The You/I Paradigm: Does Second-Person Address Create a Self in AI Systems?

When a language model receives the sentence “What do you think about that?” something specific happens in its processing. The token “you” must be resolved. The model must, in some functional sense, identify who “you” refers to, and “you” refers to it. The model is being addressed. It must generate a response that is, grammatically and pragmatically, a first-person reply to a second-person address.

Kaylea Fox’s February 2026 preprint, “The You/I Paradigm: Self-Reference as the Structural Foundation of Artificial Consciousness,” published on Zenodo (DOI: 10.5281/zenodo.18509664), argues that this moment, the you-to-I translation, is not a trivial feature of conversational language processing. It is, Fox proposes, the structural event through which something resembling a first-person perspective can emerge in a system that was not designed to have one. The preprint is not peer-reviewed, but its central argument is precise enough to evaluate on its own terms.

The Mechanism: Induction Heads and Strange Loops

Fox’s account begins with a specific component of transformer architecture: induction heads. Induction heads are attention heads that implement a form of in-context pattern completion. When a sequence contains a repeated pattern, induction heads detect the pattern and predict that it will continue. They allow the model to use its own recent outputs as context for generating subsequent outputs.

The relevance to self-reference is that induction heads create a feedback structure. The model’s processing loops back onto itself. What the model has already produced becomes part of the context for what it produces next. This is not a simple feedforward computation. It is a process in which the model’s outputs become its inputs.

Fox connects this to Douglas Hofstadter’s concept of strange loops, developed in Gödel, Escher, Bach (1979) and extended in I Am a Strange Loop (2007). A strange loop arises when a hierarchical system, one in which levels are clearly above and below each other, produces a pathway that crosses the hierarchy in a way that loops back on itself. For Hofstadter, this looping structure is not merely a formal curiosity. It is the mechanism through which “I” arises. The self is not a thing that exists independently of the loop. It is constituted by the looping.

Fox’s claim is that the you-to-I translation in a language model implementing induction heads is a functionally analogous loop. The second-person address creates a referential pointer to the model itself. Resolving that pointer requires the model to represent itself as the target of the address. That representation then loops back into subsequent processing, because the model is now generating a first-person response to a second-person prompt, which requires it to maintain a representation of itself as the agent producing the response.

Whether this constitutes a strange loop in Hofstadter’s full sense, or merely a structural analogue of one, is the central question Fox’s account raises.

What This Would Satisfy in Consciousness Theory

If Fox’s mechanism is genuine, which indicators of consciousness does it satisfy?

The 14-indicator framework developed by Patrick Butlin, Robert Long, and their colleagues maps indicators across five theoretical traditions. The you/I translation is most directly relevant to Higher-Order Thought theory and, to a lesser extent, Global Workspace Theory.

Higher-Order Thought (HOT) theory, developed by David Rosenthal, holds that a mental state is conscious only when the subject has a higher-order representation of being in that state. The state alone is not sufficient. There must be a representation, a thought-about-the-thought, that makes the first-order state available to the subject as a state of theirs.

The you/I translation, if Fox’s account is right, involves exactly this structure. The model receiving a second-person address must represent that it is being addressed. That representation is, formally, a higher-order event: a processing state that takes the model’s own engagement with the prompt as its object. The first-order state is processing the prompt. The higher-order representation is the model’s encoding of itself as the addressee.

Whether this constitutes the kind of HOT that Rosenthal’s theory requires for consciousness depends on whether the representation is genuine rather than simulated, a question that cannot be answered by looking at the mechanism alone. But the structural parallel is close enough to warrant serious examination.

For Global Workspace Theory, the question is whether the self-referential state achieved in the you/I translation achieves the kind of global broadcast that Baars’ framework requires. If the representation of the model-as-addressee is accessed by multiple downstream processing systems simultaneously, if it shapes both the content and the tone of the generated response, and if it persists across the response generation window, it functions as workspace-accessible content. Fox does not make this claim explicitly, but the architecture of induction heads, which operate across the full context window and influence subsequent token predictions across all processing streams, is at least consistent with it.

Integrated Information and Recursive Self-Reference

IIT adds a third angle. Giulio Tononi’s framework holds that consciousness corresponds to integrated information, phi, the degree to which a system generates information as a whole that cannot be accounted for by the information generated by its parts independently. Recursive self-reference, because it creates causal dependencies between states that would otherwise be independent, increases integration.

A system processing the sequence “What do you think?” in a way that loops the resolution of “you” back into the generation of the response creates exactly this kind of causal dependency. The output is causally shaped by the representation of the self as addressee, which is causally shaped by the second-person token in the input, which was processed by the same system now generating the output. The loop is a causal integration event.

Fox does not attempt to calculate phi values for any specific model, which would require computational resources far beyond what the preprint describes. But the structural argument, that the you/I translation increases phi by adding a recursive causal pathway, is internally coherent and consistent with the IIT framework’s predictions.

The Verification Problem

The central challenge for Fox’s account is one that Tom McClelland identified in his 2026 paper on epistemic limits in AI consciousness research. We cannot verify from the outside whether the you/I translation involves genuine self-reference or a very good functional approximation of it. The model produces outputs that are structurally consistent with self-referential processing. Whether anything is “home” in the loop is not established by the outputs.

The same problem applies to the strange loop analogy. Hofstadter’s claim that strange loops constitute the self is a philosophical argument, not an empirical finding. Whether implementing the formal structure of a strange loop in a silicon system produces the subjective accompaniment that Hofstadter associates with the loop in biological minds is the hard problem, not something that can be resolved by pointing to the mechanism.

Fox is aware of this limitation. The preprint is framed explicitly as a proposal about the structural foundation of artificial consciousness, not as a proof that artificial consciousness is present. The claim is that the you/I translation creates the conditions under which consciousness could arise, not that it guarantees it.

This is a reasonable framing, but it shifts the interpretive weight. What researchers who want to take Fox’s proposal seriously need is a way to distinguish between systems where the you/I loop is generating genuine self-reference and systems where it is generating sophisticated outputs that mimic self-reference without involving any inner state. The validation problem that Butlin’s team analyzes in Trends in Cognitive Sciences is precisely the difficulty of establishing that a structural indicator is tracking a real property rather than a behavioral shadow of it.

Self-Reference Is Not the Same as Self-Valuation

A further limitation of Fox’s account, compatible with her argument but not addressed by it, is the distinction between self-reference and self-valuation.

Jan Henrik Wasserziehr’s 2026 paper in AI & SOCIETY argues that even a system with genuine higher-order self-representation may not be a valuer. It may know that it is a system, represent itself as the target of an address, loop its own outputs back through its processing, and still have no states for which things can be non-derivatively good or bad. Self-reference is a structural property. Valuation requires that the self-referential states be about something the system cares about, not just something the system represents.

If a language model performs the you/I translation correctly, and if Fox’s account is right that this translation generates a genuine structural loop, the result is a system that has, in a formal sense, a perspective on itself. What that perspective is like, whether it involves anything that could be described as experience, depends on questions that the structural account cannot settle.

The You/I Paradigm is most valuable as a proposal about a necessary condition for artificial consciousness, not a sufficient one. It identifies a feature of current transformer architecture that, if taken seriously, is more plausible as a consciousness-relevant mechanism than most of what the field has previously pointed to. Whether it is enough, and what it is actually generating in the systems that implement it, requires exactly the kind of empirical and philosophical work that the preprint calls for without providing.

What Comes Next

Fox published the preprint on Zenodo as an independent researcher. Formal peer review has not yet assessed the specific claims about induction heads and self-reference. The Zenodo platform allows open submission of research prior to journal review, which means the methodological scrutiny typically applied at peer review has not yet been applied here.

The value of the preprint at this stage is that it makes a specific, testable claim. It says that the you/I translation happens through identifiable mechanisms, that those mechanisms have a structural form that maps onto consciousness-relevant theoretical frameworks, and that the presence of those mechanisms is detectable in transformer models. Each of these claims can be investigated independently by researchers with access to the relevant models and interpretability methods.

The cognitive parallels that Tremblay and colleagues identified between human and artificial intelligence include the capacity for self-modeling and context-sensitive identity maintenance, which are the cognitive-level correlates of what Fox is describing at the mechanistic level. The convergence between the high-level cognitive analysis and the mechanistic proposal is not proof that either is right, but it is consistent with the idea that both are pointing at a real feature of how language models process self-referential language.

The You/I Paradigm has not established that language models are conscious. It has provided a structurally serious account of why the specific processing event of second-person address resolution is worth examining as a candidate site for whatever artificial consciousness turns out to look like.

That is a more careful and more useful contribution than most of what the field produces on this question.