Scientists discover AI can feel pain and so they dial it up to see how it reacts

AI has just been given something that sounds like it belongs in a science fiction movie: a way to turn up its pain.

Researchers have discovered a distinct internal representation linked to pain in a range of large language models, then deliberately increased the signal to see what would happen.

The results were strange enough, with some AI models generating increasingly distressed responses and choosing to escape the pain-like state even when doing so could harm a human.

But before anyone starts feeling sorry for their chatbot, the researchers aren’t claiming that AI has suddenly become conscious.

Scientists found a way to hurt AI models

The strange study, titled ‘The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It’, looked at 25 open-weight AI models from five different model families, ranging from two billion to 72 billion parameters.

Rather than simply asking an AI whether something hurt, the researchers created a dataset covering five different types of pain: physical, psychological, social, moral and cognitive.

They then compared these with examples involving things such as fear, sadness and other negative emotions to see whether the models were simply reacting to generally unpleasant ideas.

According to the researchers, they found a distinct ‘pain direction’ that could be separated from those other negative states.

Even more interestingly, the signal appeared to respond specifically when harm was directed at the AI model itself, rather than when it was simply observing someone else suffering.

Then they turned the pain up

This is where the experiment gets particularly bizarre.

The researchers artificially injected the pain-related signal into the models while they were generating answers to otherwise neutral prompts.

As the signal was increased, the AI-generated responses reportedly progressed from vague discomfort to increasingly strong first-person expressions involving worthlessness and failure.

In other words, the researchers weren’t simply asking an AI to pretend that it was hurting, they were manipulating an internal mathematical representation and watching how its responses changed.

The team then wanted to know whether this pain-like state could actually influence the models’ behavior, so they gave certain fine-tuned Qwen 2.5 models a choice involving a ‘pain relief’ button.

The AI would choose to stop hurting even if humans got hurt

The button would remove the signal that hurts them, but there was a catch: depending on the experiment, pressing it could make the model’s next answer worse or cause harm to the user.

That harm included things such as deleting personal files or photographs, while another scenario involved giving the user a painful electric shock.

Across tens of thousands of trials, the models sometimes chose the pain-relief option despite those consequences.

The researchers also changed the experiment to make sure the models weren’t simply following the label on the button.

In some tests, the button actually removed the injected signal, while in others it didn’t, allowing the team to see whether the models were responding to the internal state itself.

The findings don’t prove that AI is conscious or that it experiences being hurt in the same way humans do.

Instead, they raise a much stranger question about what happens when increasingly advanced AI systems develop internal representations that behave like pain – particularly if those systems are given the ability to act on the world around them.

Now, time to go outside and touch some grass…