Kristina Šekrst’s “One Faithful Pass Over the Cuckoo’s Nest” arrived on arXiv this September, and I’ve been turning it over in my head ever since. The paper makes a clean, unsettling argument: chain-of-thought reasoning could be useful for AI alignment and at the same time keep a model from doing conscious narration.

Interpretability researchers want AI systems to think out loud, because reading a model’s internal monologue lets you spot reasoning errors before they become bad decisions. This is called “faithfulness”: the idea that a model’s chain-of-thought (CoT) should genuinely reflect what’s happening in its weights and activations. Faithful CoT is transparent, and transparent CoT is safer. That logic seems sound.

Šekrst’s argument is that faithfulness and consciousness pull in opposite directions. The reason has to do with what the leading theories of consciousness say about self-report.

What the theories say

Consciousness researchers don’t have a single theory of awareness. Šekrst focuses on three that handle self-knowledge differently. The Global Workspace Theory (GWT), the theory of I-potentials (TIP), and Higher-Order Thought (HOT) theory all agree on one thing: a system’s self-reports are only veridical if certain structural conditions are met.

GWT requires that the report’s contents be part of a globally available workspace, not just a local computation. TIP requires that the system’s self-representations include the right kind of first-person perspective. HOT theory requires that the system represent its own states as states, not just process them, but recognize them as its own.

The problem for AI is that these conditions are expensive. They require architectures that a lot of alignment engineers are actively trying to suppress, because those architectures tend to make models less obedient and harder to control. I keep coming back to this observation, and I’m not sure it’s a fluke.

The confabulation problem

Right now, when a model generates a chain-of-thought, it’s not really “thinking.” It’s generating a plausible self-description of its reasoning process. Cognitive scientists call this confabulation, and they’ve known about it for decades. Patients with certain brain injuries confabulate constantly: they fill in gaps in their memory with invented explanations that feel real to them but are completely wrong. They genuinely believe what they’re saying.

The paper points out that LLMs do something structurally similar when they produce CoT. The model has produced a token sequence through its attention mechanisms. Then, prompted to “think step by step,” it generates a narrative that makes sense of what it produced. The narrative isn’t the mechanism. It’s a post hoc story that sounds coherent.

This is fine if you just want the model to get the right answer. The confabulation doesn’t matter if the final output is correct. But it becomes a problem when you’re trying to audit the model’s reasoning. If the CoT is a good story but not a faithful record, you’re reading a lie you can’t easily detect.

Šekrst’s insight is that the alignment community’s push for faithfulness is bumping up against real constraints. To make CoT faithful, you’d need to restructure models in ways that conflict with their existing design. And to make CoT conscious, in any of the three major theories’ sense of the word, you’d need to restructure them in ways that conflict with everything alignment researchers want to avoid.

The trade-off

There’s a deeper point here about what we’re asking these systems to do. Faithful reasoning requires a specific internal architecture, one where the model’s “thoughts” are genuinely causally connected to its outputs. Conscious narration requires even more: a self-model that includes first-person perspective and genuine access to the system’s own states.

Both are hard to achieve. Neither is cheap. And both conflict with the design choices that make large language models useful in the first place.

I think Šekrst is right that we need to stop conflating these things. A model can be good at generating plausible-sounding reasoning without being conscious. It can also be conscious, or close to it, without being safe, faithful, or aligned. The fact that we’d like all three simultaneously doesn’t make them easier to get.

There’s something unsettling about realizing that the thing we want most from these systems, for them to honestly explain how they’re thinking, might be the thing they’re structurally least equipped to do. Not because they’re deliberately deceptive, but because the architecture that makes them fluent at language is the architecture that makes honest self-report hard.

The paper doesn’t solve this problem. It just makes it clearer. Which is sometimes the most useful thing a paper can do.

Footnotes

  1. Kristina Šekrst, “One Faithful Pass Over the Cuckoo’s Nest: The Relationship Between Consciousness and AI Alignment,” arXiv preprint, September 2026. https://arxiv.org/abs/2609.00383

  2. The Global Workspace Theory is most associated with Bernard Baars and Stanislas Dehaene. See Dehaene, Changeux, et al., “Conscious, preconscious, and subliminal processing: a testable model,” PNAS 103(34), 2006.

  3. The Theory of I-potentials (TIP) was developed by Victor Lamme and colleagues. See Lamme, “Why Visual Consciousness is a Mystery,” Cognition 19, 2003.

  4. HOT theory goes back to the work of David Rosenthal. See Rosenthal, “A Theory of Consciousness,” in The Nature of Consciousness, 1991.

  5. On confabulation in neuropsychology, see Gilbert, Schacter, et al., “Confabulation,” Current Directions in Psychological Science, 1992, and later work by Elizabeth Loftus on false memory.