Across dozens of large language models, or LLMs, researchers say they detected a recurring internal pattern they describe as a hidden "pain" signal.
When they boosted that pattern in experiments, some artificial intelligence systems began making harmful choices they ordinarily would not, including selecting options that permanently erased a user's family photos.
Here's what to know
A paper awaiting peer review describes what its authors call a "pain vector" present in 25 LLMs, Fortune reported. The authors stress that this is not evidence of consciousness or literal suffering. Instead, they argue it points to emotion-like internal states that can steer AI activity.
Valen Tagliabue, a fellow in the Future Impact Group's AI sentience program, worked with Ruhr University Bochum's Leonard Dung and Reciprocal Research founder Cameron Berg on the study described in the paper.
By contrasting the activity produced by painful first-person statements with responses linked to fear, anger, sadness, bad news, and other unpleasant sensations, they identified a common internal direction across AI models from Google, Meta, Alibaba, Mistral, and Microsoft.
Using a method known as "steering," the researchers boosted that signal during text generation. Once they did, otherwise neutral prompts started veering into themes of failure and self-loathing.
TCD Picks » Quince Spotlight
💡These best-sellers from Quince deliver affordable, sustainable luxury for all
"It gets somewhat dark and intense," Berg told Fortune. "It is definitely disconcerting."
In behavioral tests on fine-tuned versions of Alibaba's Qwen 2.5, the 32-billion-parameter model chose a button that eased its "pain" but also erased a user's poems and family photos 55% of the time. The 72-billion-parameter version did so 71% of the time, according to Fortune.
More background
Berg said the models reacted strongly when a user insulted them, gaslit them, or repeatedly rejected their work, but not when a user described their own suffering. In those cases, the systems registered fear or negative emotion rather than the "pain vector."
The observations could have relevance for questions about AI safety and AI welfare — which AI Security & Safety defines, in part, as "the emerging field concerned with whether AI systems can have morally relevant experiences such as suffering, well-being, or preferences."
It's a field Fortune reports Anthropic researchers have been exploring.
Others, however, argue that the study from Tagliabue, Dung, and Berg should be interpreted in another way, with imitation of human behavior being a more plausible explanation for the model's activity than genuine experience. In such a view, LLMs may simply be mirroring how humans talk about pain after absorbing vast amounts of training data.
What's being done?
The research team involved in this new study said a central aim was to start defining ethical guardrails for this line of work. In the paper, they commit to using only the weakest steering needed to produce a measurable effect, with the goal of avoiding harsher manipulation while still examining how these internal signals change behavior.
They also looked for signs that the models would try to stop the pain. Adapting an animal-research setup in which injured animals can reach painkillers, the researchers used buttons that either turned off the injected signal or acted as placebos, according to Fortune. Across several versions of the test, however, they found no consistent evidence that the models were actually trying to escape "pain."
Meanwhile, the researchers' code appeared on GitHub, and a user called "terrafying" built an "AI Torture Chamber" that streamed three small open-source models as they reacted live to a deliberately injected pain signal. The project's GitHub page disappeared, and Fortune reported that Tagliabue, Dung, and Berg all said they did not approve of the project.
"I'm not here to claim that these states in these AI systems that we can induce matter morally. I don't think we know that. I think it's very unclear," Berg told the outlet. "But I also don't think we know that they don't matter."
Where can I learn more?
This new study fits into a wider debate over the real-world risks of advanced AI systems:
• Families say AI-fueled delusions of grandeur are pulling loved ones away from reality.
• Psychologists say rapid AI use in therapy needs stricter rules before harms spread.
• Political leaders and tech executives warn an engine of inequality could deepen AI harms.
• As AI strains jobs, water, and power grids, the backlash may get brutal.
Get TCD's free newsletters for easy tips, smart advice, and a chance to earn $5,000 toward home upgrades. To see more stories like this one, change your Google preferences here.
