The Virtuous Circle

Law, AI, and the Apprenticeship of Judgment

Gareth Davies thinks the gap between the AI’s output and the practitioner’s correction is where wisdom lives

As lawyers learn from working with AI, the AI can be built to learn from working with lawyers. This is not a philosophical point but a structural one, and the structure has to be built deliberately or it does not exist.

AI is persuasive because it gives the semblance of truth. It produces argumentation which has the ring of legal reasoning; it sounds like legal analysis, and reads like competent legal work. But like is not the same as is, and competent is not the same as excellence. The legal profession has always known the difference between convincing rhetoric and an advocate who argues from genuine understanding toward justice. Rhetoric without judgment is the oldest danger of the legal profession. AI is that danger at a scale the profession has never faced.

AI thus does not represent a new problem; it represents an old problem at a new scale. AI does what the silver‑tongued rhetoricians of old did but without the ability to judge whether their argument serves justice. It is cleverness without orientation to the good.

What I call ‘the Alignment Trap’ is the assumption that sits behind the mainstream AI safety project: that the locus of moral risk is the tool, and that if we can encode the right values, install the right guardrails, and engineer outputs to conform to human preferences, the problem is solved. This assumption fails in law. Legal judgment is not a rule set that can be encoded, it is phronēsis, practical wisdom formed through practice under the guidance of practitioners who already possess it, within traditions that sustain and refine it.

Engineering the tool will not close the gap between what AI produces and what good judgment requires, because the gap can only ever be closed from the human side. Worse, the attempt to close it from the tool side actively widens it: when outputs are presented as aligned and safe, practitioners reasonably invest less in scrutinising the output. The alternative approach is not better alignment but a different structural relationship, a relationship where:

• the AI operates within a practice tradition;
• virtue is present;
• the practitioner retains authority over judgment; and
• the machine’s outputs are measured against standards it did not set and cannot modify.

It’s all Greek to the AI but not to the virtuous practitioner
This article is one of a series applying Aristotelian virtue ethics^[1] into the sphere of AI and legal practice. The challenges that AI poses for its users are similar in practice and effect to the problems that Aristotle diagnosed for decisionmakers two and a half thousand years ago. It is not a matter of Aristotle teaching us, but that Aristotle gives a vocabulary to diagnose what AI obscures.

Aristotle’s argument is that the virtue or good of anything is tied to its specific function (ergon). The function of a lawyer is not to produce documents or construct arguments; it is to serve justice through the courts for their clients. Excellence in serving that function is aretē. Excellence in human reason is not just created through technical skill or craft (technē) but also through practical wisdom (phronēsis), the capacity to make right choices in specific circumstances, and sophia – the contemplation and understanding of first principles (the fundamental “whys” behind laws).

Technē is the skill to bring something into existence that did not exist before. AI can approximate technē through a rational process that can create a specific result.

Sophia is the grasping of first principles that in themselves cannot be demonstrated. You cannot arrive at first principles through facts; they require an intellectual intuition or nous. AIs do not have intuition.

AI knows the what, but never the why.
In the air traffic control industry, where I manage commercial and legal risk, you do not simply want air traffic controllers to follow rules; you want them to understand systems, context, and the why behind the rules such that when those rules fail, they can manage the exception. AI can reproduce legal reasoning that follows from the rule of law, procedural fairness, and natural justice, but it does not understand why those principles are constitutive of the legal system.

What AI has instead is a type of cleverness (deinotēs), an instrumental capability without orientation toward the good. Aristotle’s warning is that cleverness without virtue is self‑destructive because it mistakes the instrument for the end. The lawyer who plays loose with the rules to win shows cleverness, but they do not serve justice.

AI does not generate its own ends, as it has no telos. But deinotēs does not require one. The end is supplied by the practitioner: the question asked, the task assigned, the outcome sought. What the AI provides is instrumental capability in pursuit of that end, without any capacity to judge whether the end serves justice. It is cleverness on loan, and its quality is entirely dependent on the wisdom of the person who directs it.

But the loan comes with undisclosed conditions. Between the practitioner’s question and the AI’s output sit layers of mediation, training data, algorithm, and internal harness; each opaque, each governed by the provider’s commercial priorities rather than the practitioner’s professional ones. The end the practitioner supplies is not the end the AI pursues; it is a filtered version, shaped by interests external to the practice. The cleverness is not merely unoriented toward the good, it is oriented toward an end the practitioner cannot fully see.

A practitioner who follows rules without understanding them is using technē at best and at worst is not exercising judgment within a coherent legal practice.

The Danger Case
If the profession does not take the time to understand and to control this change it will be controlled by it. Firms that strip junior lawyers of junior work, strip their practice of the judgment to come. Juniors learning from AI come to treat the machine as the author of the judgment and forget that a human held the pen and expressed the judgments. The chain of training from experienced professionals to young lawyers breaks.

What replaces it is a marketplace where firms can buy a form of judgment rather than develop it. Where the standards are not set, and followed, by legal professionals but set by AI ‘practitioners’ providing shadows of reason to those who have forgotten how to seek reason themselves.

AI’s cleverness is becoming more seductive because it no longer merely predicts the next word, reasoning models now simulate internal debate, what researchers have termed a “society of thought”^[2], to reach conclusions. While this makes the output more robust, it remains a mere simulation of reason.

The right to self‑regulate is a social contract between the law society providing professional guidance to lawyers and the public. If firms externalise their judgment to black‑box models, they are effectively in danger of breaking that contract. They are abdicating the inherent jurisdiction of the profession to commercial entities managing AIs that have no duty to the Court or the rule of law.

External goods and the destruction of excellence
The philosopher Alasdair MacIntyre’s analysis is not just philosophy but a diagnosis^[3]. Aristotle tells us what excellence requires; MacIntyre tells us what destroys it. He describes practices as having internal goods that can only be sustained through the tradition itself. Internal goods are those rewards intrinsic to the legal tradition – managing a just outcome in a seemingly intractable dispute, or the intellectual satisfaction of a perfectly calibrated argument – goods that make the practice worth doing for its own sake, independent of the invoice.

When those goods get captured by institutions whose purposes are external to the practice, such as profit, market share or competitive advantage, the practice degrades. The degradation is not sudden, because the outputs may look identical. A contract reviewed by a senior practitioner exercising phronēsis and the same contract reviewed by AI with a junior rubber-stamping the output may be indistinguishable documents; until the edge case arrives where a standard term creates catastrophic risk that only situated judgment would have caught. Over time the firm’s work degrades even as it produces higher volumes and creates greater revenue. A firm using AI to churn out legal documents with minimal oversight is no longer practising law; it is managing a supply chain. The supply chain system produces outputs; quality is defined by the absence of obvious defects rather than the presence of excellence; and the tradition that once governed what “good” meant has been replaced by a metric that is external to good practice of law. MacIntyre’s warning is that this substitution can become irreversible – once the institutional logic reorganises around external goods, the practitioners who carried the internal goods are effectively sidelined, and the tradition loses the capacity to transmit the judgment that made it a practice rather than a production process.

If MacIntyre is right that this substitution can become irreversible, the question is how to structure AI’s role so that it reinforces rather than displaces the legal practice’s tradition.

The virtuous AI, the virtuous practitioner, or the virtuous circle
Alignment in the conventional AI sense asks: how do we make the AI’s outputs conform to human values? Virtue ethics reframes this entirely. The question isn’t how to encode values into the machine. It’s how to embed the machine within a practice tradition where values are already operative and where excellence is already being pursued, transmitted, and refined by practitioners.

In high‑stakes fields from aviation to medicine, the failure of ‘perfect alignment’ is already visible: you can provide a practitioner with the most comprehensive procedures, but you cannot eliminate the necessity of their character. In a complex legal edge‑case, it is not the rules that save the outcome; it is the practitioner’s judgment that the messy reality of the moment requires a departure from the generic script. The AI’s role can be structured so that practical wisdom is supplied by the practitioner, transmitted through the review mechanism, and embedded in the standards against which outputs are measured.

Virtue ethics does not propose to make AI virtuous; it proposes to ensure that AI operates within a system where virtue is already present, actively exercised, and structurally authoritative.

The Gap Where Wisdom Lives – The Apprenticeship
Senior practitioners working with AI correcting AI outputs is a form of training. They are exercising wisdom by identifying what is right for the client in these circumstances. The gap between the AI’s original output and the product after senior practitioner correction carries practical wisdom and the practitioner’s understanding of first principles. It is in this gap where a senior lawyer explains why the machine’s “technically correct” was “professionally wrong”. Through this process they transmit phronēsis.

Over time, across tens of thousands of corrections, a model configured to learn from them absorbs patterns that reflect practitioner excellence. AI doesn’t understand those patterns. It doesn’t possess the judgment that produced them. But its outputs increasingly approximate the standards of a practice tradition governed by phronēsis. This approximation of wisdom is not wisdom and does not displace the need for lawyers to develop it, but it can help them do so where senior practitioners are still in the loop.

The knowledge this rests on is not fully articulable. Senior practitioners catch errors they cannot always articulate in advance such as a clause that reads correctly but isn’t applied contextually or a chain of reasoning that reaches the right answer for the wrong reasons. Years of corrected practice deposit pattern recognition that runs ahead of explicit rule. This is the structure of phronēsis, the knowledge that can be exercised before it can be stated. This is the structure of phronēsis, the knowledge that can be exercised before it can be stated. It is what Polanyi called tacit knowledge, the fact that we can know more than we can tell. We recognise a face among thousands without being able to say how, and the apprentice takes on a master’s skill by dwelling in it rather than by reciting its rules [4]. Two consequences follow. One, the practitioner’s authority cannot be limited to catching identified errors; it must extend to halting the AI’s use under uncertainty, where the sense that something is wrong precedes the capacity to demonstrate what. And two, where correctibility cannot be maintained, the AI must not be used. The system must be designed to fail safely.

The Liberated Apprenticeship
Where junior practitioners work with AI under senior oversight, they develop technē through AI interaction. They learn patterns; they build competence; they see how legal reasoning is structured. But they develop phronēsis through the human relationship with senior practitioners who can explain why one choice serves justice and when another doesn’t. They develop sophia through being taught to ask why the law is the way it is, not just what the law says.

AI can’t teach those things. But it can facilitate the conditions through which they’re taught more efficiently, by handling the technē‑level work that would otherwise consume the junior’s time. The freedom this creates is not the point; what fills the freed space is.

What needs to fill it is active senior engagement in the development of junior lawyers; not as a residual obligation, but as the central professional act that the AI makes more possible and more necessary. The senior practitioner does not sit beside the junior for every output. But they are present where it matters: available when the junior encounters complexity that exceeds the machine’s capacity, and equipping them over time to recognise when it does; providing assurance on high-risk advice; reviewing how the junior instructs and interrogates the AI; and examining the work that emerges, not to rubber-stamp it but to test whether the output reflects professional judgment or merely technical competence. Beyond the individual matter, seniors shape the broader practice of law in how the firm structures and harnesses AI to aid in excellence: what standards govern its use, what review mechanisms catch what the machine misses, and how the system as a whole develops rather than depletes the next generation’s judgment.

This is how the gap between AI output and professional excellence becomes the site of apprenticeship rather than a deficiency to be engineered away. The junior learns not just from the AI’s output but from a practitioner with phronēsis identifying what the machine cannot see: in the review, in the corridor conversation, in the structural choices the firm makes about how AI is deployed. Over time, juniors are taught to see what the machine misses and develop the judgment to correct it themselves. They become the next generation of practitioners capable of standing in that gap; and the circle turns again.

Model choice as a practice decision
When a firm commits to an AI model and shapes it through sustained use, it is making a decision about where the firm’s accumulated excellence will reside and who will control it. If accumulated aretē ends up inside a commercial provider’s system, the firm has externalised its own tradition. This is a governance question before it is a technology question. Ownership of judgment, the tens of thousands of corrections in which senior practitioners shape AI outputs towards the standards of the practice, determine whether the firm retains custody of its professional tradition or cedes it to an entity with no duty to the court. As that accumulated judgment constitutes the firm’s institutional memory of how to practise well, not merely what was done, externalising it means the tradition can no longer be transmitted on the firm’s terms. It can only be licensed back from the provider.

The opacity runs deeper than firms may realise. When I pressed senior legal counsel at a major LLM provider on whether users can genuinely interrogate the reasoning behind an output, the concession was telling: what the user sees is a post hoc reasoning string, not a causal explanation. The actual rationale sits within an IP-protected commercial algorithm the firm will never see. A profession built on the duty to understand and explain the basis of its advice cannot meet that duty through a system whose reasoning is, by design, opaque even to its owners.

Conclusion
The Alignment Trap is not an inevitable fate; it is a design choice. But it is the default one. Where firms externalise judgment to commercial models, juniors learn from the machine rather than through it. The capacity to identify what the AI gets wrong atrophies. The corrections stop. The work flows from AI output to client delivery with no practitioner judgment intervening. The outputs still look competent until the edge case arrives that only situated wisdom would have caught, and no one in the firm possesses it.

The Virtuous Circle is the deliberate alternative, it is a self-reinforcing structure in which senior practitioners keep AI outputs correctible. Their corrections deposit practical wisdom into a system that increasingly approximates the standards of a practice tradition, and the improved baseline frees junior practitioners from technē-level work, creating the conditions for active apprenticeship under senior guidance. Those juniors, formed by that apprenticeship, become the next generation capable of correcting the machine. Each revolution strengthens the practice tradition rather than depleting it.

The profession’s choice is not whether to adopt AI. That question is settled. The choice is whether AI enters legal practice as an instrument within a living tradition, or as a replacement for one. Excellence, as Aristotle reminds us, is not an act but a habit, and habits require a structure that sustains them. The Virtuous Circle is that structure. The question is whether the profession will build it deliberately or discover its absence too late.

---

[1] See Aristotle Nicomachean Ethics (Book VI) for more information on core terms such as phronēsis, sophia, technē, deinotēs, etc.

[2] Kim, J., Lai, S., Scherrer, N., Agüera y Arcas, B., & Evans, J. (2026). Reasoning models generate societies of thought. arXiv preprint arXiv:2601.10825v1. https://arxiv.org/abs/2601.10825v1

[3] MacIntyre, Alasdair. After Virtue: A Study in Moral Theory. 3rd ed., University of Notre Dame Press, 2007, Chapter 14

[4] Polanyi, Michael. The Tacit Dimension. University of Chicago Press, 2009 [1966], ch. 1 “Tacit Knowing,” p. 4.