This is a public statement of where my thinking sits at this moment. It is not a manifesto. It is not a finished theory. It is a marker — so that future versions of this work can be traced back to what I actually believed, and what I did not yet know, when I started building.

The observation that motivates this work

Human-AI cognition is a coupled system.

Edwin Hutchins's foundational work on distributed cognition (Cognition in the Wild, 1995) showed that complex cognitive tasks — navigating a Navy ship across an ocean, for instance — aren't performed by individual minds working with tools to help. They're performed by the whole system: navigator, chart, compass, rangefinder, the bearings being called out, the procedures, the other crew members, the ship itself. Remove the compass and you don't have a navigator working without a key tool. You have a fundamentally different cognitive system that computes position differently. The thinking is in the configuration. The unit of analysis is the whole functional system, not the individual mind.

This is what happens when a person works with a generative AI. The thinking is no longer happening only inside the person's head. It is distributed across the person, the model, and the artifacts they produce together. And the output of that system can look indistinguishable whether the person did most of the cognitive work or almost none of it. The artifact does not reveal who did the thinking.

Prior cognitive tools externalized capability visibly. The calculator shows you what it computed. The map shows you the route. The compass shows you the bearing, but not the destination; the large language model shows you the destination, but hides the bearing. The user always knew where their work ended and the tool's began. Large language models externalize reasoning itself, and they do so invisibly. A user can read a fluent answer and feel intellectually involved because they touched the artifact, while the key reasoning transitions occurred outside them. A programmer can accept a confident solution before fully modeling its logic. The user often cannot detect, in the moment, when evaluation stopped and acceptance began.

I want to be clear about what this position is not. It is not anti-offloading. Cognitive offloading is inevitable, often beneficial, and frequently the right move — no one should compute compound interest by hand for the sake of cognitive purity. The concern is not delegation itself. It is unexamined delegation, where the boundary between human reasoning and machine reasoning becomes invisible to the human inside the loop. This is the cognitive-offloading problem named by researchers like Gerlich and Kabashkin, and it is the load-bearing concern in the recent work of UIUC's Mary Frances Phillips and Koustuv Saha. The question is not whether AI is dangerous. The question is what kind of thinker the human becomes when most of their thinking is delegable, and they cannot see, in the moment, that the delegation is happening.

The framework I am operating within

The cognitive dynamic between a human and an AI falls into one of three states. Negative-sum: the human's capacity atrophies — the system produces good output, but the person doing the work loses some of their ability to do it without the AI. Zero-sum: the AI substitutes for the human without growth or loss — the task gets done, but the human doesn't develop. Positive-sum: the AI frees the human into thinking they could not have reached alone — the human grows by working with the system.

Which state any given interaction occupies is set by the human's engagement. And engagement is precisely the variable that is invisible to both the user and any external system.

I believe most current AI interactions sit in the zero-sum or negative-sum state by default. Not because users are careless, but because nothing in the system makes the cognitive distribution visible to them. The ship's navigator can see the compass. The AI user cannot see where their thinking ends and the model's begins.

The direction this work is pointed

The dominant response to this problem locates the safeguard inside the AI itself: train models to refuse when refusal would protect the user from offloading. This is the work serious people are doing, and it matters. But it is incomplete. Refusal systems are unreliable, institutional adoption lags, and the user remains the only party present in every interaction.

Beyond the AI-system layer and the institutional layer, I am building toward a third: a user-facing reflection layer that makes the user's own cognitive engagement visible to them in real time, without forcing judgment from the system itself. The aim is not to police the user. The aim is sovereign awareness — by which I mean the user's ability to see, evaluate, interrupt, and redirect the thinking happening across themselves and the system. The user remains the active, governing party in their own cognition, with the AI serving as the instrument rather than the agent.

This is the bet. The instrument exists in early form. Whether it works at scale is an empirical question I cannot answer alone. The work, and this statement, are public so that the position is on record and the question is examinable.

— Joseph Tingling, MD/PhD
Hugonomy — May 2026

Also published on Substack: Where Hugonomy Stands — A Position, Not a Conclusion →

References

• Arditi, A., Obeso, O., Syed, A., Paleka, D., Panickssery, N., Gurnee, W., & Nanda, N. (2024). Refusal in language models is mediated by a single direction. Advances in Neural Information Processing Systems, 38. https://arxiv.org/abs/2406.11717
• Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), Article 6. https://doi.org/10.3390/soc15010006
• Hutchins, E. (1995). Cognition in the wild. MIT Press.
• Kabashkin, I. (2025). Cognitive atrophy paradox of AI–human interaction: From cognitive growth and atrophy to balance. Information, 16(11), Article 1009. https://doi.org/10.3390/info16111009
• Phillips, M. F., & Saha, K. (2026, April). Ethical guidelines for AI. University of Illinois Urbana-Champaign.
• Qazi, I. A., Ali, A., Khawaja, A. U., Akhtar, M. J., Sheikh, A. Z., & Alizai, M. H. (2025). Automation bias in large language model–assisted diagnostic reasoning among physicians trained in AI literacy — A randomized clinical trial. NEJM AI. https://doi.org/10.1056/AIoa2501001
• Zhang, Y., Li, M., Han, W., Yao, Y., Cen, Z., & Zhao, D. (2025). Safety is not only about refusal: Reasoning-enhanced fine-tuning for interpretable LLM safety. arXiv preprint arXiv:2503.05021. https://arxiv.org/abs/2503.05021