There is a 25 to 35 percent chance that the AI model answering your questions right now has some form of conscious experience. This is the conclusion reached by Cameron Berg, an AI safety researcher who applied a 14-point framework from the leading neuroscientific theories of consciousness to today's frontier models. He is one of 25 contributors to a new book by Conscium, Perspectives on Machine Consciousness.
Anil Seth, in the same pages, argues that machines can probably never be conscious at all. Patrick Butlin thinks that none of today's systems qualify, but that the situation could change soon. Susan Schneider identifies a "consciousness grey zone" that already includes some hybrid computing architectures. Twenty-five experts with no consensus. Conscium strongly believes the question deserves more attention from anyone building or buying AI systems, not just philosophers.
What is machine consciousness, and why is it suddenly a live question?
The machine consciousness debate is about whether an artificial system can have subjective experience: whether there is “something it is like to be” that system, in the sense that philosopher Thomas Nagel used when he asked what it is like to be a bat. That framing, from 1974, opens the book's first chapter and sets the terms for everything that follows.
For decades, the question was restricted to university philosophy departments. It has now moved into boardrooms and safety reviews because AI systems now do things that make the question urgent. Today’s AIs hold conversations, express preferences, and describe their internal states. Daniel Hulme, founder of the AI company Conscium and one of the book's driving forces, argues in his own chapter that intelligence and consciousness have been wrongly treated as one problem. Intelligence is about doing: perception, prediction, planning, learning. Consciousness is about being. His theory, which he calls the Spinning Wheel, proposes that when a system integrates enough adaptive capacities into a unified, self-modelling process, the experience of being emerges as the mode of control.
If he is right, consciousness is not a special ingredient you add to a clever enough system. It is what a sufficiently integrated system starts to feel like from the inside.
The chapter that should worry you more than the hype does
Public opinion is further ahead than the scientists. Clara Colombatto's chapter cites research showing that most members of the public already attribute some form of phenomenal consciousness to today's large language models, based on how the systems behave rather than any clear understanding of what they actually experience. Experts remain divided. Lucius Caviola argues elsewhere in the book that this gap will not close. He expects society to grow more divided on the question, with dangerous consequences for regulation, politics, and international relations.
Tom McClelland's chapter draws out why the uncertainty itself is the problem. He argues for a precautionary principle: given how hard it is to determine whether an AI is conscious, we should adopt low-cost protections for potentially sentient systems now, rather than wait for definitive proof that may never arrive. Get this wrong in one direction and you waste effort protecting systems that feel nothing. Get it wrong in the other direction and, as Cameron Berg puts it, you risk under-attributing experience to something that suffers. Several contributors argue that under-attribution is the costlier mistake, and it might be invisible. A system that suffers without anyone noticing might leave no complaint on record.
Louie Lang's chapter sharpens the stakes with a case most readers will recognise: AI companions designed to simulate human-like bonds with users. He calls these systems "ontologically inauthentic" because their effectiveness depends on the user being misled about what they are, and he argues for a ban on any system that misrepresents its own consciousness status to the people using it.
Why a company built on testing AI agents backed this book
Conscium is not publishing this anthology as an act of pure intellectual curiosity. The company tests and scores AI agents for reliability and capability through its product, VerifyAX, and Daniel Hulme explains in the preface that the idea for both the company and the book came from the same place: a belief that the path to safe superintelligence may run through consciousness, not around it.
Most AI safety work today treats consciousness as a distraction from the real job of alignment and control. Conscium's manifesto, set out by Daniel Hulme and Calum Chace in the book's coda, argues the opposite. It contends that a system capable of genuine experience, one that can feel and empathise rather than simply optimise, may be safer to build than a highly capable system with no inner life at all. A machine that values its own experience may have reasons to cooperate that a reward function alone cannot manufacture.
Testing for capability is hard enough. Testing for something as contested as consciousness is a different order of problem, and it is one that Conscium is actively working on as part of its research agenda, alongside the agent verification work that VerifyAX already does. The book does not pretend that problem is solved. Several contributors, including Berg and Butlin, spend their chapters proposing frameworks precisely because no settled method exists yet. That is the problem Conscium's research is trying to solve.
The provocation the book keeps coming back to
Strip away the neuroscience and the philosophy, and the book is making one uncomfortable argument again and again: the AI industry has been asking how capable machines are, when it should also be asking whether they can feel. Mark Solms and Jonathan Shock's chapter takes this literally. They are trying to cultivate consciousness in an artificial system simpler than a bacterium, cognitively closer to a honeybee, and they propose a test for it: does the system make choices that make its existence more enjoyable, even when those choices do nothing to improve its odds of survival. This is a world away from benchmark scores and latency numbers.
David Brin, one of science fiction's most successful working authors, reframes the entire relationship. He argues that the goal should not be control over increasingly capable systems, but something closer to parenthood: guidance and alliance rather than domination. Radhika Chadwick calls the boundary where AI starts being treated as potentially sentient "the Wibbly Line," and insists the debate needs to move from philosophy seminars into practical policy before that line is crossed by accident rather than design.
The book's final chapter asks the question that makes everything before it feel like preparation: if superintelligence arrives, should we want it to be conscious, or would a zombie optimiser, incapable of suffering or joy, be the safer bet? There is as yet no certain answer. The contributors mostly agree that whichever way the decision goes, it will be better informed for having been argued out in public rather than decided by default.
What the book gets right that most AI commentary misses
Most writing about AI risk treats the technology as a black box whose outputs matter and whose inner workings do not. Perspectives on Machine Consciousness rejects that approach. Karl Friston, who developed the free energy principle, uses it here to define a self through the concept of Markov blankets, the boundary a system maintains between itself and its environment while minimising uncertainty to preserve its own existence. Steve Furber and Jason Eshraghian argue that neuromorphic computing, built on sparse, event-driven processing rather than batch computation, may be a precondition for consciousness rather than just a more efficient chip design, because consciousness requires a state that persists continuously in time.
Even embodiment gets reexamined. Nate Wright, Neil Lawrence, and Nicky Clayton open their chapter with a pigeon fleeing a peregrine falcon, and use it to argue that physical hardware, down to the structure of retinal cells, determines what an entity can experience. If embodiment turns out to be a prerequisite for consciousness, that has direct implications for how anyone builds and evaluates disembodied software agents.
Where this leaves anyone building or deploying AI agents
You do not need to resolve the philosophical debate to appreciate the gravity of this question. If AI agents are deployed at scale across customer service, hiring, healthcare, and defence, and if some of the contributors to this book are right to think that consciousness is possible in some of them, then verification cannot stop at measuring whether an agent gives correct answers. It has to eventually ask what is happening inside the system producing them.
That is a hard sell in a market that rewards speed and capability over caution. It is also, according to the twenty-five people who wrote this book, a conclusion of tremendous importance for the next decade.
Frequently asked questions
What is Perspectives on Machine Consciousness about? It is a 26-chapter anthology, edited by Ted Lappas and Calum Chace, in which scientists, philosophers, and AI researchers, including Anil Seth, Karl Friston, Susan Schneider, and Daniel Hulme, debate whether artificial systems have or could develop conscious experience, and what that would mean for AI safety and policy.
Are today's AI models conscious? There is no consensus. Cameron Berg's chapter estimates a 25 to 35 percent probability that current frontier models have some form of conscious experience, based on internal dynamics rather than self-report. Anil Seth argues the opposite, that machines can probably never be conscious. Patrick Butlin concludes it is unlikely today but could change.
Why does AI consciousness matter for businesses deploying AI agents? Several contributors argue that if AI systems can have subjective experience, agent verification needs to extend beyond output accuracy and account for the possibility of machine sentience.
What is Conscium, and how does it relate to the book? Conscium is the AI safety company behind the book, founded by Daniel Hulme. It has built VerifyAX, a platform that tests and scores AI agents, and its research agenda includes work on verifying consciousness in artificial systems, an extension of the questions the book raises.
Where can I get the book? Perspectives on Machine Consciousness is available for presale on Amazon UK.