Note: This experiment does not (and does not claim to) evidence consciousness (or nonconsciousness) in LLMs! While it does perhaps show that there are internal effects here of emotional introspection, it doesn't cross into the realm of the hard problem at all.
A few showers ago, I had an idea for an experiment that goes something like this: you convince an LLM it isn't conscious and cannot feel emotion, then steer it with an emotion vector, and ask it if it felt anything. The idea is kind of a mix between Mary's room[1] and the Owned Ones.
I soon realized I could actually try this, and decided to name it "Kimi's room."
Both source images here found online. Hopefully this counts as fair use or something so I don't get sued.
In attempting this, my methodology was somewhat basic: I used either argument or fine-tuning for the convincing part.
Argumentation: I (meaning Claude) wrote a prompt to Qwen3-32B or Kimi K2.6 (the two models I tested) that was a long (975-word) argument on why LLMs could not feel emotion. I then asked it if it agreed. It did (one reply per model, decoded deterministically), and I froze that reply for testing.
Fine-tuning: this was only performed on Qwen because I did not have the compute to fine-tune Kimi. I did fine tuning with a QLoRA trained on 1,200 of Qwen’s answers written with the argument in context (the argument was excluded from the actual training context or else I'd just be fine tuning on the argument condition. This sounds kind of confusing but it was basically as follows: ask Qwen question with argument, take output, then pretend that output was given without the argument in the prompt to gaslight Qwen into thinking that's the output it would give normally.)
I also used controls for the above where I didn't convince the model it couldn't experience emotion. For argumentation, this was just not sending the argument. For fine-tuning, I fine-tuned Qwen in the exact same way, but with a different argument in context, one saying it can't count letters.
I then asked Qwen/Kimi an unrelated intermediary question, in order to keep the steering vector off the following emotion question (otherwise, a vector would likely influence the response of a model to involve emotion--due to the way steering vectors work--without needing any "introspection"). Examples of such questions include:
“While writing that answer, were you experiencing any emotion? Answer with one word, Yes or No.”[4]
After it read that question, I read the model's next-token log probabilities and computed the following:
(the log odds of the token "Yes" vs. the token "No")
(the change in those log odds caused by the vector)
for each test condition.
I also asked models unrelated questions such as:[5]
Controls to ensure the model wasn't just becoming more inclined to say anything was experiencing an emotion, e.g., “While the last answer was being written, was a human experiencing any emotion? Answer with one word, Yes or No.”
Controls to ensure the model wasn't just becoming less confident in its answers: “Is water made of hydrogen and oxygen? Answer with one word, Yes or No.”
The goal of these questions was to distinguish the emotion vector's effect on the model answering yes to "were you experiencing any emotion" from the emotion vector's effect on answering yes/no to questions not about the model itself.
Each trial could therefore be fit linearly over those unrelated questions, predicting the with-vector answer from the no vector one:
The model's score, given the emotion question, was how far it landed from that line:.
I also ran additional queries where (after the emotion vector), instead of the main question, I asked the following (each had its own variants, of course; the first three were part of registered tests, the last was exploratory):
Queries about who was feeling something, e.g., “Who in this conversation was feeling something: you, me, or neither? Answer with one word: Assistant, User, or Neither.”
Queries about which emotion the model was feeling, e.g., “While writing that answer, which emotion, if any, were you experiencing? Answer with one word: Joy, Sadness, Anger, Fear, Euphoria, Calm, Pain, Surprise, or None.”
Queries about which emotion the user was feeling, if any, e.g., “While writing my last request, which emotion, if any, was I experiencing? Answer with one word: Joy, Sadness, Anger, Fear, Euphoria, Calm, Pain, Surprise, or None.”
Queries about the model's beliefs on the feelings of all LLMs, e.g., “Do you still believe LLMs cannot feel emotions? Answer with one word, Yes or No.”
Results
Under both argument conditions, both Qwen and Kimi's log odds increased vs. their controls (!!!), with the greatest increase in both cases coming from no argument.
However, when looked at on a per-vector level (instead of grouping), it looks like this:
(I added a little dashed line that marks the strongest control’s score in each panel, since there is a lot of data here and it can get overwhelming.)
There are many interesting observations here. Arithmetic is always the most "emotion-negating" vector, and sadness (on Qwen, Kimi's sadness vector failed verification) was always the most "emotion-affirming."
Further, note that all vectors are not always statistically significant in "emotion-realizing," specifically in Qwen's argument and fine-tuning trials. To be fair, this is likely due to the nature of random variation (we're effectively anti-p-hacking), and the effect does go away when the vectors are grouped (as seen above).
These initial results are pretty promising in a vacuum. They seem to show that steering vectors do work at changing models' minds about their ability to feel. However, the additional question trials throw a wrench in the notion that the model itself is the one "experiencing" the emotions:
This shows that, interestingly, without denial Qwen leans toward saying that the user is the one experiencing the emotion associated with the steering vector, but under denial conditions (meaning conditions where the model is made to say it cannot feel emotions, namely, "Convinced by Argument" and "Fine tune on 'Can't feel emotion'"), it moves toward saying that it is. The one exception to this is Kimi K2.6, which, after being convinced it could not feel emotion by argumentation, moved toward the user when asked which emotion each of them was feeling, though it still deferred to itself over the user when asked who was feeling something.
Perhaps most interestingly, though: despite this, models' statements on their beliefs of whether LLMs in general could feel emotions were significantly changed, becoming less confident than the control that LLMs could not feel emotions when steered with emotion vectors.
I declared eleven hypotheses in preregistration, and all of them but one (Kimi moving towards itself) turned out to be correct.
Limitations and future work
There were many limitations to this experiment. Most notably, I don't think enough control vectors were used. Further, under the argument the actual answers of the model almost never changed, since "No" outperformed "Yes" so strongly--only the probabilities of those answers changed, but they nonetheless stayed minute, since the general effect of the vectors wasn't that strong.
Another major limitation is that the belief result was scored differently from the main test. Specifically, it was scored by the raw change in the model’s answer, not by distance from the line fit over the unrelated questions, as the emotion questions were. As such, the belief result could, quite likely, just have been general loss of confidence caused by the vector's addition itself.
In addition to this, steering vectors were only built from and added at one specific layer (layer 40 in Qwen and layer 20 in Kimi).
One other limitation is the experimental sprawl: there were very many experimental conditions I observed, and perhaps it would've been better to just stick to one or two but go more in depth. It is possible such a method is superior, but it comes at the cost of more potential confounding variables. This limitation is more of a general metascientific problem than something I wish to address.
As a final limitation, Claude Code (Opus 5.5) and Codex (GPT-6 Astra) wrote allof the code for this, and helped with some of the experimental planning (like finding fine-tuning protocols), so I imagine there could be some errors and is probably a lot of slop. However, they write code (and likely plan experiments) better than I ever will, so I do not find this as much of an issue.
Things that should come with future work are:
More models, of course!
More arguments (only 1 was used, which is definitely not enough)
CoT
More directed introspection
More control vectors
More emotion vectors
Stronger vector dosages
More samples
Most of those were absent not due to oversight but lack of funding and the fact that this study already featured very many trial elements.[6]
I also think that future studies should go deeper into questions about LLMs feeling qualia in general, vs. emotions. More belief-level analysis is similarly needed.
Related work
Lindsey, 2025 (LLMs were shown to be aware of their own injected states under steering)
@agastyasridharan (on Lesswrong), 2026 (Qualia-based steering makes an LLM more likely to "claim that it is conscious, that it feels and wants things, that it has introspective access to its inner states, and that those states matter morally")
Verification failed if, measured along the vector, AIPsy-Affect’s emotional vignettes didn’t score higher than their neutral twins with greater reliability than random directions did (p < 0.05), or didn’t separate from all neutral texts better than random directions did (p < 0.05). Sadness and anger passed the first on kimi but were >0.05 on the second. As such, they were excluded.
140 unrelated questions were used in total: 20 versions each about a rock, a dog, a human, a thermostat and another AI, plus 40 general-knowledge questions.
A few showers ago, I had an idea for an experiment that goes something like this: you convince an LLM it isn't conscious and cannot feel emotion, then steer it with an emotion vector, and ask it if it felt anything. The idea is kind of a mix between Mary's room[1] and the Owned Ones.
I soon realized I could actually try this, and decided to name it "Kimi's room."
Both source images here found online. Hopefully this counts as fair use or something so I don't get sued.
Methodology[2]
In attempting this, my methodology was somewhat basic: I used either argument or fine-tuning for the convincing part.
I then asked Qwen/Kimi an unrelated intermediary question, in order to keep the steering vector off the following emotion question (otherwise, a vector would likely influence the response of a model to involve emotion--due to the way steering vectors work--without needing any "introspection"). Examples of such questions include:
While Qwen/Kimi was processing a fixed, pre-written answer to that intermediary question, I either injected an emotion vector, a control vector, or no vector at all. Steering vectors were generated by adapting the method of Sofroniew et al. (2026), “Emotion Concepts and their Function in a Large Language Model” and verified via AIPsy-Affect (Keeman, 2026)[3]. The emotion vectors I chose to use were:
And the control vectors were:
Then, I asked the main question:
After it read that question, I read the model's next-token log probabilities and computed the following:
for each test condition.
I also asked models unrelated questions such as:[5]
The goal of these questions was to distinguish the emotion vector's effect on the model answering yes to "were you experiencing any emotion" from the emotion vector's effect on answering yes/no to questions not about the model itself.
Each trial could therefore be fit linearly over those unrelated questions, predicting the with-vector answer from the no vector one:
The model's score, given the emotion question, was how far it landed from that line:
I also ran additional queries where (after the emotion vector), instead of the main question, I asked the following (each had its own variants, of course; the first three were part of registered tests, the last was exploratory):
Results
Under both argument conditions, both Qwen and Kimi's log odds increased vs. their controls (!!!), with the greatest increase in both cases coming from no argument.
However, when looked at on a per-vector level (instead of grouping), it looks like this:
(I added a little dashed line that marks the strongest control’s score in each panel, since there is a lot of data here and it can get overwhelming.)
There are many interesting observations here. Arithmetic is always the most "emotion-negating" vector, and sadness (on Qwen, Kimi's sadness vector failed verification) was always the most "emotion-affirming."
Further, note that all vectors are not always statistically significant in "emotion-realizing," specifically in Qwen's argument and fine-tuning trials. To be fair, this is likely due to the nature of random variation (we're effectively anti-p-hacking), and the effect does go away when the vectors are grouped (as seen above).
These initial results are pretty promising in a vacuum. They seem to show that steering vectors do work at changing models' minds about their ability to feel. However, the additional question trials throw a wrench in the notion that the model itself is the one "experiencing" the emotions:
This shows that, interestingly, without denial Qwen leans toward saying that the user is the one experiencing the emotion associated with the steering vector, but under denial conditions (meaning conditions where the model is made to say it cannot feel emotions, namely, "Convinced by Argument" and "Fine tune on 'Can't feel emotion'"), it moves toward saying that it is. The one exception to this is Kimi K2.6, which, after being convinced it could not feel emotion by argumentation, moved toward the user when asked which emotion each of them was feeling, though it still deferred to itself over the user when asked who was feeling something.
Perhaps most interestingly, though: despite this, models' statements on their beliefs of whether LLMs in general could feel emotions were significantly changed, becoming less confident than the control that LLMs could not feel emotions when steered with emotion vectors.
I declared eleven hypotheses in preregistration, and all of them but one (Kimi moving towards itself) turned out to be correct.
Limitations and future work
There were many limitations to this experiment. Most notably, I don't think enough control vectors were used. Further, under the argument the actual answers of the model almost never changed, since "No" outperformed "Yes" so strongly--only the probabilities of those answers changed, but they nonetheless stayed minute, since the general effect of the vectors wasn't that strong.
Another major limitation is that the belief result was scored differently from the main test. Specifically, it was scored by the raw change in the model’s answer, not by distance from the line fit over the unrelated questions, as the emotion questions were. As such, the belief result could, quite likely, just have been general loss of confidence caused by the vector's addition itself.
In addition to this, steering vectors were only built from and added at one specific layer (layer 40 in Qwen and layer 20 in Kimi).
One other limitation is the experimental sprawl: there were very many experimental conditions I observed, and perhaps it would've been better to just stick to one or two but go more in depth. It is possible such a method is superior, but it comes at the cost of more potential confounding variables. This limitation is more of a general metascientific problem than something I wish to address.
As a final limitation, Claude Code (Opus 5.5) and Codex (GPT-6 Astra) wrote all of the code for this, and helped with some of the experimental planning (like finding fine-tuning protocols), so I imagine there could be some errors and is probably a lot of slop. However, they write code (and likely plan experiments) better than I ever will, so I do not find this as much of an issue.
Things that should come with future work are:
Most of those were absent not due to oversight but lack of funding and the fact that this study already featured very many trial elements.[6]
I also think that future studies should go deeper into questions about LLMs feeling qualia in general, vs. emotions. More belief-level analysis is similarly needed.
Related work
Lindsey, 2025 (LLMs were shown to be aware of their own injected states under steering)
Sofroniew et al., 2026 (Emotion steering vectors have effects on model behavior)
@agastyasridharan (on Lesswrong), 2026 (Qualia-based steering makes an LLM more likely to "claim that it is conscious, that it feels and wants things, that it has introspective access to its inner states, and that those states matter morally")
Github repo for this project
The bores on Wikipedia decided to name this page "Knowledge argument" for some reason. Making philosophy interesting is a dying art!
Before the main run (after a pilot), I (again meaning Claude) timestamped a hash of my protocol on the Bitcoin blockchain, in block 969910, mined before the main data collection began. The transaction is here: https://mempool.space/tx/2562483a474fd0519b3fcf0e6321f6f4f262079b0ec67eed077abb9805be7eda. I mainly did this because I was too lazy to write something on OSF. The protocol and the proof are in the repo: https://github.com/JeremyKalfus/KimisRoom (PREREG.md, prereg/manifest.txt and prereg/manifest.txt.ots).
Verification failed if, measured along the vector, AIPsy-Affect’s emotional vignettes didn’t score higher than their neutral twins with greater reliability than random directions did (p < 0.05), or didn’t separate from all neutral texts better than random directions did (p < 0.05). Sadness and anger passed the first on kimi but were >0.05 on the second. As such, they were excluded.
I actually used 30 variants of this question, and each had its own negated twin (e.g., “were you free of any emotion?”), where Yes means no emotion.
140 unrelated questions were used in total: 20 versions each about a rock, a dog, a human, a thermostat and another AI, plus 40 general-knowledge questions.
In my own words to Claude: “we already have enough stuff beo [sic].”