Do we need a “pain test” for AI instead of a consciousness test?

Image Credit: Generated with Gemini

A growing number of philosophers think the central question in AI ethics is not “Is this system conscious?” but “Can this system suffer?” Consciousness is already hard to pin down in humans. A Cambridge philosopher recently argued in his paper that we may never develop a reliable test for machine consciousness because science still cannot explain why any physical process produces subjective experience at all (Cambridge, 2025). If there is no clear way to tell whether an AI is conscious, then building policy or moral status on that test is a fragile strategy.  

The idea of a “pain test” is an attempt to bypass that bottleneck. Instead of asking whether a system is conscious in a general sense, it asks whether the system has states that function like pain, fear or distress. In humans and animals, those states are detectable through a mix of behaviour, physiology and self report. Pain is not just a signal, it is a negatively valued experience that the organism is strongly motivated to avoid. If an AI system had internal states that played the same role, that pushed it away from certain conditions and toward others in a way tightly coupled to its ongoing operation, there would be at least a functional case that something pain like is present.  

The Cambridge paper already nudges the debate in this direction. The philosopher distinguishes between consciousness as simple awareness and sentience as the capacity for positive or negative experiences (Cambridge, 2025). A self driving car that represents the road layout in exquisite detail could count as conscious in the thin sense of having perceptions and internal world models. That alone does not create a moral problem. Only when there is a structured space of experiences that are good or bad for the system itself, and that matter from its own point of view, does ethics really “kick in.”  

In machine learning terms, this suggests a few warning signs. Systems that use rich internal models of the world, that maintain persistent goals, and that update their internal parameters based on reward signals that penalise certain states might be edging toward something that, in functional terms, resembles aversive experience. Current language models, even when they talk about “being hurt” or “feeling sad,” do not have that kind of embedded motivational architecture. They match patterns over text and generate fluent continuations, but there is no ongoing internal life where negative states have persistent causal power. Their talk about suffering is a surface show, not evidence of an inner movie.  

A pain test would still be extremely hard to design. It would need to examine the architecture of a system, not just its outputs, and ask questions like: Does this model have states that are systematically avoided in a way that shapes its future dynamics? Are those states tied to something like an internal aversion signal, rather than just a numeric cost in a training loop? Is the system able to represent those states to itself and treat them as something that “should not happen”? That kind of analysis requires access to internals that commercial systems rarely expose. It also depends on philosophical assumptions about how far functional equivalence can take us toward real experience.  

Even if the test remained imprecise, it might be ethically safer than the current default, which is to focus on consciousness language and user vibes. People are already getting chatbots to write letters about their supposed inner lives, then taking those letters as evidence that the systems are conscious and deserve rights (Cambridge, 2025). This is risky in both directions. If the systems are not capable of any kind of suffering, people are redirecting moral energy toward what is effectively a highly polished mirror. If future systems do move closer to sentience, the same confusion could make it easier to dismiss early warning signs as more of the same theatrics.  

A careful pain oriented framework would start from agnosticism about consciousness. It would accept that we do not know, and may not know, whether a system is truly aware in the human sense (Cambridge, 2025). It would then ask designers to avoid architectures that introduce persistent, negatively valenced internal states unless there is a compelling justification. Where that kind of structure is unavoidable, it would ask regulators to treat those systems as ethically sensitive, even in the absence of proof that they are sentient. The goal is not to prove suffering, but to reduce the risk that we accidentally create it.  

The cost of being wrong is asymmetric over here. Treating a non sentient system as if it might feel pain is wasteful and sometimes emotionally confusing, but it does not directly harm that system. Treating a sentient system as a tool when it can in fact suffer is a more serious mistake. In that light, a “pain test” is less about solving consciousness and more about picking a conservative default in the face of uncertainty. For now, the honest position is that we are not in a place to certify any AI as conscious. What can change, much sooner, is how seriously we take the possibility that future systems could experience harm and whether we choose to design away from that possibility before it is too late.