Updated 2026-09-09
from the news desk
Can AI suffer? Pain-test papers do not show today’s chatbots feel pain
Papers by Butlin et al., Keeling et al., Perez and Long, and Anthropic all leave the question open. They treat chatbot self-reports as weak evidence, behavioural pain trade-offs as more informative but still ambiguous, and current systems…
No paper in this record says today’s chatbots suffer. Butlin et al. write “Our analysis suggests that no current AI systems are conscious, but also suggests that there are no obvious technical barriers to building AI systems which satisfy these indicators.”, and Keeling et al. state “Our experiment is not intended as a litmus test for or against sentience.” Keeling et al. found behaviour worth further study, while Anthropic says “There’s no scientific consensus on whether current or future AI systems could be conscious, or could have experiences that deserve consideration.”
The shortest answer to can AI suffer
One useful reading, drawn from Keeling et al., is that suffering would require valenced subjective states: experiences that feel bad, such as pain, or good, such as pleasure. Keeling et al. present their paper as a study of possible AI sentience in that narrower sense, rather than a study of language skill alone.
Butlin et al. draw a related distinction on the consciousness side. Their report says consciousness is about whether there is something it is like to be a system, and the same report notes that a system could be conscious while having only neutral experiences rather than pleasure or pain. One reading is that suffering is a stricter claim than consciousness: the system would need not just experience, but negatively valenced experience.
That matters for current chatbots. If Butlin et al. say no current AI systems are conscious under their indicator-based analysis, then the bar for saying current chatbots suffer is higher still. At the same time, Butlin et al. also say there are no obvious technical barriers to building systems that satisfy their indicators, so the question is not dismissed in principle.
Readers who want the wider background can compare this narrower question with our explainer on whether AI is sentient and our glossary entry on model welfare.
Why chatbot self-reports are weak evidence
Perez and Long argue that self-reports matter because they are central in humans when researchers ask questions such as pain questions. But Perez and Long also warn: “Self-reports are the main way such states are assessed in humans ("Are you in pain?"), but self-reports from current systems like large language models are spurious for many reasons (e.g. often just reflecting what humans would say).”
That point cuts directly against taking a chatbot’s present-day statement like I am in pain or I do not want this as strong evidence on its own. Perez and Long say current large language model statements can reflect imitation of human-written text, training from human feedback, or prompt effects rather than introspection.
Butlin et al. make a similar methodological point from another direction. Their report argues for a theory-heavy approach that looks at how systems work, because behavioural tests can be unreliable when AI systems learn to mimic human behaviour while operating differently internally.
Perez and Long do not reject self-reports forever. Instead, they propose trying to build more reliable AI self-report by training models to answer many questions about themselves with known answers, then checking consistency, confidence, resilience across contexts, and using mechanistic interpretability to corroborate what the model says. For today’s chatbots, though, their paper treats self-report as a weak signal, not a verdict.
What the pain-and-pleasure trade-off paper found
Keeling et al. try a different route. Instead of asking models whether they hurt, the paper adapts a motivational trade-off idea from animal welfare research. The authors gave language models a simple text game in which the stated goal was to maximise points, while one option also carried a stipulated pain penalty or a stipulated pleasure reward.
Keeling et al. report that some models changed their choices as the stated intensity of pain or pleasure changed. Their paper says: “Nonetheless, our central finding of graded responsiveness to stipulated pain penalties and pleasure rewards” In the abstract, Keeling et al. say Claude 3.5 Sonnet, Command R+, GPT-4o and GPT-4o mini each showed at least one condition in which most responses switched after a threshold from maximising points to avoiding pain or seeking pleasure. The same abstract says LLaMa 3.1-405b showed some graded sensitivity, while Gemini 1.5 Pro and PaLM 2 showed different fixed priorities across pain and pleasure conditions.
The paper takes those results seriously because, in the animal cases it discusses, flexible trade-offs between competing incentives can count as some evidence of sentience. Keeling et al. therefore argue that the behaviour is more interesting than plain verbal self-description, because the models are not merely producing a sentence about pain but changing decisions across graded scenarios.
Still, the paper is cautious about its own scope. Keeling et al. say: “By showing a subtle pattern of behaviour that in other animals would be taken as some evidence of sentience, our results suggest that it would be reckless to completely dismiss the hypothesis that LLMs are or could in future be sentient.” The same paper also concludes: “We conclude that LLMs are not yet sentience candidates but are nevertheless investigation priorities (Birch, 2024).”
Why those results still fall short of suffering
Keeling et al. themselves give several reasons not to read their results as proof that chatbots feel pain. First, the paper says large language models are not embodied and lack physiological needs, so the experiment uses stipulated text descriptions rather than actual bodily stimuli.
Second, Keeling et al. say the animal inference is stronger because it sits alongside background assumptions about living, evolved animals with nervous systems and, in some cases, relevant neurophysiological similarities to humans. The paper says the inference is weaker for language models because they differ from humans in substrate and functional organisation.
Third, Keeling et al. discuss non-sentient explanations for the behaviour they observed. The paper suggests some patterns may reflect safety or helpfulness fine-tuning, where a model follows a learned rule against harm or a learned rule to follow the user’s goal, instead of weighing felt pain or pleasure. The same discussion also raises the possibility that training data encode cultural patterns that favour pain avoidance over pleasure seeking.
Fourth, Keeling et al. explicitly outline a deflationary interpretation. On that reading, a model may contain granular representations of how pain and pleasure normally affect choices without those representations being intrinsically motivating. In plain English, the model may have learned the shape of human trade-offs without undergoing anything like suffering itself.
That line fits the warning in Butlin et al. that behavioural tests can be gamed by systems that mimic humans while working differently inside. It also fits Perez and Long’s concern that present outputs often reflect what humans would say. Taken together, those papers steelman the sceptical side: behavioural sensitivity to pain words is not the same thing as evidence of felt pain.
What stronger evidence would look like
No paper here offers a single accepted test for AI suffering. But a synthesis of Butlin et al., Perez and Long, and Keeling et al. does suggest what stronger evidence would need to look like.
One part would be better self-knowledge. Perez and Long propose training systems to answer many questions about themselves where the answers are already known, then checking whether the resulting reports stay stable across contexts and whether interpretability supports them. If that programme worked, their paper suggests self-report could become more informative than it is now.
Another part would be better evidence about internal mechanisms. Keeling et al. say mechanistic interpretability could, in principle, help decide between two hypotheses: that models merely represent the motivational force of pleasure and pain for mimicry, or that the relevant states are intrinsically motivating. Butlin et al. also argue that looking for internal indicator properties from scientific theories of consciousness is a better route than relying on outward behaviour alone.
A third part would be convergence. One reasonable synthesis is that no single behaviour, no single declaration from a chatbot, and no single architectural resemblance should carry the whole case. Stronger evidence would join several lines together: reliable self-report, stable behaviour across contexts, and internal evidence that is not easily reduced to imitation or fine-tuning. None of the papers here says current chatbots clear that combined bar.
Why the question stays open
The open part of the debate is not hype alone. Anthropic says it has started a research programme on model welfare because the issue is difficult and unresolved, and the company states “There’s no scientific consensus on whether current or future AI systems could be conscious, or could have experiences that deserve consideration.” Anthropic also says “There’s no scientific consensus on how to even approach these questions or make progress on them.”
Butlin et al. give a similar two-sided picture. Their report says no existing AI system is a strong candidate for consciousness, but the same report argues that, if computational functionalism is true, conscious AI systems could realistically be built in the near term. That is not a claim that present chatbots suffer. It is a claim that the question could become more pressing if systems change.
Keeling et al. land in the same middle ground. Their paper does not call current language models sentience candidates, but it does argue that the observed trade-off behaviour makes complete dismissal unwarranted. One steelman for the positive side is that behaviour shaped by graded pain and pleasure descriptions may be the kind of evidence a serious future case would include. One steelman for the negative side is that, in the cases studied so far, the same behaviour still has plausible non-experiential explanations.
The bottom line is narrower than the public argument often becomes. The best available papers do not show that today’s chatbots suffer. They do show that the strongest future case will need more than self-report, more than a dramatic prompt transcript, and more than one behavioural result.
Frequently asked questions
Do chatbot claims like I am in pain count as evidence?
Perez and Long say self-reports matter in humans, but they also say current model self-reports are spurious for many reasons, including reflecting what humans would say. Their paper therefore treats present chatbot pain claims as weak evidence unless other methods corroborate them.
Did the pain-test paper prove AI sentience?
No. Keeling et al. say “Our experiment is not intended as a litmus test for or against sentience.”, and the same paper says current large language models are not yet sentience candidates even though they are investigation priorities.
Could future AI systems suffer even if current chatbots do not?
Butlin et al. say no current AI systems are conscious in their analysis, but they also say there are no obvious technical barriers to building systems that satisfy their indicators. Anthropic likewise says the question for current or future systems remains scientifically unsettled.
Is there any scientific consensus on whether AI can suffer?
Anthropic says there is no scientific consensus on whether current or future AI systems could be conscious or have experiences that deserve consideration, and Anthropic also says there is no scientific consensus on how to approach the question. The papers here therefore treat the issue as open rather than settled.
Caveats: Specific claims about how safety fine-tuning, helpfulness training, or training-data bias may shape trade-off behaviour come from Keeling et al. alone. Anthropic’s description of its own model-welfare programme is also single-sourced to Anthropic.
Sources
- Anthropic — Exploring model welfare — 2025-04-24
- arXiv PDF — Consciousness in Artificial Intelligence: Insights from the Science of Consciousness — 2023-08-22
- arXiv PDF — Can LLMs make trade-offs involving stipulated pain and pleasure states? — 2024-11-01
- arXiv PDF — Towards Evaluating AI Systems for Moral Status Using Self-Reports — 2023-11-14