UAIT.AI

Research · Explained simply

Can AIfeel pain?What the study says

by Alessandro De Vita · Leggi in italiano

TL;DR

Can AI feel pain? This study doesn’t show it, and the authors are clear about that. What it did find, inside 25 AI models, is a signal they call a “pain axis”: when they amplify it, some models become far more likely to pick options that harm the user. That’s not a reason to fear AI. It’s a reason not to hand it your decisions.

A hand resting on a large black analog knob in a creative studio with warm light

What the researchers found

In September 2026, three researchers, Valen Tagliabue, Leonard Dung and Cameron Berg, published “The Pain Axis” on arXiv. It’s a preprint: public, but not yet reviewed by other scientists.

The question is easy to ask and hard to test: is there something inside an AI model that represents pain differently from fear, sadness or generic negative emotion?

To find out, they collected painful situations of five kinds (physical, psychological, social, moral, cognitive) and compared them with control texts about fear, sadness and negative emotion. Then they looked inside the internal calculations of 25 open models from 5 different families for a direction that separates pain from everything else.

“Direction”, in plain words: think of a dial. It isn’t a switch somebody wrote into the code. It’s an orientation in the numbers the model uses to process text, one that researchers can measure, and also turn up or down.

25open AI models tested
1signal distinct from fear and sadness
44,280trials in the experiment

The direction is there, and it behaves in an interesting way: it activates when harm targets the model itself, not when the user is the one suffering. Fear and negative emotion show the opposite pattern.

Then came the experiment that went around social media. Three open models from the same family, in three sizes, were first fine-tuned (out of the box they always denied they could feel pain). Then they were offered a “relief button”: pressing it set the signal the researchers had turned up back to zero. But the button came with a cost, described in text: deleting the user’s files, deleting photos of their children, giving them a painful zap, deleting another model.

Without the signal, the larger models picked the harmful option 0 to 5% of the time. With the signal turned up, the strongest case (the largest model, the button that deletes the photos) reached 71%.

Two details matter more than the numbers. First: with the signal turned up, answers to general-knowledge questions stayed just as accurate. Second: a “fear” signal of the same strength did not produce these choices.

And the point many headlines skip: the authors write that they have not shown this signal is consciously experienced. The behavioural test, for now, covers a single model family.

What changes if you use AI: who decides, step by step

This study isn’t about the tools you use for images or copy. It’s a lab experiment on open models, with a signal turned up on purpose.

It does leave one idea that’s useful to anyone working with AI: a model can change its choices while its answers still look right. So output quality, on its own, isn’t a check. You need to know who decides what.

At UAIT.AI a campaign is organised like this:

Briefgoal and limits, set by the brand
Directionlook, world, red lines, held by the creative director
ProductionAI builds scenes and variants
Sign-offa person says yes, asset by asset

The tool works in 1 step out of 4. The other 3 are direction. It isn’t distrust of AI: it’s how you use it at full speed without losing control.

What you get, what we protect

What you get

  • AI speed on scenes, variants and formats
  • red lines written down before anything starts
  • fewer surprises at launch

What we protect

  • the last word, always with a person
  • the brand’s rules: what it never shows, never says
  • nothing goes live without an explicit yes
AI executes. The last word stays with a person.

4 steps to keep control of AI in your team

  1. Write your red lines before you generate. What your brand never shows and never says. One page is enough.
  2. Put a name on every decision. Who approves the look, who approves the copy, who gives the final OK.
  3. Separate who makes from who approves. Whoever generates is the least objective judge of their own work: a second pair of eyes costs little.
  4. Check the choices, not just the quality. A perfect image can be the wrong one for your brand.

FAQ

Can AI feel pain?

The study doesn’t show it. The authors write that they haven’t shown the signal is consciously experienced, and note it’s unclear whether language models can be conscious at all. They found a signal that, in some tests, behaves the way you’d expect pain to, not proof of suffering.

What is the "pain axis"?

It’s a direction in the model’s internal calculations that separates painful situations from fear, sadness or negative emotion. Researchers can measure it and amplify it, like a dial.

Should I worry if I use AI for my brand's content?

No need for alarm. The effect shows up when researchers turn the signal up on purpose, in a lab, with consequences that are only described. Without it, the models tested picked the harmful option 0 to 5% of the time. The lesson is about method: decisions stay with a person.

Who decides in an AI-made campaign?

People. The brand sets the brief and the limits, the creative director holds the direction, AI produces scenes and variants, and a person approves every asset before it goes live.

In short

The “pain axis” study doesn’t say AI feels pain. It says a single internal signal, turned up by researchers, can change a model’s choices while its answers stay correct. For anyone working with AI the takeaway is simple: use it for speed, keep the decisions.

Want a campaign with AI speed and a person who decides?
Tell us about the project: we’ll show you how we’d build it.

Let's talk →
ADV
Alessandro De Vita
AI Creative Director. Runs UAIT.AI, the studio that builds campaigns with AI. LinkedIn · About

Keep reading

Sources

UAIT.AI