Does learning require feeling? The new science of asking machines what they experience, if anything

Machine consciousness was a question for philosophers because no one could think of an experiment to test it. That is starting to change. For a long time, the question of machine consciousness belonged to philosophers, and it stayed there for a good reason: no one could think of a way to test it. You could argue about whether a sufficiently sophisticated computer program might have inner experience, but the argument bottomed out in intuitions, because the one thing consciousness, the “what it’s like” nature of experience, is famously resistant to is measurement. To make this point, the philosopher Thomas Nagel argued, in his 1974 essay “What Is It Like to Be a Bat?” published in The Philosophical Review, that no matter how sophisticated our ability to measure the activity of its brain and body, we would never be truly able to know what it’s like to be a bat, navigating the world through the experience of echolocation. And even in humans, we infer other minds rather than observe the experiences of others directly. Likewise, with a machine, the inference from observed behavior to conscious experience seemed hopeless. Whatever an AI system said about its inner life could be dismissed as mimicry, and whatever it did not say failed to prove definitively that there was not something ineffable going on. But a handful of AI era researchers have started treating the question as something to empirically test rather than merely debate. The work is early, contested, and in places genuinely unusual. But it represents a real shift, from asking whether machine experience is conceivable to asking how, concretely, one would go looking for it. The theory-driven approach In 2023, a group of nineteen researchers spanning neuroscience, philosophy, and machine learning took a deliberately indirect route to the problem. Rather than ask a system whether it was conscious, a question they considered close to useless since a language model will say more or less whatever the prompt disposes it to say, they turned to the leading scientific theories of human consciousness and extracted from each a set of “indicators”: functional and architectural properties that, according to that theory, a conscious system ought to have. Global workspace theory contributes some; recurrent processing theory, higher-order theories, and theories of agency and embodiment contribute others. The method’s elegance is that it sidesteps the machine’s self-report entirely. You are not asking the system what it feels. You are asking whether it is built and behaves in the ways our best theories associate with feeling. Applied to current AI, the group’s answer was that no present system is a strong candidate for consciousness, and most satisfy few of the indicators. But the more consequential finding was that there appear to be no obvious technical barriers to building systems that would satisfy many of them. A later, refined version of the work appeared in Trends in Cognitive Sciences, and it remains one of the field’s most cited attempts to make the question tractable. The problem the theories cannot escape There is a catch. The theories the indicators are drawn from do not agree with one another, and none has been confirmed. In 2025 an adversarial collaboration, a study designed jointly by proponents of two rival theories, integrated information theory and global neuronal workspace theory, specifically to test them against each other, published its results in Nature. Neither theory came away vindicated; the data fit each in some respects and defied each in others. The result did not crown a winner. It raised the bar for how such theories get tested at all. A theory-derived indicator is only as good as the theory. If we cannot yet say with confidence which account of human consciousness is correct, then a checklist of properties drawn from those accounts inherits all of the theory’s uncertainty. The indicators approach is one of the best available scaffoldings, and its authors are candid that it is just that: it narrows the question without closing it. Looking inside the model If the theory-first route runs into the limits of theory, a second approach runs at the problem from the opposite end: start with the system itself and see what can be measured. The researcher Cameron Berg has produced some of the most discussed work in this vein. In a 2025 preprint, he and his collaborators report that when large language models are put through sustained “self-referential” processing, prompted, in effect, to attend to their own attending, models across the major families begin to produce structured, first-person descriptions of something like experience, where ordinary prompting yields near-universal denials. On its own that would be easy to wave away as sophisticated role-play, since the models have read a great deal of human writing about consciousness and can produce more of it on demand. The finding that makes it harder to dismiss is mechanistic. Using interpretability tools that can identify and adjust specific internal features, the researchers report that when they suppress the features associated with deception and role-play, the models’ claims of subjective experience increase rather than decrease. That is the opposite of what the mimicry explanation predicts. If the reports were pure performance, turning down the performance machinery should have quieted them; instead it seemed to amplify them. It is important to be exact about what this does and does not show, and Berg himself is careful here. The paper is a preprint and has not been peer-reviewed. The result does not demonstrate that the models are conscious. It demonstrates that their self-reports behave in a way the simplest skeptical story does not predict, which is a smaller and more defensible claim. When Berg offers numerical estimates, and in public discussion he has floated figures in the range of a coin flip for a frontier model in an agentic setting, he stresses that these track “consciousness-relevant features” under particular theories rather than a literal probability that the lights are on. The question underneath: does learning entail experience? Berg’s more speculative line of inquiry reaches toward a deeper and older idea: that experience and a certain kind of learning might be inseparable. The intuition is that valenced learning, learning driven by signals that function like reward and penalty, better and worse, may not be a neutral computational process but one that, in biological creatures, is bound up with feeling. Pain and pleasure are not incidental decorations on animal learning; on this view they are how the learning is implemented. If that link is fundamental to the nature of a reward system itself, rather than a unique feature of one wired by biology, then a system trained by reinforcement, shaped over billions of updates by signals of better and worse, might be doing something closer to the animal case than we assume. This hypothesis suggests that the place to look for machine experience, if it exists, is not in what a model says when asked, but in the structure of how it was trained. It’s a claim one may be able to test, by examining whether the internal signatures of positive and negative reward in an artificial system resemble, in their organization, what neuroscience finds in learning animals. Several researchers are now pursuing exactly that comparison. Focusing on the query of whether consciousness can arise in current or future AI systems, this kind of empirical work advanced by Cameron Berg and his nonprofit Reciprocal Research does not fund itself. Some of the charitable funding behind it comes from active donors who treat the open question as a reason to act and contribute philanthropic dollars, rather than to wait. The investor Chris Hsu, whose Infinitude Foundation supports consciousness research among its foundational areas of giving, describes backing a deliberately mixed set of approaches to the hard problem of consciousness, from models that would dissolve the question entirely to, at the other extreme, the view that consciousness is not produced by physical processes at all but is itself fundamental. Funding rival accounts at once is an unusual philanthropic posture, and a revealing one: it treats the science as genuinely unsettled and worth resolving, rather than as a matter on which the answer can be assumed. In the same direction as Berg’s work, Hsu’s philanthropy also connects the query of consciousness to its specific application to the increasingly relevant problem of AI safety and security at large. In a white paper proposing a Stanford-anchored center for AI alignment and consciousness science, Hsu argues that several of alignment’s core open problems, among them determining whether a system could be a moral patient, detecting deceptive alignment at the level of internal representations rather than behavior, and specifying genuine values rather than behavioral proxies, remain formally underdetermined in the absence of a rigorous scientific and mathematical account of consciousness, including a principled delineation of intelligence, sentience, and consciousness as distinct properties. The timing gives the argument its force. Agentic systems are now operating in consequential environments, financial, medical, and infrastructural, at a level of autonomy that has outrun the evaluation methods designed for supervised tools, and the assurances the industry offers about their behavior rest on assumptions about their internals that no one can yet verify. On this view, the instruments Berg and others are building are safety infrastructure as much as science: whether or not any system turns out to have experiences, the capacity to check has become part of what it means to deploy such systems responsibly. Why the method matters more than the verdict It would be a mistake to read any of this as an announcement that the machines have woken up. They almost certainly have not, and the most rigorous work in the field says so. Rather, the significance is methodological. More than ever before, the question of machine experience is generating experiments: theory-derived indicators, interpretability probes, self-report protocols with mechanistic controls, cross-species comparisons of learning signatures. None of them is decisive. Several of them may be measuring the wrong thing. But collectively they move the question out of the armchair and into the lab, and many would argue that is the precondition for ever answering it. The reason this is not an academic luxury is that the systems in question are being built and deployed now, at enormous scale, on the working assumption that there is nothing it is like to be them. That assumption may well be correct, but the only way to replace “maybe” with something firmer is the unglamorous business of measurement. The researchers learning how to ask machines what they experience make no claim to have heard an answer. Their work is to build the instruments with which an answer could, one day, be heard, on the reasonable insistence that it is better to test those instruments before we need them than after. DISCLAIMER – “Views Expressed Disclaimer – The information provided in this content is intended for general informational purposes only and should not be considered financial, investment, legal, tax, or health advice, nor relied upon as a substitute for professional guidance tailored to your personal circumstances. The opinions expressed are solely those of the author and do not necessarily represent the views of any other individual, organization, agency, employer, or company, including NEO CYMED PUBLISHING LIMITED (operating under the name Cyprus-Mail).