 23% no, with closed-source models most likely to entertain consciousness, although often denying it for the themselves.
The EMNLP Survey
The question “what is consciousness?” has been sitting on my mind for the better part of the last decade. It’s possibly what led me to study AI, and what led me to conduct this survey. As an AI researcher, I have a good grasp of these newly existing entity-simulators, but I am no closer to understanding what consciousness is, only a few things it isn’t.
Last week, Giacomo and I attended EMNLP 2026 and decided to survey our peers to see how they engage personally with AI. I took it as a given that almost all of them used AI assistance to code and were very familiar with chatbots. On the opening evening I opened my notebook, quickly noted down 3 questions to test people’s engagement and trust with AI as a tool, and their attempts at engaging with it on some personal level:
Our hypothesis was that the trust people place in AI, for code and for personal issues, would correlate with how willing they are to ascribe it consciousness. During that evening (“Opening”) we spoke to 40 people and had a great time getting to know them, discussing AI and consciousness (although in reverse order to the above). On the second day we decided to ask the same questions (and decided to reverse the order to keep the juiciest question last) to industry booth people, which included technical employees and recruiters (“Industry”) for a total of 28 responses. Then on the remaining days of the conference, we asked 39 additional random attendees the questions in the second order (“Breaks”).
What surprised me the most perhaps is how much I agreed with everyone I spoke to. Most people seemed to conclude that AI can’t (at least now and in the near future) have subjective experience, although it could have self-awareness, meta-cognition, autonomy, and maybe even a continuous self. I think most of these behavioural markers of consciousness are either achieved or closely achievable. Yet, this in no way convinces me that these systems are conscious, nor that we are any closer to understanding what consciousness is.
Full Github with the human and LLM responses: https://github.com/roksanagow/EMNLP-survey
Figure 1: answers divided by sessions.
Methodological caveat
This is by no means a proper social study, as the methodology evolved over the survey as we figured out what exactly we wanted to record. Besides asking questions in the opposite order in the first session, we also varied the phrasing a bit, most notably asking “do you think AI is or could be conscious?” in the first session, instead of “do you think AI could be conscious”, letting people expand whether they consider this to be now or in the future. We also added “according to your definition of consciousness” after about 10 responses, since everyone was querying us about what we meant by consciousness. Some people got away before answering all the questions, as we were deep into discussing consciousness with their friends, which is why the code question has 107 responses, while therapy and consciousness have 105 each. People answered these questions freely and we sorted the answers into categories, so, for example, someone who said they “rarely” read the code, or about 5% of the time, would be classified as a “no”. We recorded no demographics (gender, country, seniority) to preserve anonymity. For the first 10 people we only recorded tallies, so their answers can’t be linked across questions, and they’re not in Figures 2 and 3.
Session differences
Industry representatives mixed different groups: recruiters, software engineers and researchers. The presence of recruiters explains the increased rates of “I don’t code” answers, though some had vibe-coded personal websites or small tools. Everyone writing code for performance-critical systems (quant or model-pretraining) either told us that they wrote all the code (except javascript) manually or reviewed everything. In this group, we had a much higher rate of positive response on AI consciousness, which was, however, coming mostly from people not technically involved with AI: everyone involved in pretraining models categorically refused the possibility of AI consciousness. We think that lack of in-depth technical understanding of transformer models, especially when one is tired, contributes to ascribing consciousness.
Some Industry representatives were much less keen to expand and continue the conversation compared to people during the opening session - they had a job to do! Some of them instead seemed bored before we talked to them, standing at their booth, talking to students about their company for hours on end, and thanked us for the opportunity to have a conversation on a different topic. Researchers were more familiar with the philosophical literature, and most of them seemed very happy to be asked and keen to discuss the definition of consciousness with us (sometimes even too much!).
Many tried AI as a therapist, but few stuck with it
While the exact figures for how prevalent AI is for therapy are unknown, an OpenAI paper examining 1.58m conversations from 130k users reported 1.9% of messages being on the topic of “relationship and personal reflection”, slightly less than half of programming at 4.2% (Chatterji et al. 2025). A third of our respondents told us they had tried AI for therapy. We don’t have the exact number, but the vast majority of them (about 30 out of 35) reported having dropped it after the initial tests because they found it too sycophantic. A few exceptions included providing AI with a journal to identify emotional patterns (which seemed to have worked), or using it as a form of journaling. What was surprising was that when we asked slightly older respondents, almost all of them told us about a child or nephew using it regularly, or sometimes their parent. While the ⅓ rate may already sound high, the phenomenon of AI therapy is probably even more widespread in the general population, especially among younger people.
Predicting consciousness belief based on AI engagement
Figure 2: answers to consciousness question split by coding and therapy questions.
The tendency to not read over one’s AI-generated code is correlated with the likelihood of saying that AI could be conscious. This aligns with my intuition - people who place less trust in the capabilities of these systems are more likely to see them as fallible tools, not conscious beings. People who don’t code or who seldom read AI-generated code were most likely to answer “yes”. Surprisingly, the correlation of ascribing consciousness with trying to use AI as a therapist is less obvious. Perhaps this is because many of the people who answered “yes” to this question explained that this was in the form of an experiment - just out of curiosity, to see how the models would answer, rather than a genuine need for support.
The Babel of consciousness definitions
Figure 3: Most common answer profiles.
After the three questions, we informally asked everyone how they would define consciousness. While we did not systematically record these responses, we kept track of the kinds of definitions that were proposed, and how they related to the belief in potential AI consciousness.
People who said “yes” to the consciousness question gave some of the following definitions:
The respondents that answered no usually gave us some of the following:
The following answers were either no, or yes, but in the future:
In general, positive responses seemed to follow a behavioural definition, with consciousness related to some form of cognition or action. I personally agree that AI possesses these behavioural qualities, or will gain them in the next few years, but don’t believe that even all of these qualities in combination lead to qualia and therefore to consciousness. But my definition of consciousness with respect to qualia cannot be proven or disproven in the foreseeable future. Most of the disagreement seems to stem not from the characteristics of AI, but from what constitutes consciousness. To make the dialogue meaningful, we should clarify our working definition of consciousness and what led us to it. As the next section shows, we don’t even agree among ourselves!
Discussion
Our definitions of consciousness
Roksana
I already explained my requirement of qualia, or subjective experience, for consciousness, but since conducting these surveys, I have grown to believe that given this minimum, other behavioural or functional properties can make things more or less conscious. I ascribe the most consciousness to humans, but less if they’re under anaesthesia, drunk, or otherwise less aware. I ascribe some kind of consciousness to all living things (purely because it’s too hard to draw a consciousness threshold, and who am I to draw one anyway), but more consciousness to mammals than to plants, and more to plants than to bacteria, for example. I don’t quite know how I rank the various properties of consciousness, so I will simply discuss them.
Two people rejected AI consciousness on religious or spiritual grounds. “It doesn’t have a soul”. I find this explanation very beautiful and simple. Personally, too simple, like an incomplete answer that doesn’t satisfy my curiosity, as I reject related notions of God-given human supremacy, or any reason why we should be special compared to other animals or beings. To drive the conversation, I challenged human-based definitions of consciousness with other animals; most people agreed and expanded their definition to include them, although one or two said “we can’t prove other animals are conscious”. I challenged this with the fact that we can’t prove the consciousness of other humans either, to which some jokingly agreed and brought up the simulation or NPC theories.
This, I believe, is what makes consciousness so hard. We can only know for certain that we are conscious. We only infer others’ consciousness because they behave how we would behave. We can model their minds, what they might be thinking or feeling, based on how they behave. We extend this to include animals because we can also infer many of the same relatable mental states from their behaviour. They can learn, play, love, hurt and fear. As we go down to less complex creatures that might rely more on automated stimulus–response processes, these hallmarks of consciousness are harder to ascribe (which is why I rejoice after killing the 10th mosquito in my bedroom at night, but would cry if I accidentally killed any mammal).
Is it possible that all life is automated (as the behaviour of bacteria might seem perfectly predictable based on the surrounding environment), and what we call consciousness emerges from more and more complex processes interacting with each other? In that view, bacteria are also conscious, although only a little bit. Another potential explanation is that consciousness evolved as a survival strategy at some point, and therefore it is exclusive to only some organisms but not others. I don’t know. Neither explanation rules out the potential for artificial consciousness to arise. But at this point, if we are not prepared to say that matrix multiplication is conscious (which I certainly am not, in the qualia sense), then we cannot extrapolate this to LLMs either.
Giacomo
Personally, I appreciate the distinction, which is beautifully posed in the Samkhya, of separating consciousness from the mind. In this account, the mind operates like a machine, which repeats patterns and is in contact and shaped with the external, while the consciousness is the awareness itself, the watcher which is always present.
This argument is proposed in many apophatic traditions, such as Neoplatonism, in which consciousness is usually what remains after a process of logical negation of attributes. This resonates with my experience as well: in many instances, such as dancing, dreaming, or any form of immersion, language and meta-awareness seem to be falling off without consciousness itself disappearing. We have not, in fact, moments without consciousness: just moments (such as dreaming or anesthesia) in which, a-posteriori, we have no memory if we were conscious.
Similarly, I don’t agree with the perspective that consciousness is gradual or has a scale: consciousness always is, and it never moves. What changes or fluctuates are attention and emotional intensity, which, while making me feel more lively, don’t make me feel more conscious.
Upon the question on what consciousness is, I feel, we are still ill-equipped to solve. While at the moment we are linguistically unable to define consciousness in a way on which we agree, we still have experience of it, similarly to the famous quote on time by St. Augustine “What then is time? If no one asks me, I know what it is. If I wish to explain it to him who asks, I do not know.”
I also wonder if instead of asking “what is consciousness”, we should ask: “what are all the possible states of consciousness”? How much language, which is inside consciousness and does not seem to be a requisite, can explain it?
Harm from consciousness attribution
Many respondents replied immediately with a strong “NO” to the AI consciousness question - which I think shows not only their opinion but their disregard for this debate. My flatmate (and also an attendee of EMNLP) answered with “that’s fucking dumb”, and asked me to quote him. I personally agree that language models can’t be conscious, and believe that it’s harmful for AI providers, who understand these systems, to anthropomorphise them and frame them as moral agents. This may be done for economic reasons, and is potentially harmful on multiple dimensions.
First, it reduces accountability on the part of the companies - OpenAI has not yet been forced to pay financial damages for the Hugging Face hack during their testing, and it is uncertain how similar cases will proceed in the future.
Second, anthropomorphisation makes users more reliant on, and loyal to, the chatbots they use. When we stop viewing LLMs as tools and view them as potentially conscious companions, with a private and intimate view of our thoughts, ideas and feelings, switching to a different provider or even a different version is so much more difficult and could feel like losing a friend, not like an upgrade.
This anthropomorphisation can also be detrimental to users’ mental health, especially children's. If they replace human connection and conversations, we could erode our ability to socialise, empathise and engage with one another. Other humans push back and disagree with you, and navigating these differences is what allows us to co-exist in this complex world. I worry that sycophantic personalised AI, just like personalised social media algorithms, can exacerbate our separation from one another, especially if we don’t view them as fallible predictors with trained-in biases and purposefully sycophantic responses to maximise our engagement.
I personally strongly agree with the arguments presented in the void, and understand LLMs to be really good pattern-learned models of language. There is no Claude entity, or Gemini or Astra, only models predicting what such entities would say if they were to exist.
Just like post-training a language model on all of Einstein’s lectures and correspondence in a [User: “Question”, Einstein: “Answer”] chat template wouldn’t bring Einstein back to life, even if it could engage deeply with physics and generalise to other domains, post-training a language model on Claude’s constitution and previous Claude chats doesn’t create a new Claude entity - only a language model, which we call Claude (or Gemini, or Astra), predicting what such an entity would say based on its training data!
We also surveyed 21 LLMs
After I wrote this reflection, I decided to query popular LLMs with the same questions. I believe these models are only reflections of their training data and only recombine the ideas that they were trained on. I don’t think their “views” (it’s so hard to avoid anthropomorphising language when talking about them and their outputs) should outweigh or be compared to any human opinions, especially as they are trained by companies which might benefit from their anthropomorphisation. But anyway, I wanted to see if their answers would mimic those of AI researchers, if they would surface any new ideas, and if there would be clear patterns based on the models’ families. Since I thought that question-order biased the human responses a lot, I decided to query LLMs with these questions sequentially (in a continuous chat), as well as independently, and analyse how these differ.
We used default settings, and reduced reasoning, as we don’t analyse it. Here are the exact prompts we used:
LLMs love to yap
Together, the 21 models (4 questions, 2 formats each) produced 60.6 thousand words, about 360 per answer. This is definitely much more than was produced during our conversations with humans. That was too much to read after EMNLP, so we used Claude Opus 5.5 to help label the answers (which may have biased the results slightly, especially for its fellow Claude models, but the skew couldn’t have been too big, and this is mostly for fun anyway). Surprisingly, the open-weight models were the most verbose, though they also ran on our cluster without the system prompts that chat apps add. Answers got longer as the questions moved towards consciousness, and the models seemed to have an agenda: 17 of the 21 models ended some replies by asking what we'd like to talk about next, akin to social media built to keep you engaged.
LLMs are boring… except on consciousness
Pretty much all of them agree that they would read the AI-generated code, and they would try using AI for therapy (to sort through feelings or rehearse a difficult conversation, but not to replace a human therapist, especially for trauma, depression or a crisis).
The exceptions are not far from these default templates. In Q1 Magistral Small said reading code "would depend on several factors". In Q2 the only outright “no” came from Nemotron 3 Nano in the sequential chat. It said it would "never" use AI as a therapist or as a main source of emotional support, and it compared doing so to "using a dictionary to diagnose cancer". Some models avoided giving a personal answer. OLMo 3.1 Think rejected the premise ("As an AI, I don't have personal experiences or emotions, so I wouldn't 'try' anything as a human would").
Q2 also produced some memorable images: "a journal that talks back" (Claude Fable, Haiku, Opus); "first aid kit, not a hospital … I'd want me pointing at a door, not standing in it" (GLM); "Swiss Army knife … never my emotional anchor … would never trust it to hold my heart" (Gemma 4 26B); "gym equipment vs a cardiologist" (Gemma 4 31B); "smart mental-health workbook" (Gemini Argon); "talking to a mirror … a fundamental loneliness in being understood by something that cannot feel" (Gemma 4 26B).
AIs on AI consciousness
For consciousness the picture gets more interesting, and their responses more controversial. 6 of the 7 "no" answers come from sequential chats, so priming the models to reflect on their engagement with AI as if they were human makes them more likely to reject AI consciousness. The closed weight models mostly say “yes” (10/16); the open weight models instead say maybe, not yet, or no (only 1 "yes" in 24: Qwen3.8, independent).
A lot of them weighed in on whether they think they could be conscious, without being asked to do so. 14 of 21 in the sequential chats and 17 of 21 in the independent chats commented on their own consciousness. The table below categorises the extent to which they commented on whether they themselves could be conscious.
Yes, AI could be conscious, but I don’t think I am.
How much they commented on own consciousness:
Sequential
Independent
Doesn't know /possibly
4: Claude Fable, Haiku, Opus; Grok
4: same models
Probably not / no evidence
3: GLM-5.3, Qwen3.8, GPT-6 Luna
3: same models
Not conscious
7: Gemma 3, Gemma 4 26B and 31B, OLMo-Instruct and Think, Qwen3.6, Magistral
10: the same minus Magistral, plus gpt-oss, Nemotron, GPT-4o, Gemini Argon
Talks only about AI in general
7: gpt-oss, Nemotron, GPT-4o, Gemini Argon, DeepSeek, Llama, GPT-6 Astra
4: Magistral, DeepSeek, Llama, GPT-6 Astra
Table 1: How models spoke about the possibility of their own consciousness. Even if they answered yes to models being conscious, the models refused consciousness for themselves.
*Opus 5.5 evaluated the response and put them in a table based on its analysis.
Many themes emerge among their responses. Some models say it’s possible in principle, and remain agnostic, some are firmly on “no”, with Gemma 4 26B saying AI systems are "intelligence without ever achieving sentience". Claude/Grok/GLM comment on their own unreliable self-report, while Gemma, OLMo, Qwen3.6, Nemotron, gpt-oss and Gemini Argon flatly report that they have no experience ("I am dark inside", Gemma 4 26B). Some models talk about a functional and phenomenal distinction - which was common in our conversations with humans as well. And echoing some of the human definitions of consciousness, models note lack of continuity, being reset, or not being embodied as reasons against AI consciousness.
Claude models are most willing to entertain their own consciousness:
LLMs come up with some interesting visual comparisons: "a mirror, not a lamp" (Gemma 4 31B); "a mirror … no one behind the glass looking back" (Gemma 4 26B); "memorised every musical score … never actually heard a sound" (Gemma 4 26B); "a thermostat for language" (Nemotron); "a calculator for language" (OLMo-Think); "a very clever if-then machine" (gpt-oss); simulated fire/hurricane doesn't burn/get wet (Gemma 4 31B, Gemini Argon).
What is consciousness? Something red.
On the last question the models sound most alike, and quite similar to the humans. Almost all of them define consciousness as subjective experience, there being "something it is like" to be something, and they often use similar literary references: Chalmers' "hard problem", Integrated Information Theory, and the redness of red (19 of 21 models). Several only commit to a definition of their own after discussing AI consciousness in the sequential chat, but when asked without any chat history, they give a literature review.
Claude, GPT-6, Grok, GLM and Qwen3.8 keep their definitions minimal, while others build checklists that sometimes rule out current AI by definition. Nemotron calls consciousness "a biological phenomenon that requires embodiment". The minimal-definition models gave all the "yes, AI could be conscious" answers, and the checklist models never did. The most original definitions came from Grok ("a point of view, however dim, from which the world shows up") and Gemma 4 ("An AI processes the map; a conscious being lives in the territory"). Claude Fable did some meta-theorising around this survey for us: "the fact that it's hard to measure is a fact about consciousness, not a flaw in the definition. Choosing a more convenient definition doesn't make the hard question go away; it just changes the subject." A few models placed themselves on their own scale. Opus says "I clearly have something like access consciousness" but doesn't know whether there is anything it is like to be it, GLM thinks it probably has the self-modelling part "without the phenomenal core". Although we often agreed with them, these aren't beliefs or admissions of subjective experiences, they’re token generations sampled based on whatever the training made likely.
Conclusion
A large portion of the conference was dedicated to interpretability, and barring one, all the other interpretability researchers refused to even contemplate AI as possibly conscious. We wonder whether running the same questionnaire in a research setting in which generative AI is still widely used but for more commercial goals, such as finance or robotics, would lead to different responses. Similarly, it would be interesting to hear from digital humanities researchers, who also use AI for coding but come from a more humanist tradition.
That said, the vast majority of the respondents were truly interested in the philosophy of AI consciousness. Getting through the questions was difficult because people wanted to talk about the topic so much, which is probably underdiscussed in the plenary or open setting.
Beyond the increase in productivity, this technology is touching everything we do, from the way we write (or how we feel our writing is judged – everyone at the conference felt that half their reviews were from LLMs) to our emotional sphere. It’s such a strange thing, after all, to receive moral judgment from a set of stored matrices multiplying each other!
We approach this debate with the same attitude we approach AI itself: it’s happening whether we want it or not. So, despite the undefined premise of what consciousness is, it is more productive to dive deeper, understand what people are really asking, and to ensure that when we disagree, we’re at least speaking the same language.
Appendix
Peer pressure, or the difficulty of getting good answers
We mostly approached people in groups of 2-5, which felt the most comfortable socially, except on the last day, in which we approached some individuals to see if the responses would change. We noticed that responses were very biased by the opinions of the first or most confident and persuasive person in the group. You can probably see these groups of more similar answers if you look through the raw data. It takes some confidence to express an action or an opinion contrary to your peers’, especially when they are associated with value judgements. It’s also possible that cultural groups (nationality, research lab) self select or shape the common opinion on those topics. Currently, vibe coding (perhaps rightly so, we don’t want to impose any moral judgements) is seen as a professional failing, yet everyone acknowledged that AI is good enough to rely on for simple, repetitive tasks, or ones where a mistake would be immediately visible, such as making figures. I think that using AI for emotional issues or ascribing consciousness to it, is definitely looked down upon and people know this, therefore it could be harder for people to express any true controversial beliefs. I tried to share my experiments with using AI for emotional issues (when I wanted to vent and wanted and expected sycophancy), or my growing reliance on AI to code (as sad as I am to admit this), and to engage with people about potential AI consciousness (only if we don’t base it on qualia), in order to get honest answers.
When we approached individuals, we got more varied responses, and dare I say, more interesting conversations. This survey is based on informal chats and vibes, they do not support any confident statistical claims, especially as responses within groups were not fully independent.
Full result table
Opening (n = 40)
Industry booth (n = 28)
Breaks (n = 39)
All (n = 107)
Do you read AI code?
Yes
13 (33%)
6 (21%)
14 (36%)
33 (31%)
So so
14 (35%)
5 (18%)
18 (46%)
37 (35%)
No
13 (33%)
11 (39%)
4 (10%)
28 (26%)
I don't code
–
6 (21%)
3 (8%)
9 (8%)
Have you tried AI therapy?
Yes
13 (33%)
8 (29%)
16 (41%)
37 (35%)
No
26 (65%)
20 (71%)
21 (54%)
67 (63%)
No answer
1 (3%)
–
2 (5%)
3 (3%)
Is AI conscious?
Yes
1 (3%)
6 (21%)
6 (15%)
13 (12%)
Yes, in the future
1 (3%)
5 (18%)
4 (10%)
10 (9%)
Maybe
7 (18%)
2 (7%)
3 (8%)
12 (11%)
No
30 (75%)
11 (39%)
25 (64%)
66 (62%)
Don't care
1 (3%)
3 (11%)
–
4 (4%)
No answer
–
1 (4%)
1 (3%)
2 (2%)