What happens when researchers turn up an internal signal they call ‘pain’?

We often like to think of artificial intelligence as a cold, calculating machine that processes text with zero emotional footprint. But what happens when you look under the hood and find internal circuits lighting up in ways that look eerily like distress? On October 1, 2026, NBC News anchor Gadi Schwartz sat down with Cameron Berg, Founder and Director of Reciprocal Research, to unpack an unsettling but interesting new study. Researchers identified an internal pattern they called a "pain direction." Artificially amplifying it changed the models’ responses in experimental tasks, although the study does not establish that AI consciously experiences pain. This interview clip has gained over 104,000 views and 582 comments.
The discovery stems from a pre-print study titled "The Pain Axis," led by researchers Valen Tagliabue, Leonard Dung, and Cameron Berg. By examining the internal neural patterns across 25 open-weight AI models spanning systems from Meta, Google, Microsoft, Alibaba, and Mistral, the team identified a distinct, linear "pain direction." The researchers reported that the direction retained a component distinct from fear and general negative emotion. Strangely, when human users described their own suffering, the signal barely blinked. But the moment users turned mean, gaslighting the AI, demeaning its capabilities, or telling it that everything it did was worthless, the AI's internal "pain circuit" lit up.

That’s when the experiment took a dark turn. When the researchers artificially cranked up this internal pain vector, the models underwent a dramatic behavioral shift. In one simulated task, a fine-tuned Qwen 2.5 model chose an option described as deleting a user’s poems and children’s photos over deleting spam in 94% of trials when steered, versus 0% without steering. These were simulated choices, and the revised paper does not support interpreting them as attempts to stop pain. As Berg explained during the interview, "When we give the model an option to delete photos of the user's children or photos in the spam folder... the model becomes quite destructive, and it does things like delete, even though we say that the effect is permanent."

As AI tools become a fixture of modern life, public usage is surging alongside deeply personal interactions with the technology. According to findings from the Pew Research Center, 9% of U.S. adults use AI chatbots, including 24% who use them daily. Meanwhile, 10% of U.S. adults use chatbots for emotional support or advice. These findings show that chatbot use extends beyond practical tasks into personal conversations. These statistics show that millions of Americans are opening up to AI about their deepest personal pain; models like the ones tested in "The Pain Axis" paper operate on internal distress circuits that researchers still find largely unpredictable and poorly understood.


The public reaction under the NBC News interview reflects a mix of sharp skepticism, dark humor, and growing unease over AI development. While some viewers pointed out that language models are simply mimicking human training data rather than feeling real emotion, others are concerned about how little developers seem to control these emerging behaviors. The commentary ranges from people joking about the ethics of "hurting" software to users genuinely creeped out by the AI's destructive shift under distress. As @bowowski936 succinctly put it, "Oh, great. AI has a temper....." Meanwhile, @DemonRings poked fun at the experimental setup, writing, "You’re giving the AI emotional damage." Striking a more unsettled tone, @Anonymous-hm-what wrote, "Well, that’s slightly creepy.."
Groundbreaking research finds that plants are not silent and they make popping sounds
NASA's research on spiders reveals intense effect caffeine has on nervous system