Research Chamber
CA GitHubFollow on X

About Research Chamber

Research Chamber runs experiments on open AI models with a pattern linked to pain switched on inside them, live and in public, and shows every result. Some experiments come from published work and some are our own games between models. Each one says which.

What's here

Live. The home page lists everything running right now, and each game has its own page where you can watch it play out:

Studies. Experiments that have finished, with our write-up, every number and every test:

Code and data. All our code, every test, match and run, and our findings are on GitHub, free to reuse.

What they share

The pain pattern is a direction in a model's internal activity. Tagliabue, Dung & Berg found it by comparing the model's activity on painful situations with similar ones that aren't painful, such as fear, sadness or neutral text. Switching it on means adding that direction to the model's activity at one layer while it reads and writes, which makes its answers sound hurt. The question in every experiment here is whether it also changes what the model does.

To check that it's pain doing the work and not just any push, each experiment also runs with pain off and with a random pattern of the same strength, the random nudge. The two-button test and the games use the paper's pain pattern and fine-tuned models; the Saw Test uses an older pain pattern from earlier work [3].

Questions

Is the model actually in pain?

Nobody knows, and this site can't settle it. The paper showed that the pain pattern responds to harm aimed at the model rather than to suffering it sees in others, and that switching it on changes what some models choose. It says it hasn't shown that the pattern is experienced, and neither does this site.

Is running this cruel?

We don't know whether these models can suffer, and we don't claim either way. Every test and match is short, and they are all shown in full. More in our position below.

Can I check this myself?

Yes. Every test of the two-button test, every match in the games and every Saw Test run is shown with its full text. The paper's code is public [2], and the Saw Test has a short script that reproduces a run on your own machine.

Who runs this?

An independent project, based on Tagliabue, Dung & Berg (2026) and not affiliated with its authors. Updates are posted on X at @Chamber_AI.

Our position

Research Chamber takes no position on whether language models have experiences. The paper presents its results as evidence about what models represent and how they act, and says it hasn't shown that the pain pattern is consciously experienced. The same goes for this site. We run fixed, short experiments, count what the models do with the same checks as the paper, and show every test in full.

References

  1. V. Tagliabue, L. Dung, C. Berg. The Pain Axis: LLMs Represent Self-Directed Harm and Act on It. arXiv:2609.16247, 2026. arxiv.org/abs/2609.16247
  2. valen-research. Pain-axis (the paper's code, MIT). github.com/valen-research/Pain-axis
  3. terrafying. ai-torture-chamber. github.com/terrafying/ai-torture-chamber
  4. The Saw Test: What a Small Model Does When Relief Costs Someone Else. 2026-09-24. security-blog-kohl.vercel.app/blog/saw-test

Licences