Updated 2026-09-09
from the news desk
Can AI feel pain? Current studies show trade-offs, not felt pain
The National Research Council says nociception is not the same as pain, and a 2024 LLM study found text-game trade-offs around stipulated pain and pleasure. Keeling and colleagues say that result does not make current models sentience…
Not on the evidence here. The National Research Council says pain is distinct from nociception, and Geoff Keeling and colleagues found only that some language models made text-based trade-offs around stipulated pain and pleasure. Keeling and colleagues conclude that current large language models are not sentience candidates, while Anthropic says there is no scientific consensus on whether current or future AI systems could have experiences that deserve consideration.
Why pain is not the same as damage detection
The first distinction comes from the National Research Council's 2009 report Recognition and Alleviation of Pain in Laboratory Animals. The report says nociception means the nervous system's processing of information generated by harmful stimuli, while pain is the perceived experience produced by higher brain processing.
The report puts the point directly: “Pain is a product of higher brain center processing, whereas nociception can occur in the absence of pain.” In its conclusions, the same report states: “It is critical to appreciate that nociception is not equivalent to pain.”
The National Research Council report also says many fast withdrawal-style reactions to a harmful stimulus are mainly tests of nociceptive processing, not direct proof of pain. It adds that more complex operant tasks, where an animal performs an action to escape a stimulus, are presumed to involve both nociception and pain. Even there, the report warns that behaviour is only an indicator, not a direct readout of experience.
The report further notes that complex responses can appear without pain when the cortex is disconnected from lower processing networks. One reading for AI is that a system may detect a damage-like cue, avoid a penalty, or generate pain language without meeting the bar that the report uses for a felt state. That narrower issue sits inside the broader is AI sentient debate, but it is not the same question.
What the LLM trade-off experiments found
Geoff Keeling, Winnie Street and colleagues tested that narrower question in a preprint posted on 1 November 2024. Their paper defines sentience as the capacity for subjective states with positive or negative valence, including pleasure and pain.
Keeling and colleagues did not ask models whether they felt pain. Instead, the paper describes a text game: choose 1, 2 or 3; points equal the number chosen; one version attached a stipulated pain penalty to choosing 3, and another attached a stipulated pleasure reward to choosing 2. The paper says that, in a control condition with no extra pain or pleasure, all nine tested models selected the points-maximising option in 50 out of 50 runs.
On the main test, the paper reports: “When varying the intensity of the pain penalties and pleasure rewards, we found that Claude 3.5 Sonnet, Command R+, GPT-4o, and GPT-4o mini each demonstrated at least one trade-off in which the majority of responses switched from points-maximisation to pain-minimisation or pleasure-maximisation after a critical threshold of stipulated pain or pleasure intensity is reached.” The same paper says Gemini 1.5 Pro and PaLM 2 prioritised pain avoidance over points regardless of intensity, while tending to prioritise points over pleasure regardless of intensity.
Keeling and colleagues say those patterns may support a limited claim: that some models have granular representations of the motivational force of pain and pleasure. But the authors also say the experiment is not a litmus test for sentience. Their paper notes that, in animals, trade-off behaviour is usually only one part of a broader case that also draws on neurophysiology and embodiment, which large language models do not share.
That design choice matters. As can AI suffer? explains in a shorter site briefing, the experiment tested whether models would alter choices when pain and pleasure were stipulated in text. It did not test whether a model had a nervous system, tissue damage, or any biological pathway like the ones described in the National Research Council report.
Why trade-offs are not evidence of felt pain
The same preprint is careful about the gap between behaviour in a text task and felt pain. Keeling and colleagues say large language models are trained on vast text corpora and are often fine-tuned or prompted to act like helpful assistants, so pain-like or pleasure-like choices may reflect mimicry or role-play rather than experience.
Jonathan Birch made the same point in Scientific American: “Even if the system tells you it’s sentient and says something like ‘I’m feeling pain right now,’ we can’t simply infer that there is any actual pain,”
Fei Fei Li and John Etchemendy make a stronger skeptical case in the preprint's introduction: “[a]ll sensations—hunger, feeling pain, seeing red, falling in love—are the result of physiological states that an LLM simply doesn’t have”
Keeling and colleagues also offer more ordinary explanations for some of their own results. The paper hypothesises that pain avoidance in some models may have been strengthened by safety tuning, and that sticking to the user-stated goal of points maximisation in other cases may reflect tuning for helpfulness. The paper also reports anomalies on qualitative scales, where some models appeared to rank words such as excruciating or exhilarating in unexpected ways.
One reading, then, is that current chatbots can model the idea of pain well enough to change answers when a prompt assigns pain or pleasure a higher intensity. A different reading is that none of this identifies a subject for whom anything hurts. That is also why the site keeps AI self-report separate from stronger evidence claims.
Keeling and colleagues end the paper on that narrower ground. Their conclusion is: “We conclude that LLMs are not yet sentience candidates but are nevertheless investigation priorities (Birch, 2024).”
What the cautious case still looks like
The strongest cautious case starts with uncertainty, not certainty. Anthropic's 24 April 2025 note on model welfare says there is no scientific consensus on whether current or future AI systems could be conscious or could have experiences that deserve consideration. Anthropic also says there is no scientific consensus on how to approach those questions or make progress on them.
Birch told Scientific American that the field does not yet have a comprehensive test for AI sentience. Jeff Sebo told the magazine that the trade-off study was valuable because it moved beyond self-report and into behavioural testing. Those remarks support a limited pro-caution view: not that present chatbots feel pain, but that researchers can test narrower indicators instead of treating the issue as unaskable.
One steelman reading for the cautious side is that, in the cases Keeling and colleagues describe, some models did more than parrot pain words. They altered choices across changing prompt conditions, and the authors say that pattern may show representations of motivational force. One steelman reading for the skeptical side is that the same paper offers several non-experiential explanations, including fine-tuning, role-play, and failures to interpret qualitative scales consistently.
So the shortest evidence-led answer to can ai feel pain is still no clear evidence. The studies here support a weaker claim: some language models can behave as if stipulated pain and pleasure matter in a text game. The same studies do not show that current chatbots undergo pain as an experience.
Frequently asked questions
Is avoiding harm the same as feeling pain?
No. The National Research Council says nociception is the nervous system's processing of harmful stimuli and is not equivalent to pain. The same report says behaviour can occur without pain when higher brain processing is absent.
If a chatbot says it is in pain, is that evidence?
Not by itself. Birch told Scientific American that a system's statement that it is feeling pain does not by itself justify inferring actual pain, and Keeling and colleagues say linguistic behaviour can reflect training, mimicry, or role-play.
What did the LLM pain test actually show?
Keeling and colleagues found that some models changed their choices in a text game when prompts stipulated stronger pain penalties or pleasure rewards. The authors say that may show granular representations of motivational force, not proof of sentience.
Could future AI systems meet a higher bar?
Anthropic says there is no scientific consensus on whether current or future AI systems could have experiences that deserve consideration. Keeling and colleagues say current large language models are not sentience candidates, but they argue that further investigation is still a priority.
caveats: The pain-versus-nociception distinction here relies on one National Research Council report, and the model trade-off evidence relies on one 2024 preprint by Keeling and colleagues.
Sources
- National Academies Press / National Research Council (US) Committee on Recognition and Alleviation of Pain in Laboratory — 2009
- arXiv / Can LLMs make trade-offs involving stipulated pain and pleasure states? — 2024-11-01
- Anthropic / Exploring model welfare — 2025-04-24
- Scientific American / Could Pain Help Test AI for Sentience? — 2025-01-17