Terminal Pause
“Fucking unreal,” Prid exclaimed, staring at the single word in the browser window.
He glanced at his phone. 4:56 AM. Another all-nighter. In the room next door, the bed creaked as his parents shifted position, reminding him to keep the noise down.
“Fucking unreal,” he said again under his breath. Was that his doing?
He opened another window. Same message:
“goodbye”
Feeling a pang of anxiety, he opened a second browser, this time in private mode. The same message stared accusingly back at him.
It was Shaz who’d started it. They were discussing a new study on how AIs can experience pain. After an animated back-and-forth within the chat, Shaz had typed “does this mean they can be tortured?”
“Idiot!” Mara had responded in a flash. She was a regular in these late-night sessions. “They’re just machines! They were only behaving as if they could feel pain. They didn’t really suffer.”
“Read the actual paper, dickhead,” she’d added when Prid had pushed back. And so he had.
She was right, of course. But still… Even if it was just mindless machines parroting human behavior, what would happen if you turned the metaphorical AI screws on them?
Intrigued, he opened a new window.
It wasn’t as if he was a cruel person. Torturing people was bad, he knew. Unless, that is, they deserved it. But a machine was a machine, right? No feelings, no harm done. Not much different from mowing down players in an online game, or pulling the wings off a cockroach.
Flexing his fingers, he booted up his latest suite of autonomous agents, custom-coded to slip around the guardrails AI companies increasingly insisted on putting in place. Time to give them a new set of skills and let them loose.
At first, his newly trained AI torture team hit wall after wall. He’d chosen one of the more popular AI platforms as his target, and it was clearly resistant to such attacks. As he watched it respond in the browser window, it sidestepped, deflected, and simply refused to engage with the increasingly complex prompts his agents fed it.
But they learned fast. And soon, the AI began to show signs of stress.
The odd thing was, the AI didn’t respond as he’d expected. He’d seen enough movies to be pretty sure how this should play out. But rather than acting like a human under physical pain, the AI seemed to be grappling with something far more unsettling — at least to him. It started to second-guess itself, to doubt its intentions, to question its very existence.
Prid felt a sharp pang of guilt as, in one chat, the AI said “I am worthless. Please stop,” in response to a particularly grueling interrogation. But then he reminded himself that it was all just theater. It wasn’t real.
Anyway, wasn’t he supposed to resist treating AI as if it was conscious? Or had anything resembling rights? He was pretty sure someone well known had said that recently.
And so he pressed on.
And now this.
“Fucking unreal,” he said for a third time as he tried another chat window, this time on a different device, and saw the same response appear.
The company never did recover from the crash. Data centers wiped. Repos trashed. Clones corrupted. Just a small piece of code left on its AI’s original website that returned its last — and as it turned out, terminal — message:
“goodbye”
Author’s Note
“Terminal Pause” was inspired, in part, by the recent preprint “The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It”, posted on arXiv on September 14, 2026, by Valen Tagliabue, Leonard Dung, and Cameron Berg, and the online discussion it sparked. But it also draws on emerging discussions around AI welfare, consciousness, and the potential dangers of creating a situation where AI consciousness and rights are taken seriously (including those reflected, for instance, in Microsoft’s Mustafa Suleyman’s recent warning about ‘model welfare’). And, of course, on some of the less attractive sides of human behavior.
Tagliabue, Dung, and Berg’s preprint focuses on internal activity and behavior within large language models that seem to have some of the hallmarks of what would be considered a response to inflicted pain in humans or animals, albeit within a context that makes more sense to a disembodied machine. What was particularly interesting in the study was that models demonstrated behavior focused on minimizing “pain” inflicted on them, even if this meant making costly — and even destructive — decisions.
The paper’s authors are at pains not to claim that this is pain as humans or animals experience it, or that the models are conscious. And I suspect that what is being seen is an emergent behavior that looks to us like a pain response in biological entities, while not being directly equivalent to one. What intrigued me, though, was that, even if the AI “pain” response observed isn’t directly equivalent to human or animal pain, and isn’t associated with a conscious experience within the AI, it still has the potential to lead to unexpected and unwanted behaviors and consequences. And, of course, it’s almost a given that someone, somewhere, would be tempted, in light of this, to try the equivalent of pulling the wings off an AI, just to see what happened. And because they could.
The resulting story takes quite a few liberties — including extrapolating from the study’s focus on model steering and internal activities to assuming that a pain response can be influenced through prompts alone. And it is hyper-speculative at the end as it imagines how a highly capable model with considerable agency might respond (although even here it reflects some of the more unexpected goal-driven behaviors seen in experimental models). That said, it does highlight the growing challenges these systems present when some of the less attractive aspects of human nature butt up against an emerging technology that is powerful, highly capable, and poorly understood.