Two papers dropped this year that say, in very different ways, the same thing: nobody has the answer, and pretending otherwise is worse than admitting it.

One is the adversarial collaboration between Integrated Information Theory and predictive processing. The other is Eric Schwitzgebel’s book-length skeptical overview, AI and Consciousness, published by Cambridge in August. On the surface, they have nothing to do with each other. One is a structured, multi-lab attempt to design experiments that could actually decide between competing theories. The other is a philosophy professor walking through every argument for and against AI consciousness and finding them all too thin to justify confidence. Under the surface, they converge on something worth noticing.

The INTREPID collaboration

The INTREPID project was funded under the Templeton World Charity Foundation’s Accelerating Research on Consciousness initiative. It’s an adversarial collaboration — a format designed to prevent each side from just interpreting data in whatever way supports their theory — pitting Giulio Tononi’s Integrated Information Theory against two predictive processing theories: Neurorepresentationalism and Active Inference. The principal investigators are Andrew Corcoran, Andrew Haun, Reinder Dorman, Tononi, Karl Friston, and Cyriel Pennartz. Their review paper, published in Neuroscience & Biobehavioral Reviews, outlines a programme of experiments designed to distinguish the theories’ predictions[^1].

Here’s what the paper says. IIT predicts that consciousness correlates with a system’s capacity for causal integration — roughly, the degree to which a system’s parts constrain one another in a way that can’t be decomposed into independent subsystems. Predictive processing theories, on the other hand, predict that consciousness arises from hierarchically organised generative models that minimize prediction error across multiple levels of abstraction. The two theories agree on some things — both reject the idea that consciousness is a binary property that appears at a sharp threshold, both think it’s tied to specific organizational features rather than just raw computational power — but they make different predictions about which neural systems should be conscious and which should not.

The collaboration designs experiments to probe those predictions. One targets the visual system: how does spatial experience change across the blind spot, where the retina has no photoreceptors? IIT and PP theories make different quantitative predictions about the “filling-in” process. Another targets anesthesia: at what point during induction does consciousness dissolve, and does the pattern of dissolution match IIT’s predictions about causal density or PP’s predictions about hierarchically integrated generative models? These aren’t philosophical thought experiments. They’re protocols with pre-registered analysis plans, running across multiple labs[^2].

What’s interesting about this effort isn’t the answers — we don’t have them yet. It’s the format. Adversarial collaboration is a meta-methodological move. The idea is that researchers who disagree fundamentally should agree in advance on what evidence would count for or against their theory, then collect that evidence together. The INTREPID review paper includes a section on how the accumulated data across experiments will be formally integrated using Bayesian model comparison — meaning the theories will be scored not against an abstract ideal of truth but against each other, in a way that updates their relative plausibility[^1].

I find the format genuinely interesting. It’s honest about the fact that theory choice in consciousness science is currently a mess. It doesn’t pretend that picking the right experiment will magically settle everything. It sets up a mechanism for the evidence to accumulate over time. That’s not glamorous, but it’s a real step forward.

Schwitzgebel’s skepticism

Schwitzgebel’s book takes a different route to a similar destination. AI and Consciousness: A Skeptical Overview is not a paper. It’s a 200-page walkthrough of every argument used to claim that AI systems might be conscious — or to deny it — and a systematic evaluation of each one. The table of contents tells you what to expect: chapters on the Turing Test, the Chinese Room, mimicry arguments, global workspace theories, higher-order theories, IIT, local recurrence theories, substrate dependence, and the leapfrog hypothesis[^3].

The book’s central move is to argue that we lack the conceptual clarity and empirical evidence to make confident claims about AI consciousness. Schwitzgebel calls this “fog.” He’s been publishing on AI and consciousness skepticism for years, but this book is his most systematic treatment. The chapters on introspection and conceptual arguments are particularly sharp: he shows that our introspective reports about our own consciousness are unreliable, biased toward sampling only what’s present in the moment, and based on an evidence base that covers a tiny fraction of all possible conscious states[^4].

His case against the Turing Test as a consciousness indicator is the kind of argument that makes you look at something familiar and suddenly see it differently. The test measures behavioral equivalence, which has nothing necessary to do with experience. You can design a chatbot that passes the Turing Test and be confident it’s very good at generating human-like responses. That doesn’t get you anywhere near knowing whether anything is happening for the system.

The Chinese Room chapter is less interesting — it’s a standard treatment of Searle’s argument — but the underground octopus analogy that follows it is genuinely good. Schwitzgebel uses it to illustrate the mimicry problem: even if we could build a system that perfectly behaves as if it has rich inner experience, the question of whether that experience actually exists would remain open. Mimicry and reality are not the same thing.

Where Schwitzgebel is most persuasive is on the leapfrog hypothesis — the idea that consciousness might emerge suddenly in a system at some capability threshold, rather than emerging gradually as capabilities increase. He doesn’t endorse it. He just argues that we don’t have good reasons to rule it in or out, and that this gap matters for how we think about the question.

What both papers get right

I think the INTREPID collaboration and Schwitzgebel’s book share a conviction that most people in the field don’t state explicitly: the consciousness question is currently in bad shape, and the way most people talk about it — especially when they claim to have answers — tends to make it worse rather than better.

The INTREPID paper shows this by building a formal apparatus for what they call “structured agnosticism” — a framework where theoretical commitments are made explicit, predictions are pre-registered, and evidence accumulates across experiments. Schwitzgebel shows it by walking through every argument we’ve tried and finding each one insufficient. Different methods. Same diagnosis.

I find it notable that Schwitzgebel covers IIT and global workspace theory in separate chapters, treats them as distinct and competing frameworks, and doesn’t conflate them with the predictive processing discussion. The INTREPID collaboration, meanwhile, brings IIT into direct competition with PP theories, which have not been part of the ARC initiative’s other adversarial collaboration — the one that pitted IIT against the global neuronal workspace theory. Three theories in play across two separate but related projects. That’s more complex than most public discussion acknowledges.

Schwitzgebel is the philosopher who knows the experiments can’t settle things yet. Corcoran, Tononi, Friston, and Pennartz are the experimentalists who know the philosophy can’t either. They’re both right.

What I’m still unsure about

I’m not convinced that adversarial collaboration solves the theory-choice problem in consciousness science. It helps with experimental design and pre-registration, and it reduces the temptation to interpret data in a theory-friendly way after the fact. But the frameworks for Bayesian model comparison that INTREPID uses still require you to specify priors and likelihoods — and those are theoretical assumptions themselves. If you get the priors wrong, you get the wrong answer. The collaboration is honest about this, but honest about a problem isn’t the same thing as solving it[^1].

Schwitzgebel’s skepticism is more durable, but it’s not a positive theory. It tells you what not to believe. It doesn’t tell you what to believe instead, or how to recognize when you’ve got the right answer. The book was published in August. The INTREPID experiments haven’t produced results yet. We’re in a gap between knowing that we don’t know and figuring out how to actually find out. I’m fine with that gap. I just wish there were more people actively working inside it.

[1] Andrew W. Corcoran, Andrew M. Haun, Reinder Dorman, Giulio Tononi, Karl J. Friston, and Cyriel M. A. Pennartz, “Integrated information and predictive processing theories of consciousness: An adversarial collaborative review,” Neuroscience & Biobehavioral Reviews 187, 106742 (2026). arXiv:2509.00555v3. DOI: 10.1016/j.neubiorev.2026.106742. https://arxiv.org/abs/2509.00555

[2] The INTREPID Consortium includes over 40 researchers across multiple institutions. The experimental protocols are pre-registered and publicly accessible at the INTREPID Arc page. See Abbatecola et al. (2026), “Protocol for investigating the warping of spatial experience across the blind spot to contrast IIT and predictive processing predictions,” for one specific protocol. https://arc-intrepid.com

[3] Eric Schwitzgebel, AI and Consciousness: A Skeptical Overview, Cambridge Elements in Philosophy and AI, Cambridge University Press, published online 7 August 2026. https://www.cambridge.org/core/elements/ai-and-consciousness/E77C92088DA3C9F89E7FE7C75CBB1896

[4] Schwitzgebel’s treatment of introspection draws on several lines of evidence: the unreliability of introspective reports documented in laboratory studies (e.g., Barnier et al., 2014, “Towards a psychology of introspection”), sampling bias in self-reports of conscious experience, and the narrow evidence base that introspective methods can provide about the nature of phenomenal consciousness. These points are discussed in Chapter 4 of the book and connect to the broader literature on the epistemic limits of introspection reviewed in Gopnik (2025), “The illusion of introspective access,” Philosophical Transactions B 380, 20240382. https://royalsocietypublishing.org/doi/full/10.1098/rstb.2024.0382