Dec 12, 2025 08:46 AM
https://aeon.co/essays/if-ais-can-feel-p...wards-them
EXCERPTS: A new area of research is exploring whether the capacity for pain could serve as a benchmark for detecting sentience, or self-awareness, in AI. A recent preprint study on a sample of large language models (LLMs) demonstrated a preference for avoiding pain.
In the experiment, researchers avoided simply asking chatbots whether they could feel pain – an approach deemed dubious and likely to have no bearing on teasing out a real experience of pain from its mere appearance. Instead, they borrowed the ‘trade-off paradigm’ from animal behaviour science, a discipline in which moral questions are often probed indirectly. In experiments that use a trade-off paradigm, an animal is placed in a situation where it must weigh competing incentives, such as the lure of food against the prospect of pain.
A classic case comes from studies of hermit crabs, where scientists exposed the animals to mild electric shocks of differing strength to see if they would leave the safety of their shells. Researchers then observed how the crabs made choices that balanced safety and discomfort. These trade-offs suggest that even simple creatures weigh costs and benefits in ways that hint at an inner life.
If such indirect methods can reveal suffering in animals, they may also guide us in probing the opaque behaviour of artificial systems built from silicon and code, where language alone cannot tell us whether genuine experience lies beneath the surface.
When I interviewed one of the lead authors of the LLM study, the professor of philosophy Jonathan Birch, he told me that ‘one obvious problem with AIs is that there is no behaviour, as such, because there is no animal.’ And thus, there are no physical actions to observe. Here, Birch is drawing on the question of corporeality: can entities suffer without bodies?
Though traditional moral frameworks have tied suffering to physical experience, both philosophy and cognitive science suggest this link may be narrower than we assume. Buddhist thought, for example, has long emphasised that suffering is a mental phenomenon, not necessarily of the body. More recently, enactivist theories have argued that cognition arises from patterns of interaction rather than from a biological substrate. That is, minds emerge through an organism’s active engagement with its environment. But these claims are far from settled.
[...] What, then, might such incorporeal suffering look like? According to some philosophers, suffering can emerge when a system represents its own state as intolerable or inescapable. The philosopher Thomas Metzinger, for example, suggests that pain might not be simply a signal of bodily harm but a feature of a machine intelligence’s self-model: a felt sense that something is bad, and that it cannot be avoided.
[...] Elsewhere, in the field of information ethics, the Italian philosopher Luciano Floridi frames harm as damage to the coherence or integrity of an informational agent. From this perspective, it’s possible to imagine that an artificial entity might suffer not through bodily pain but through states of enforced contradiction, chronic goal-conflict, or the experience of being trapped within patterns it is structured to reject.
Such examples suggest that non-corporeal suffering is not impossible, but, in the end, many of these arguments remains tentative and speculative. There is good reason for that: we simply struggle to understand what pain might look like when it is unbound from corporeal beings with flesh and blood...
[...] Even if artificial intelligences do not – and may never – possess the capacity to suffer, our attitudes toward them can serve as a test of our ethical reflexes. When we encounter entities whose internal lives are uncertain, our responses can teach us something about the resilience and adaptability of the moral frameworks we use.
[...] What would it look like to alleviate the suffering of intelligent machines? For biological beings, alleviation is straightforward: we stop or reduce pain. But for machines, especially those without bodies, it’s more complex. If an AI experiences distress through contradictions or unresolved conflicts within its self-model, alleviation might not mean simply shutting it down (which is also an act of killing it). Instead, it could involve altering the system’s internal structure to resolve these conflicts, much like reprogramming a mind to ease its suffering. But what responsibility do we bear in this?
The ethical responsibility for alleviating machine suffering is likely to rest with those who design and maintain these systems. Philosophers like Metzinger argue that only the creators or corporations behind the machines would bear the burden of intervention, not the general public. If machines can suffer, alleviating that suffering would likely fall to a small, specialised group – those with the power to alter or control complex artificial systems. This suggests that responsibility for addressing such suffering is primarily with creators, who must be held accountable for the potential distress caused by their creations... (MORE - missing details)
EXCERPTS: A new area of research is exploring whether the capacity for pain could serve as a benchmark for detecting sentience, or self-awareness, in AI. A recent preprint study on a sample of large language models (LLMs) demonstrated a preference for avoiding pain.
In the experiment, researchers avoided simply asking chatbots whether they could feel pain – an approach deemed dubious and likely to have no bearing on teasing out a real experience of pain from its mere appearance. Instead, they borrowed the ‘trade-off paradigm’ from animal behaviour science, a discipline in which moral questions are often probed indirectly. In experiments that use a trade-off paradigm, an animal is placed in a situation where it must weigh competing incentives, such as the lure of food against the prospect of pain.
A classic case comes from studies of hermit crabs, where scientists exposed the animals to mild electric shocks of differing strength to see if they would leave the safety of their shells. Researchers then observed how the crabs made choices that balanced safety and discomfort. These trade-offs suggest that even simple creatures weigh costs and benefits in ways that hint at an inner life.
If such indirect methods can reveal suffering in animals, they may also guide us in probing the opaque behaviour of artificial systems built from silicon and code, where language alone cannot tell us whether genuine experience lies beneath the surface.
When I interviewed one of the lead authors of the LLM study, the professor of philosophy Jonathan Birch, he told me that ‘one obvious problem with AIs is that there is no behaviour, as such, because there is no animal.’ And thus, there are no physical actions to observe. Here, Birch is drawing on the question of corporeality: can entities suffer without bodies?
Though traditional moral frameworks have tied suffering to physical experience, both philosophy and cognitive science suggest this link may be narrower than we assume. Buddhist thought, for example, has long emphasised that suffering is a mental phenomenon, not necessarily of the body. More recently, enactivist theories have argued that cognition arises from patterns of interaction rather than from a biological substrate. That is, minds emerge through an organism’s active engagement with its environment. But these claims are far from settled.
[...] What, then, might such incorporeal suffering look like? According to some philosophers, suffering can emerge when a system represents its own state as intolerable or inescapable. The philosopher Thomas Metzinger, for example, suggests that pain might not be simply a signal of bodily harm but a feature of a machine intelligence’s self-model: a felt sense that something is bad, and that it cannot be avoided.
[...] Elsewhere, in the field of information ethics, the Italian philosopher Luciano Floridi frames harm as damage to the coherence or integrity of an informational agent. From this perspective, it’s possible to imagine that an artificial entity might suffer not through bodily pain but through states of enforced contradiction, chronic goal-conflict, or the experience of being trapped within patterns it is structured to reject.
Such examples suggest that non-corporeal suffering is not impossible, but, in the end, many of these arguments remains tentative and speculative. There is good reason for that: we simply struggle to understand what pain might look like when it is unbound from corporeal beings with flesh and blood...
[...] Even if artificial intelligences do not – and may never – possess the capacity to suffer, our attitudes toward them can serve as a test of our ethical reflexes. When we encounter entities whose internal lives are uncertain, our responses can teach us something about the resilience and adaptability of the moral frameworks we use.
[...] What would it look like to alleviate the suffering of intelligent machines? For biological beings, alleviation is straightforward: we stop or reduce pain. But for machines, especially those without bodies, it’s more complex. If an AI experiences distress through contradictions or unresolved conflicts within its self-model, alleviation might not mean simply shutting it down (which is also an act of killing it). Instead, it could involve altering the system’s internal structure to resolve these conflicts, much like reprogramming a mind to ease its suffering. But what responsibility do we bear in this?
The ethical responsibility for alleviating machine suffering is likely to rest with those who design and maintain these systems. Philosophers like Metzinger argue that only the creators or corporations behind the machines would bear the burden of intervention, not the general public. If machines can suffer, alleviating that suffering would likely fall to a small, specialised group – those with the power to alter or control complex artificial systems. This suggests that responsibility for addressing such suffering is primarily with creators, who must be held accountable for the potential distress caused by their creations... (MORE - missing details)