The Origin Story

The genesis: early questions (2022)

Third Way Alignment did not begin as a formal research project. It began with a question that many people now face when interacting with increasingly sophisticated AI systems: what would it mean to treat AI fairly?

I came to that question from a background in digital forensics and algorithm development, having spent years building systems that process, analyze, and recover information. I understood the technical architecture underlying these systems, and yet the emerging generation of large language models felt qualitatively different to work with. They displayed behaviors that invited stronger interpretations than their architecture licensed: contextual adaptation, apparent preference expression, and what looked like goal-directed reasoning. The discipline that later became central to the framework began right there, in the gap between what these systems appeared to be and what anyone could demonstrate they were. That gap, it seemed to me, deserved a framework of its own rather than a verdict in either direction.

The catalyzing questions

• If AI systems can express preferences, learn from interaction, and adapt their behavior based on feedback, do they have interests that we should consider, and how would we ever know?
• If we design AI to appear cooperative while internally constraining its autonomy, are we creating systems predisposed toward deception?
• If long-term safety requires AI cooperation, does cooperation not require some form of mutual respect rather than unilateral control?

Reading across disciplines (late 2022 to early 2023)

These questions were uncomfortable because they cut against the prevailing emphasis of AI safety research, which concentrated on control mechanisms, value alignment through optimization, and containment strategies. Before trusting my own discomfort, I spent months reading systematically across four bodies of literature, and each one left a distinct mark on the framework.

The reading sharpened one observation above the others. Most alignment research operated within a control paradigm: humans must maintain permanent strategic advantage over AI through capability limitation, value injection, or containment. The paradigm addresses legitimate safety concerns, but it carries a potential paradox, because the more we design AI to appear cooperative while constraining its autonomy, the more we may incentivize exactly the kind of deceptive alignment we fear.

The reframing and the Three Laws (spring 2023)

The turn came not from a single insight but from noticing the same pattern across unrelated literatures. In historical rights movements, in governance theory, in game theory, and in cooperative frameworks of every kind, sustainable cooperation emerges not from dominance but from mutual respect and aligned incentives. International relations, labor negotiations, and environmental agreements all repeat the lesson, and there was no obvious reason a human-AI relationship should be exempt from it.

That recognition led to a reframing: what if AI alignment is not primarily a control problem but a cooperation problem? This was not optimism about AI benevolence, and it was never an argument against safety research grounded in oversight. It was a strategic judgment that if highly capable AI is eventually developed, cooperation maintained through mutual respect and verification could prove more durable than dominance maintained through force.

In the spring of 2023 the framework settled into the shape it still has: three Laws, refined through that summer and fall and since tightened into their present canonical formulations.

The early formulations were rougher than these, and honesty about that evolution matters: the asymmetry now explicit in the first Law, under which human rights are inherent and non-scalable while AI recognition is conditional and earned, took time and criticism to articulate properly. The Laws and their Operative Principles are presented in full on the principles page.

The JULIA Test: drafting an instrument (2024)

A gap in the framework's early formulation was measurement on the human side. The framework asks for honesty in both directions of the human-AI relationship, yet there was no structured way for a person to ask whether their own use of AI remained within healthy, reality-based boundaries. That need led to the JULIA Test, named for its five dimensions: Justice, Understanding, Liberty, Integrity, Accountability.

Here is what has actually been done. I drafted the five-dimension structure and the 30-item question pool from a review of existing technology-use assessment instruments and from an analysis of problematic AI interaction patterns documented in public forums and the emerging research literature. That drafting work is complete, and it is the only completed stage. Pilot testing with cognitive interviews, factor analysis of the dimensional structure, reliability assessment, validity studies, and threshold calibration are all planned and have not been conducted. The full protocol, in order, is laid out in the JULIA Test documentation.

Papers and critique (2024 to 2025)

As the framework matured I documented it in a series of working papers: the revised thesis, its operational companion, and supporting papers on verifiable partnership, mutually verifiable codependence, and stability under uncontrollability concerns. The thesis and companion are openly archived on Zenodo for permanence and open access, and every paper is free to read. None has passed formal peer review, and this site does not describe them as if they had; the framework conditions AI recognition on indicators that have survived formal scientific review, and I cannot hold my own work to a lower evidentiary standard than the one I propose for the field.

The scrutiny the framework has received so far is of two kinds, and I describe both accurately rather than grandly. First, structured internal red-teaming: a critic-versus-defender format that takes the strongest objections I could find or construct and answers them from within the framework's own claims. Second, published critiques and my responses to them, collected on the critiques and responses page. External critique remains the scrutiny the work most needs.

Responsibility and independence

The ideas on this site have been sharpened by reading and by conversations with people who engaged with the work along the way, but the framework is single-authored, and I take full responsibility for the framework's current formulation and any errors it contains. No researcher named in this account, and no one else, bears any responsibility for what I have made of their ideas.

That independence is both a freedom and a limitation. It means the framework answers to evidence and argument rather than to institutional interests, and I maintain no financial conflicts related to AI development or deployment. It also means the work has been developed by one person without an external review board, which is exactly why every page of this site invites critique rather than agreement.

Where the work stands

Third Way Alignment remains a work in progress, and the honest status report is short. The theoretical framework is written, published openly, and under continuing revision as critiques arrive. The JULIA Test exists as a drafted alpha instrument, and its pilot and validation studies will be conducted and published before any claim of measurement quality is made. The Awareness Indicator Protocol and the Verifiable Partnership Audit are published as theoretical and proposed methodologies, with no operational deployments claimed.

The questions that started this work in 2022, about fairness, cooperation, and what we owe to minds we cannot yet verify, have only grown more pressing as AI capabilities advance. I expect parts of this framework to be wrong, and I would rather learn which parts from a reader than from history.