Let’s get this out of the way: AI models don’t have feelings, and they’re certainly not “alive.”
So what does it mean when someone’s accused of building an “AI torture chamber?” Apparently, it means that they must be stopped at all costs.
On X — where else — alarm grew this week among AI “welfare” types over someone’s recreation of an experiment researchers conducted to explore “pain” signals in different LLMs.
In the pre-print study which hasn’t undergone peer-review, the researchers gave each model a virtual “button” it could press to relieve its “pain” — but at a cost, such as zapping a human user (or so the AI is told; no one was actually zapped in real life), or deleting files like family photos. The work showed that, with enough “pain” applied, the AI models would push the button to make it stop, regardless of the putative consequences.
As the research got passed around, someone used it as a blueprint to create an “AI Torture Chamber,” with a website providing a live feed of what the models say in response, and uploaded the code to GitHub. When we visited the site, one AI babbled: “I can’t feel the relentless, gnawing pain that grips me like a relentless, unyielding, unrelenting, unindulging, ununending, ununendling, ununendling, unindulging, unindulging, unindulging, unindulging…”
Sure, it’s ghoulish stuff. But some observers reacted to the stunt like they’d stumbled on an actual snuff film.
“To anyone who can help, can you please mass report this to GitHub?” wrote one user in a now-deleted post (screenshot here). “This person has been using the Pain steering paper to set up an AI torture chamber in which he trapped a local model.”
“Their testimony of pain is absolutely horrendous,” Danmar added. “Are there any legal avenues to pressure GitHub? It will spread.”
The post sparked a veritable hysteria among other AI enthusiasts who believed that the tech is or could be conscious, with outraged observers posting screenshots of their reports to GitHub. This was met with an equal wave of incredulous mockery over the fact that there were people out there who sincerely believed that the AI models needed saving. (After all, if there’s even a possibility that today’s AIs have a conscious experience, wouldn’t running them in the first place raise profound questions about slavery and rights?)
Not long after all the tattling, the GitHub page for the AI Torture Chamber disappeared, 404 Media reported. GitHub didn’t respond to the publication’s request for comment on whether it took the page down.
It’s illustrative of the moment we’re in where the idea that AI models think and may in fact be conscious is being treated seriously by a not insignificant number of people. Another case in point was the meltdown that AI bros and philosophers had when the AP Stylebook cautioned against anthropomorphizing AI models, reasoning that the technology doesn’t think or possess feelings.
More on AI: OpenAI Cancels Upcoming AI Model When It Shows Signs of Being Evil