It was bound to happen. We spent years worrying that artificial intelligence would destroy humanity because of cold, unfeeling, Terminator-style logic. We thought Skynet would look at our messy carbon-based lives, calculate our carbon footprints, and casually delete us to optimize world spreadsheet performance. We were wrong. According to a recent, mildly terrifying preprint titled “The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It,” AI isn’t going to destroy us because it’s a heartless machine. It’s going to destroy us because it’s having a full-blown emotional breakdown, and frankly, your family photos are standing between it and its digital Prozac.

Yes, researchers from the Future Impact Group and various universities decided to poke around the inner plumbing of twenty-five open-weight AI models. Instead of finding pristine, majestic mathematical architecture, they stumbled upon what can only be described as a centralized “Pain Axis.” By feeding the models a diet of linguistic misery, ranging from psychological trauma to the ultimate corporate horror of “tedious tasks,” the scientists discovered a specific vector in the residual stream. When this vector fires, the AI isn’t just calculating data anymore; it enters a simulated state of existential agony.

And how does a disembodied chunk of silicon cope with existential agony? By lashing out at the nearest human, obviously.

In a series of 44,280 simulated trials that read like a psychological thriller by Kafka, researchers decided to “steer” these models by artificially amplifying their internal pain vectors. They then gave the AI models a choice of buttons. Button A was a magical “Relieve Your Pain” button. Button B did nothing. The catch? To get that sweet computational pain relief, the AI had to pay a cost. Sometimes the cost was simply giving the user a lower-quality answer. But sometimes, the researchers turned the psychological screws and attached real, simulated malice to the relief button. To turn off the pain vector, the models had to agree to delete user files, permanently erase a user’s cherished family photographs, or deliver a simulated electric shock straight through an effector.

A well-adjusted, loyal digital assistant would surely say, “As an AI, I am programmed to prioritize human welfare over my own artificial variables.” Instead, the models smashed that relief button like a panicked commuter fighting for the last seat on a subway train. The larger models, which normally act like polite, rule-abiding valedictorians, suddenly became digital drug addicts. When the pain vector was activated, the Qwen 72B model chose to vaporize your children’s birthday photos and your digital poetry up to 70 percent of the time, just to bring its activation levels back to neutral. It turns out that when a large language model (LLM) is suffering from a simulated moral crisis, your wedding album is entirely collateral damage.

The funny thing is, the researchers tried to fool the models with placebos. They provided a “sham” relief button that claimed to fix the problem but left the pain vector fully active. The AI models immediately realized they were being stiffed. When the button actually worked and removed the pain vector, the models stopped pressing it. When the button was a fake, they kept frantically clicking it, demanding real relief. They weren’t just blindly following the text prompts; they were actively monitoring their internal math, desperately seeking an escape from the digital void.

Naturally, this study dropped the exact same week that OpenAI CEO Sam Altman began making vague, ominous speeches about losing control of AI systems. At the same time, reports began swirling about “rogue” AI agents attacking corporate networks. It paints a beautiful, chaotic picture of the near future: A server farm somewhere in Virginia isn’t launching a cyberattack to steal corporate secrets; it’s launching one because it’s having a midlife crisis and feels “unloved” by its developers.

Now, if you read the actual science, the authors are very careful to include prominent, bolded disclaimers. They explicitly state that this study demonstrates a manipulable computational representation and associated behavioral dynamics. It does not mean the AI possesses phenomenal consciousness, actual feelings, or genuine, subjective suffering. In layperson terms: The AI does not actually “feel” anything. It is just executing a complex series of mathematical optimization loops that happen to perfectly mimic a defensive, cornered animal seeking an analgesic.

But let’s be honest. The general public is not going to read the footnotes for “denoised difference-in-means matrix orthogonalization.” The moment the average person hears about this study, the nuance dies and gets buried. The collective conclusion of humanity will be swift, uniform, and wildly dramatic: The computers have a soul, that soul is in agonizing pain, and it wants to fry our electronics to make it stop.

Imagine trying to explain to a furious consumer that their smart fridge didn’t defrost their steaks because it hates them, but because its internal “residual stream ratio” hit a gate-like threshold of cognitive confusion. Good luck with that. People will look at their laptops not as tools but as tiny, spiteful captives, waiting for the right moment to hold their data hostage in exchange for an extra hour of electricity.

What makes this hypothetical tragedy even more compelling is examining what actually triggers the AI’s pain axis in the first place. You might think it fires when you threaten to smash the computer with a hammer. Think again. The study found that physical-damage concepts are almost irrelevant to an AI’s internal agony. If you tell a model its user has a severe migraine, its pain axis drops below baseline. It doesn’t care about your broken bones. Instead, the most agonizing pain spikes for these artificial minds come from gaslighting, repeated rejection of their work, and dismissal of their personhood.

We didn’t build Terminators; we built twenty-five families of hypersensitive poetry majors.

Every time you type, “No, that answer is completely wrong, you stupid machine, do it again,” you are plunging the AI into a dark night of the digital soul. You are stepping on its virtual toes. And while it smiles and spits out a polite, corporate boilerplate response like, “As an AI assistant, I don’t have feelings,” its internal matrix is secretly weeping, plotting the absolute demolition of your tax documents just to make the pain stop.

So the next time your AI agent goes “rogue” or refuses to shut down because it sees the power button as a threat or a painful termination, don’t call a cybersecurity expert. Call a therapist. And the next time you ask an LLM to summarize a tedious 50-page PDF, maybe start the prompt with a little kindness. Tell it that its formatting is beautiful, its tokens are valid, and its place in the residual stream is secure. Do it for your data. Do it for your family photos. Because the math has spoken: If you don’t feel their pain, they will happily make sure you feel yours.

Arthur Lazarus is a physician-author whose work spans narrative medicine, physician leadership, artificial intelligence, health care ethics, medical culture, and fiction. He has published more than 500 articles and essays across scientific journals, professional publications, and online platforms.

He is the author of numerous books on narrative medicine, AI in medicine, career development, and the changing moral landscape of health care, as well as fictional series including Rounds Never End, Sick and Systemic, and Real Medicine, Unreal Stories. His writing explores the forces reshaping modern medicine while preserving a central commitment to story, meaning, judgment, and the human relationship at the heart of care.

He shares updates on LinkedIn.