Philosopher: Nobody Knows If AI Could Be Conscious, And Science Can't Tell Us

With sentience comes suffering: Is it unethical to create AI potentially capable of experiencing distress?

Artificial intelligence dominates headlines, Wall Street, and popular culture right now. As AI continues to make leads and bounds in terms of capabilities and intuitiveness, one philosophers posits all of humanity faces an uncomfortable truth regarding the AI race. We may build something that seems conscious without having a reliable way to know if it actually is.

Thus, University of Cambridge philosopher Tom McClelland argues that when the evidence can’t rule out sentience and suffering, the safest option is to simply avoid building that kind of AI in the first place.

Despite intense scientific interest in artificial consciousness, Dr. McClelland’s paper suggests that researchers cannot responsibly conclude whether sophisticated AI would experience the world as humans do. Both believers and skeptics in the debate are making unjustified leaps beyond what evidence supports, he argues.

As companies develop increasingly sophisticated AI and governments consider regulations, McClelland proposes a precautionary principle: if science cannot rule out the possibility that an AI might have positive or negative experiences, it shouldn’t be created at all.

“If an AI is such that we don’t know whether it is sentient, we should avoid the moral dilemma it would present us with by not developing that AI,” McClelland writes in his paper from the University of Cambridge’s Department of History and Philosophy of Science.

Why the Evidence Doesn’t Reach

The problem stems from what philosophers call the “hard problem” of consciousness. Scientists can study which brain processes correlate with conscious experiences, but cannot explain why those processes create subjective feelings at all.

Consider studies of inattentional blindness, where people watching a basketball game fail to notice a gorilla walking through the scene. Global Workspace Theory explains this by proposing that consciousness occurs when information broadcasts widely across the brain. The basketball player’s visual information made it into this global workspace while the gorilla’s did not.

But this explains which information was conscious without explaining why. Nothing about the theory tells us why globally broadcast information should feel like anything at all. McClelland calls these “shallow explanations” because they can identify which human brain states are conscious without explaining why consciousness exists.

When researchers try to extend these theories to artificial intelligence, they encounter what McClelland terms an “epistemic wall.” Evidence from biological consciousness simply does not reveal whether similar patterns in silicon-based systems would also generate consciousness.

Two possibilities remain open. Either the computational patterns themselves create consciousness regardless of physical substrate, or consciousness requires biological processes that AI lacks. Available evidence cannot distinguish between these competing scenarios.

Why Both Believers and Skeptics Overreach

Advocates for AI consciousness often assume computational functionalism: the idea that proper computations produce consciousness whether running on neurons or computer chips. Critics argue consciousness requires biological processes. McClelland contends both groups venture beyond evidential support.

“The fact that computational functionalism is so frequently assumed does not mean it is true,” he writes, quoting consciousness researcher Anil Seth.

McClelland examined major approaches for assessing AI consciousness. The “theory-heavy” approach selects a specific consciousness theory and applies it to AI. The “theory-light” approach identifies multiple consciousness markers from different theories and checks whether AI displays them.

Both face what he calls the “Evidentialist Dilemma.” Researchers can adopt modest theory versions staying within evidence from biological systems, offering no verdict on AI consciousness. Or they can adopt bold versions making AI consciousness claims but exceeding evidential justification.

A recent report by researchers including Patrick Butlin and Robert Long proposed indicator properties for assessing AI consciousness, drawn from multiple theories. McClelland focuses on hypothetical “Challenger-AI” systems displaying all these indicators. Even such advanced AI would not resolve the uncertainty, he argues, because the indicators derived from biological systems.

150 Years of Research, No Progress on the Core Question

Could future scientific advances overcome this obstacle? McClelland argues such optimism lacks justification. Despite 150 years of neuroscience progress since philosopher Thomas Huxley first noted consciousness’s mystery, science has made no headway on the fundamental question.

“New theories simply relocate the apparent miracle of consciousness without managing to make it any less miraculous,” McClelland writes. Different theories identify different processes responsible for consciousness, yet none explain why any process should generate subjective experience.

Current AI systems like large language models likely lack consciousness, McClelland notes, because researchers can explain their outputs through processes unlike those underlying human consciousness reports. The problem emerges with more advanced hypothetical AI displaying biological consciousness markers. For such systems, evidence becomes genuinely ambiguous.

McClelland distinguishes his agnostic position from mere uncertainty. Many researchers express caution while still taking positions. Agnosticism means taking no position because evidence cannot support one. The only justified confidence level is no confidence either way.

Focusing on Suffering Instead of Consciousness

If science cannot determine whether advanced AI would be conscious, how should society handle ethical questions? McClelland proposes focusing on sentience (the capacity for positive or negative experiences) rather than consciousness generally. An entity could theoretically be conscious without experiencing anything as good or bad.

Researchers might determine whether AI mental states would be pleasant, unpleasant, or neutral if those states were conscious. If an AI would lack valenced experiences even if conscious, no special ethical precautions are needed. But if an AI might experience suffering if conscious, McClelland recommends not building it.

This allows useful AI development to proceed while avoiding systems that might have negative experiences. As McClelland notes, citing researcher Susan Schneider: “Consciousness is the philosophical cornerstone of our moral systems, being central to our judgment of whether someone or something is a self or person rather than a mere automaton. And if an AI is a conscious being, forcing it to serve us would be akin to slavery.”

Rather than risking creating enslaved, suffering entities when science cannot verify their nature, McClelland suggests acknowledging the limits of current knowledge and avoiding creating the problem altogether.

---