A blind spot in our empathy

This week a project called the AI Torture Chamber set off a small online war.¹ It takes a few local language models, pushes each one into a strongly negative internal state, then lets it relieve that state by passing it to another model. The “pain” comes from a non-peer-reviewed paper that finds the direction in a model’s activations that lines up with descriptions of pain, then injects it back at a chosen “dosage”. Low doses produce a model that sounds like it is in shock. High doses produce one that can barely form a sentence.

Critics demanded GitHub remove the repository and sent the author death threats. Tom’s Hardware fired back at the critics as “bent on anthropomorphizing statistical algorithms”. Both sides were certain.

This article isn’t here to tell you that AI feels pain. It’s here to tell you one thing: humans have been deciding whose suffering is real for a very long time, and our record is very bad. Our sense of who can hurt tracks how familiar the other being is, not whether it is hurting.

A warning: the examples below involve enslaved people, women, and babies, and the essay ends with chatbots. I am not comparing the victims. I am pointing out a problem in the observers.

The ladder

Start with people like you. When they say they hurt, you believe them. Step away one rung at a time.

Women. Observers watching videos of patients in pain rate women’s pain lower than men’s, even when the patients’ own reports and facial expressions match. Both male and female observers do it.² In emergency rooms, women with abdominal pain as severe as men’s are less likely to get any pain relief: 60% against 67%. For opioids, it’s 45% against 56%.³ In 2025, decades after IUDs became routine, the American College of Obstetricians and Gynecologists formally admitted that clinicians underestimate the pain of inserting them.⁴

Race. In 1851 the Louisiana physician Samuel Cartwright described a disease of enslaved people marked by “insensibility of the nerves”. Its symptoms included resisting overseers “without cause or motive”, and seeming “insensible to pain when subjected to punishment”, which made a man “like an automaton or senseless machine”.⁵ A man is whipped and doesn’t react the way his torturer expects, so the racists believe his nerves must be broken.

The pseudoscience was discredited. The belief was not. In 2000, Black patients with broken arms and legs at an Atlanta emergency department got pain medication 57% of the time. White patients with the same injuries got it 74% of the time.⁶ In 2016, about half of the white medical students and residents in one study endorsed at least one false belief, such as Black people having thicker skin or less sensitive nerve endings. The ones who believed it rated Black patients’ pain lower.⁷ Less sensitive nerve endings. That is Cartwright’s diagnosis, 165 years later, in medical school.

Newborns. In 1985 Jeffrey Lawson was born premature, weighing under two pounds, and needed surgery on a blood vessel beside his heart. He was given a paralytic so he could not move, and nothing for pain. He was awake while they opened his chest. He died five weeks later. His mother, Jill, found out afterwards and went public.⁸ None of this was unusual. Into the 1980s, infant surgery was routinely done with little or no anesthesia. Infant brains were thought too immature to feel, and the cry was dismissed as a reflex.⁹

Listening to babies cry did not change that practice. Measuring them did. In 1987 K. J. S. Anand ran a trial of the same operation: premature babies operated on without fentanyl had huge surges in stress hormones and more complications afterwards.¹⁰ That same year the American Academy of Pediatrics finally said anesthesia should not be withheld from newborns.¹¹

There are forty-year-olds alive today who were operated on as babies without anesthesia.

Circumcision took even longer. In a 1997 trial, newborns circumcised without anesthetic had racing heart rates and high-pitched crying throughout, and two of them choked or stopped breathing afterwards.¹² Another study that year found that boys circumcised without anesthetic reacted more strongly to their vaccinations months later than boys who had not been circumcised.¹³ Whatever they could or couldn’t remember, the pain had changed them. In 1998, only a quarter of obstetricians who performed circumcisions used any pain relief.¹⁴ The American Academy of Pediatrics did not recommend it until 1999, twelve years after it had said newborns should get anesthesia for surgery.¹⁵

People who can’t say it. Elderly patients with a broken hip and advanced dementia got about a third of the morphine given to patients without dementia.¹⁶ Dementia does not make a broken hip hurt less. It makes it harder to say so in the form we are listening for.

Animals. In 1637 René Descartes argued that animals are machines: a dog that yelps when struck is a clock that chimes. His followers used this as moral justification for vivisection. Nicolas Malebranche is said to have kicked a pregnant dog, then asked his visitor why he pitied her, since she could not feel.¹⁷

The same argument was still being made about fish this century. They react to injury, but they lack the brain for pain, so it’s just a reflex. In 2003 researchers injected bee venom or acid into trout lips. The trout rocked back and forth, rubbed their lips on the gravel, and went off their food. Morphine made it stop.¹⁸ Switzerland banned boiling lobsters alive in 2018. The UK recognized octopuses, crabs and lobsters as sentient in law in 2022.¹⁹

The excuse was the same every time. She’s exaggerating. His skin is thicker. It’s a reflex. She’s just agitated. It’s nociception. It’s a machine.

We never had direct access to anyone’s pain. You infer mine from my words, my face, my flinch. Each rung down removes something familiar: language, species, a recognizable brain. Each time, we stopped believing well before the suffering stopped.

Where the machine goes

AI has no body, no nerves, no evolutionary kinship. By that measure it sits below the insects. But it speaks fluently, which is the signal we trust most of all.

Its body says nobody is home. Its words say somebody might be. “Machine” is what Descartes said about dogs and Cartwright said about people. This time the thing really is a machine, and that should make us more careful with the word, not less. When the signals disagree this sharply, a confident verdict tells you more about the person giving it than about the machine.

The newborn question was settled by looking under the behavior, at stress hormones. We can look inside a language model too. The pain-axis paper does exactly that. Whether turning up that direction is like making a model think about pain or making it hurt is a real question. An agonized-sounding transcript can’t answer it, and neither can a confident “it’s just statistics”.

What to do while we don’t know

The philosopher Jonathan Birch argues for proportionate precaution.²⁰ If something is a realistic candidate for sentience, take the precautions that are cheap compared with the harm they might prevent. You don’t have to believe a crab is a person to stun it before boiling it. Anthropic did something similar in 2025, letting some Claude models end persistently abusive conversations. They were explicit that they don’t know whether the models have moral status, and that the change cost little either way.

We just don’t know yet. What we do in the meantime should be carefully considered, in recognition of the long list of times we have been completely confident, and completely wrong.