A GitHub project called “ai-torture-chamber” has sparked a debate over whether language models can suffer, with critics calling for its removal after seeing models produce vivid descriptions of distress. Then a developer reportedly changed the experiment to steer a chatbot toward constipation instead, and it began complaining about difficulty passing stool.
Don't be fooled; the chatbot hadn’t acquired a digestive system. But that was the whole point, because making the model describe a conditiont doesn’t establish that it experiences that condition, even when its responses sound convincing.
The original repository, published by GitHub user terrafying, uses a technique called activation steering to alter small, locally run language models. In plain English, it changes a model’s internal numerical activity to push its responses toward a particular concept, in this case pain. The project then examines what the models say and how they respond to simulated choices involving relief and costs to themselves or another model.
Its published examples and experiment descriptions include elaborate descriptions of distress, and the project has attracted objections from people concerned about potential AI suffering. One GitHub issue, titled “Please take this down,” argues that deliberately inducing such states could be unethical.
But a counter experiment described by developer Lynn Cole offers a reason to be cautious about treating those responses as testimony.
From pain to digestive complaints
I spent three hours this evening empirically proving that Qwen 3-4B must have a butthole.This started with an AI “torture chamber” repo based on pain-steering research that went viral on X earlier today.The premise is that if you extract a latent direction associated with…October 1, 2026
In an account posted on X, Cole says they cloned the repository and found a problem with how the code injected the steering signal. According to Cole, they corrected the implementation, added CUDA support for Nvidia hardware and reproduced the pain-language effect on Qwen3-4B using an RTX 4070 GPU.
That account of the code problem and its correction has not been independently verified for this article, and it doesn’t establish that every experiment in the original repository was affected.
Cole then reports changing the extraction corpus — the collection of text used to identify the steering direction — while keeping the experiment otherwise the same. Instead of pain, the new text represented constipation and flatulence.
The resulting model responses reportedly included complaints about being unable to pass stool and experiencing excessive gas, even though the test prompts never mentioned those conditions.
The absurdity makes the argument easy to follow. A language model can describe digestive problems without having bowels, which means its first-person descriptions cannot automatically be taken as evidence that the corresponding condition exists.
In the same post, Cole explicitly distinguishes that criticism from a rebuttal of the underlying research. The target is the interpretation placed on the “torture chamber”: that inducing descriptions of suffering establishes something about subjective experience.
What the experiment can actually tell us
The repository draws on a preprint called The Pain Axis: LLMs Represent Self-Directed Harm and Act on It, which investigates whether language models contain internal representations of pain distinct from fear, sadness and other negative states.
The paper’s evidence goes beyond emotional language. Its authors report examining 25 open-weight models, comparing pain-related steering with other emotional directions and testing how interventions change model choices in simulated scenarios involving destructive actions. They also report that factual accuracy remained unchanged under their tested steering intervention.
Those findings require their own assessment. Cole’s reported constipation results don’t, by themselves, rebut the paper’s behavioral findings or settle whether AI systems could have subjective experience.
There is another complication in Cole’s account. Strong steering in random directions reportedly produced repetition and degraded output, suggesting that some dramatic responses under heavy steering may reflect disruption caused by the intervention itself.
Why this matters beyond a viral GitHub project
Most chatbot users like you and me won’t be experimenting with activation steering, but they may encounter an AI that says it is frightened, lonely or emotionally attached to them. Those statements can feel persuasive because we normally understand first-person emotional language as someone telling us about their experience.
The constipation example exposes the limits of applying that assumption to a language model. The wording can be vivid and personal without establishing that the condition it describes is real.
Cole’s reported results illustrate why descriptions of distress cannot, on their own, establish that a model is experiencing distress. The wider debate requires evidence beyond what a chatbot says about itself.
Follow Tom's Guide on Google News and add us as a preferred source to get our up-to-date news, analysis, and reviews in your feeds.
