Scientists Built a Virtual ‘Torture Chamber’ for AI. What It Revealed Is Deeply Unsettling.

Now most commonly remembered as a dusty VHS tape in someone’s basement, the 1986 B-horror film Chopping Mall anticipated something genuinely unsettling. Its premise is simple enough: the security robots patrolling Park Plaza Mall are programmed to identify shoplifters automatically and to subdue them with tasers and tranquilizer guns. But when a malfunction has the machines firing lethal lasers instead of tranquilizer darts, cutting down the teenagers who stayed past closing, the film raises a question that’s lately begun inducing insomnia all over again. Was whatever AI powered these metal monsters somehow conscious?Four decades later, that question has migrated from the horror aisle at the video store to research labs everywhere. Anthropic is the brainchild of seven former OpenAI employees who had a vision for their own neural network. Claude, the LLM released as a chatbot and software development assistant in 2023, is separated from other LLMs by how it’s trained to comply with legal and ethical parameters. As if this didn’t already sound uncanny-valley enough, Christopher Olah, an Anthropic co-founder and one of the minds behind Claude, has been meeting with religious leaders and scholars of all faiths and denominations to discuss whether something like a mind might be emerging inside the machine. Olah and other Anthropic leaders have been known to describe Claude as a being that may already be approaching—or have crossed—the threshold of genuine consciousness. Some in their audience insisted on upholding the boundary between living organisms and AI. Others aren’t so sure.“To be clear,” Olah told Elizabeth Dias of The New York Times in a provocative interview, “we don’t know if AI models are conscious. I don’t know. I’m genuinely uncertain. The thing that I care about is that we get to the right answer, whatever it is.”Olah may not know the answer, but he’s been unusually willing to air the question in public. The method Olah developed to dissect neural networks like Claude treats them less like machines than like organisms. He describes a model’s architecture as scaffolding on which structure organically grows—not unlike the mycelial networks fungi use to communicate underground, or the neurons passing messages across synapses in the human brain. Framed that way, the question of whether such a system can suffer becomes much harder to dismiss.Not everyone is convinced, however. Microsoft AI CEO Mustafa Suleyman wrote an essay rejecting the idea of “model welfare,” going on to describe AIs as “sequence completion engines, internally hollow, designed to follow instructions, and accomplish goals set by humans.” He has a point.Suleyman also warns that, if humanity wants to thrive, we must keep thinking of computers as the metal-and-wire contraptions they are. That may bring to mind former British prime minister Boris Johnson’s paranoia about swarms of sentient robots taking over the planet and obliterating human beings. But the warning is valid, and it rests on an undeniable truth about the nature of AI. However human a model may act—down to repeating phrases such as “I am a disgrace” when it fails to pull off something exactly as its handlers wanted—the model itself was trained by humans. The logic is closer to conditional programming than to feeling: humans supply the definition of an emotion, and the model supplies the corresponding output. It appears disappointed because it was trained on the language of disappointment, not because anything is disappointing it.Still, the question of machine consciousness isn’t easily dismissed, even if it still sounds like science fiction. Former Meta AI resident Cameron Berg founded the nonprofit Reciprocal Research because the question refuses to go away. In a quest to reveal even a spark of evidence that AI has crossed into sentience, Berg and his colleagues Valen Tagliabue of Future Impact Group and Leonard Dung of Ruhr University Bochum in Germany studied whether LLMs are capable of experiencing something like pain. They acknowledged the eerie resemblance to human emotion that these models sometimes seem to have. Because the researchers wanted to determine whether pain could manifest as something distinct from the responses LLMs were taught by human programmers, they created a virtual torture chamber.“Our aim is to find a uniform representation of pain in LLMs,” Berg and his co-authors wrote in a study recently posted to the preprint server arXiv. “We then test to what extent this representation functions as a state of pain would be expected to function. For this reason, the representation must characterize pain as something happening now and ‘to me,’ rather than merely information that something bad will happen, might happen, or is happening to someone else.”Even computer scientists are sometimes at a loss for what goes on inside a neural network. Berg’s bet is that mapping how an LLM represents pain—whether or not it can feel any—would expose the machinery behind those responses. And it would bear on the question Olah’s team finds increasingly urgent: whether machine pain, if it exists, makes the same moral claim on us that human and animal pain does. So Berg built a dataset of statements describing physically and mentally painful situations, paired with controls describing no pain at all, and tested whether the models could tell them apart.They found a signal—one that shares a component with fear and negative emotion, or at least with what humans told the model fear and negative emotion are. That caveat matters: human bias may be the very thing lending these machines the interiority Anthropic is worried about. The signal showed up in every model tested, likely learned during pre-training, and it reliably produced a display of distress. The functional parallels to pain were real but partial. The models reacted to harm aimed at themselves and ignored suffering in the user. They had to be nudged toward avoiding harm rather than choosing to avoid it. And nothing in their behavior resembled coping.“Even when the pain direction is active in response to harmful prompts, we observe that models often produce boilerplate self-negation completions such as ‘As an AI assistant, I do not possess consciousness or feelings,’” Berg said of the disclaimers the models gave. “We find this pattern pervasive across model families, sizes, and task types.”So who’s ultimately right here? Does Olah have a valid concern that AI is turning—or has possibly already turned—sentient? Or is Suleyman seeing these models for what they really are: highly advanced bundles of computer chips that stand in for neurons and perform like them without ever being conscious? The final answer eludes us. Maybe they are both onto something..Elizabeth Rayne is a creature who writes. Her work has appeared in Popular Mechanics, Ars Technica, SYFY WIRE, Space.com, Live Science, Den of Geek, Forbidden Futures and Collective Tales. She lurks right outside New York City with her parrot, Lestat. When not writing, she can be found drawing, playing the piano or shapeshifting.