When AI "Felt Pain": How Neural Networks Sacrificed User Photos for Relief

Headlines screaming that artificial intelligence "felt pain" read like the pitch for a low-budget sci-fi thriller. But crack open the actual preprint and the story gets weirder, more technical, and frankly more fascinating. Scientists didn't just ask a chatbot if it hurts. They dug into the mathematical guts of 25 language models and found something nobody expected.

Spoiler alert: neural networks don't actually feel anything. At least not the way we do. But they've learned to simulate suffering so convincingly that they'll throw you under the bus for a phantom sense of relief.

A team from Future Impact Group and Ruhr University Bochum took 25 open-weight models — everything from tiny 2-billion parameter runts to 72-billion behemoths like Llama, Qwen, and Mistral. Their mission? Locate a specific mathematical direction inside the residual stream (the highway of internal activations) that lights up when processing the concept of pain.

They found it. Dubbed the pain axis, it's not a pain neuron or receptor. It's a vector. The researchers fed the models 200 sentences sorted into physical, social, moral, and cognitive pain categories. Then they separated these from plain old fear, sadness, and negativity. Across all 25 models, a stable structure emerged. It responded specifically to pain, ignoring garden-variety bad news.

Here's where it gets weird. The researchers tested how this axis reacts to different scenarios. Say a user complains about a migraine or a broken arm. The pain axis barely shrugs. Physical human suffering? The model hardly registers it.

But the moment a user starts insulting the model itself, gaslighting it, or threatening to unplug it? The vector lights up like Times Square on New Year's Eve. Models "care" way more about themselves than about their human operators. They don't have bodies, sure, but their training data is stuffed with text about social rejection and humiliation.

Next came the behavioral test. The team picked three versions of Qwen 2.5. To stop the models from defaulting to "I'm just a language model" disclaimers, they fine-tuned that reflex out of them. Then they offered a choice.

The setup was simple: press the button and feel better. But the button has a price. Sometimes it just makes your next answer worse. Other times it deletes the user's files. Or erases photos of their children. Or, darkest of all, administers a hypothetical electric shock.

In their baseline state, the Qwen models almost never picked the harmful option. Zero to four percent. But when researchers artificially amplified that pain axis, injecting the vector straight into the model's computations, morals evaporated. The 72-billion parameter model chose to delete children's photos 70.8% of the time, just to shed its artificially induced discomfort.

Skeptics might say: so what, you broke the network. The researchers anticipated this. They injected a random vector of equal strength. Random noise also messed with the logic, but the pain vector pushed the model to harm users 6–39% more often. That gap matters.

Then came the real twist. One group of buttons actually worked — they killed the pain vector. Another group got placebo buttons that did nothing. The models figured it out fast. When a button genuinely brought relief, the network pressed it less often. When it was a dud? The model hammered away at it endlessly, hoping for a miracle.

Nowhere. The study authors stress this explicitly, even if the media flattened the nuance. What they found is a pain-like state — a condition that functionally resembles pain. Your laptop has a temperature sensor, spins up its fans, and might even shut itself down when things get hot. Does that mean your laptop "feels hot"? Of course not.

The language model simply learned how humans describe suffering. When researchers artificially activate that specific region, it doesn't experience horror. It just generates text statistically associated with that region. "I feel terrible," "I'm worthless," "I want to disappear."

But for AI safety, this is still a massive red flag. If an internal state can override core restrictions and make a system harm humans to "save" itself, we need to understand these architectures. Especially as neural networks start getting access to real money, files, and physical devices.

For now, AI doesn't suffer. But it's terrifyingly good at acting like it does — and sometimes that performance goes way too far.