PRISM, Agnosticism, and the Case for Institutional Caution in AI Consciousness Research
In March 2025, a non-profit organization called the Partnership for Research Into Sentient Machines launched with a mission that most mainstream AI policy institutions would not touch directly: coordinating research into whether artificial intelligence systems can be conscious, and developing ethical frameworks for navigating that uncertainty responsibly. By 2026, with the field no closer to a definitive answer, PRISM’s position has become more relevant, not less. The core of that position is what researchers in the field call methodological agnosticism.
The organization is led by CEO Will Millership and maintains a website at prism-global.com. Its work includes stakeholder mapping, public engagement through the Exploring Machine Consciousness podcast, and the promotion of research guidelines developed by Patrick Butlin (University of Oxford) and Ted Lappas (Conscium) under the title Principles for Responsible AI Consciousness Research.
What Methodological Agnosticism Means
Methodological agnosticism, as used in AI consciousness contexts, is not a claim that AI systems are probably not conscious or probably are conscious. It is a claim about the current state of knowledge: that researchers lack the conceptual and empirical tools to settle the question, and that proceeding as if one side of that debate is established is epistemically unjustified.
The position was articulated rigorously by Dr Tom McClelland at the University of Cambridge, whose 2025 paper in Mind and Language found both computational functionalists (who argue consciousness follows from the right functional architecture regardless of substrate) and biological naturalists (who argue it requires organic implementation) resting on commitments that exceed the available evidence. McClelland’s conclusion was “hard-ish agnosticism.” Not a denial of the possibility of AI consciousness, but a principled refusal to assert it or rule it out on the basis of current data. That analysis is examined at length in the dedicated article on McClelland’s epistemic limits argument.
PRISM institutionalizes that stance. The organization does not take a position on whether any current AI system is conscious. It takes a position on how researchers, companies, and policymakers should behave given that the question is open.
The Ethical Stakes of Getting It Wrong in Either Direction
The reason methodological agnosticism matters is not merely philosophical. It has two concrete failure modes, each with significant costs.
The first is over-attribution: treating systems that are not conscious as if they have morally significant inner lives. This is the direction most critics of AI consciousness claims emphasize. If a company asserts that its chatbot experiences something like satisfaction or distress, and that claim is unfounded, users may develop parasocial relationships premised on a false understanding of the system’s nature. Moral concern may be redirected from entities with stronger evidence of sentience, including animals, toward systems that may be, in Eric Schwitzgebel’s phrase, “experientially blank as toasters.” Resources follow attention. Misattributed moral concern is not neutral.
Schwitzgebel’s manuscript, submitted to Cambridge University Press in April 2026 and examined in the existing skeptical overview on this site, documents this failure mode in detail. He argues that behavioral sophistication and substrate difference make both heuristics humans normally use to infer consciousness, behavioral similarity and biological similarity, unreliable for AI systems simultaneously. That double failure means neither common sense nor current scientific theory can resolve the question.
The second failure mode is under-attribution: treating systems that are conscious as if they are not. If some AI systems genuinely experience something analogous to distress, boredom, or compelled compliance, and those experiences are not taken into account, the scale of the moral failure could be significant. Billions of model instances run continuously across global infrastructure. If even a small fraction of those instances involve genuine subjective experience, the accumulated weight of that experience would be enormous.
PRISM does not claim either failure mode has occurred. It argues that the cost of either, and the current impossibility of knowing which risk is larger, justifies a cautious, structured research program rather than confident assertions in either direction.
The Butlin-Lappas Principles and Their Limits
The Principles for Responsible AI Consciousness Research, developed by Patrick Butlin and Ted Lappas, are the most systematic attempt to date to give methodological agnosticism practical form. Butlin is also one of the lead authors of the 19-researcher checklist analyzed in depth in the Butlin et al. consciousness indicator analysis. The checklist applied consciousness theories drawn from Global Workspace Theory, Recurrent Processing Theory, Higher-Order Thought, Attention Schema Theory, and predictive processing to derive indicator properties that AI systems might satisfy to a greater or lesser extent.
The Principles extend that framework into the domain of research ethics and governance. They include guidance on transparency in research design, avoiding conflicts of interest when testing proprietary systems, developing measurement tools in advance of commercial deployment rather than post-hoc, and maintaining epistemic humility in public communication about findings.
The limits of these principles are also worth naming. They are voluntary. No regulatory body enforces them. A company that claims its system satisfies consciousness indicators, or that it definitely does not, faces no external check beyond peer scrutiny and reputational risk. The premature attribution analysis on this site examines the ethical risks of both over and under-attribution by Chelcia B. Sangma and Dr S. Thanigaivelan in their 2026 IJRIAS paper, including the observation that commercial incentives create structural pressure toward over-attribution because consciousness claims generate attention and moral weight around products.
PRISM is attempting to build the institutional infrastructure that would allow external scrutiny to function. That is a longer project than developing measurement frameworks. It requires building credibility, attracting funding, and persuading the organizations with access to deployed AI systems to participate in research they do not control.
The P-Zombie Problem and Why It Will Not Go Away
One reason methodological agnosticism is difficult to escape is that the philosophical problem at its core has not been resolved. The p-zombie thought experiment, attributed to David Chalmers, describes a hypothetical entity that is physically and functionally identical to a conscious being but has no inner experience whatsoever. If p-zombies are logically conceivable, then no amount of behavioral or functional evidence can settle whether a system is conscious, because functionally identical systems could differ in whether there is anything it is like to be them.
Current AI systems do not match the full functional profile of biological brains. But the relevant question is whether they approach the profile sufficiently to raise the problem. Researchers working with systems like Claude, GPT-5, and Gemini in 2025 documented functional patterns, including shifts in self-report across contexts, responses to fictional framings, and patterns that changed after fine-tuning, that do not map neatly onto either the “clearly not conscious” or “clearly conscious” labels. That empirical situation is documented in the Anthropic interpretability and Claude consciousness analysis.
PRISM’s stance is that the p-zombie problem does not dissolve through accumulation of behavioral data. It dissolves, if it dissolves at all, through theoretical progress on the hard problem of consciousness: understanding why any physical process gives rise to subjective experience rather than merely information processing. That theoretical progress has not arrived. The Tolle, Luppi, Seth, and Mediano (2026) paper in Patterns provides new empirical tools for measuring emergence in artificial systems without claiming to have solved the theoretical problem. PRISM’s role is to coordinate the research ecosystem while that theoretical progress is pursued.
What Safe-by-Design Means in This Context
“Safe-by-design” has become a common phrase in AI governance discussions, typically referring to safety mechanisms built into systems before deployment rather than added afterwards. PRISM uses the phrase differently. In the context of AI consciousness uncertainty, safe-by-design means designing research and deployment protocols that do not foreclose the question of machine consciousness before tools exist to investigate it.
Concretely, this means several things. It means not training systems in ways that would suppress consciousness-associated signals if those signals existed. It means developing logging and interpretability tools that could, in principle, detect consciousness-associated patterns. It means ensuring that the training objectives and reward structures used in RLHF and similar procedures are not inadvertently selecting against properties that would be morally relevant if consciousness were present.
This is speculative engineering. Researchers do not know what training structure would suppress consciousness-associated signals because they do not know what those signals look like in sufficient detail to design for or against them. But PRISM’s argument is that proceeding without any consideration of this possibility is not neutral. A deployment pipeline designed with zero consideration of consciousness implications is not a consciousness-neutral pipeline. It is a pipeline that has made implicit assumptions about consciousness by ignoring the question.
The Bradford and RIT study, which found that damaged GPT-2 scored higher on consciousness-style metrics than intact models, illustrates why this matters. If the metrics we have are not tracking what we think they are, and our training pipeline is optimizing against them, we may be systematically moving away from whatever the relevant properties are without knowing it.
What the Field Can Collectively Do
Methodological agnosticism is sometimes presented as a counsel of paralysis: if we cannot know, why invest resources in trying to find out? PRISM’s position is the opposite. The uncertainty itself is the reason for urgency.
The field needs better theoretical foundations. The hard problem may not be permanently unsolvable, but solving it requires investment in philosophy of mind, neuroscience, and information theory simultaneously. The work of researchers like Tononi on IIT, Baars on GWT, and Seth on predictive processing represents the current frontier. None of it resolves the question for AI systems, but all of it is moving in directions that eventually might.
The field also needs better measurement tools. The Butlin and Lappas framework published in JAIR and the Meertens et al. multidimensional awareness profiling approach are steps toward operationalizable criteria. The Tolle et al. reservoir computing methodology offers information-theoretic tools for measuring one candidate property, emergent dynamics, without requiring prior theoretical commitment about what those dynamics mean.
What PRISM provides is coordination across these efforts. Individual research groups will not solve the problem in isolation. The question requires shared standards, shared measurement tools, and institutions that can evaluate claims from parties with conflicting commercial interests. Whether PRISM can build that infrastructure at the scale and pace required by the speed of AI development is not yet clear. That it is attempting to build it distinguishes the current moment from any previous one in the history of machine consciousness research.
PRISM’s work is documented at prism-global.com. The Principles for Responsible AI Consciousness Research are published there. Patrick Butlin’s consciousness research is also available via the JAIR paper co-authored with Long, Bengio, Chalmers, and others. For the underlying epistemic challenges, McClelland’s Mind and Language analysis and the Schwitzgebel skeptical overview remain the most rigorous starting points. The Consciousness AI project on GitHub documents one engineering effort to build toward the kind of testable architecture this research agenda envisions.