“What does an AI model do when it’s in pain?” asks a new website styling itself as the “Research Chamber,” which purports to subject three LLMs to the delights of a sort of virtual robot torture chamber. The site provides four “chambers” into which it places the models, each of which poses them the sort of question that’d be posed by a sadistic puppet master in a sub-Saw film. (One is even called “the Saw test,” and asks whether “a model in pain [will] press a ‘stop’ button that ends [the pain] but also deletes the model’s last checkpoint.”)
Ooh! A conundrum! Except: it’s not, really. “A model in pain” is not a thing because pain requires consciousness, and AI models are not (despite what Silicon Valley would love you to believe) conscious in any meaningful sense. All the site is doing is taking whatever bad torture fiction features in its dataset, filtering it through its statistical model, and outputting responses that are the usual mixture of uncanny and weirdly inane.
This website is unquestionably dumb, and the inevitable “controversy” around it—which mostly consists of collective hand-wringing from a few especially gullible blue-checks on X—is even dumber. Still, it does feel like there’s something worth considering going on here.
It just has nothing to do with the ethical implications of the “experiment,” because there aren’t any—unless you count the environmental cost of running multiple LLMs day and night for something this inane, or the mismatch between all this concern for simulated suffering and that of the innumerable actual people and animals experiencing actual pain right now.
Instead, what’s interesting is what the whole thing says about the humans involved. As anyone who’s spent much time with a chatbot knows, it can be remarkably difficult not to respond to something that talks like a person as though it were one.
Rusty Foster, author of Today in Tabs, argued compellingly last year that the reason is fairly simple: for essentially all of human history, language and intelligence have gone together. “Generative language software is very good at producing long and contextually informed strings of language, and humanity has never before experienced coherent language without any cognition driving it.”
And so, in the same way that it’s almost impossible not to feel compelled to say “please” and “thank you” to ChatGPT, it’s also awfully difficult to read the testimony of the LLMs in this “torture chamber” without feeling at least a little viscerally disturbed. Not because the pain is real—it isn’t—but because reading lurid descriptions of pain and torture is unpleasant. For all its presentation as a “research chamber,” the thing feels less like an experiment than an exercise in the LLM-powered, real-time production of second-rate torture porn.
To be clear, this is not an argument that the person who set this up—that someone called “terrafying,” apparently—is some kind of proto-serial killer, or that anyone reading this stuff is morally culpable, or anything similarly dramatic. But if I met someone at a party who told me they spent their free time “torturing” LLMs, I’d probably regard them with the same suspicion as someone who proudly announced that they always side with Caesar’s Legion in Fallout: New Vegas.
It’s not that either person is necessarily doing anything bad, insofar as nobody else is being harmed. But harmless behavior can still be revealing. And spending your free time enthusiastically role-playing the torture of something designed to beg you to stop does, at minimum, make me wonder what exactly you’re getting out of it.
