There’s been a lot of buzz lately about AI policies and whether they should prohibit or encourage generative AI use among legal practitioners and law students. Many policies land at one pole or the other. Early on, the restrictive end dominated, with rules built around preventing data exposure and warning about hallucinations. We’re still seeing some, most notably UC Berkeley School of Law’s much-debated policy, which, while not an outright ban, is framed almost entirely around what students may not do. The word “prohibit” appears eleven times in a two-page document.
Part of what makes Berkeley’s approach so newsworthy is that it cuts against the current. More and more policies have now swung toward acceptance or even encouragement, especially at law firms, where leadership increasingly asks lawyers to incorporate AI into their practice and market the firm’s AI capabilities. Restrict or encourage, permit or prohibit: that’s the axis most policies argue along.
A new paper from David Simon suggests that’s the wrong axis altogether. Simon is a partner at Foley & Lardner and an adjunct professor at UW Law School, and in Managing GenAI’s Uneven Performance: A Governance Framework for Law Firms, he reframes the question. The thing to govern isn’t whether AI is used, or even what it produces. It’s what cognitive function the AI is replacing. This is the key question policymakers and educators should be considering.
Whether-or-not-to-use genAI is a false binary. The same lawyer drafting the same brief might use AI in ways that are generally safe and ways that are genuinely dangerous, sometimes within the same hour. A policy pitched at the level of permission or prohibition can’t tell those apart. And while Simon writes for firms, the logic lands just as squarely on legal education, where students need to learn not just how to use AI but when to set it aside and reason on their own.
Why deliverable-based rules fail
Simon’s sharpest practical point is that classifying whole work products as AI-permitted or AI-prohibited cannot work, because legal work isn’t built that way. A Supreme Court brief is a composite of dozens of smaller tasks, some mechanical, some judgment-laden. No firm would let a lawyer prompt a chatbot to draft the brief wholesale, yet that same brief might legitimately use AI to map case law or pull procedural histories. Treat the brief as one undifferentiated unit and you’ll block the safe uses and miss the dangerous ones. The same holds for legal education.
His framework instead sorts AI use into seven recurring modes by cognitive function, with governance controls calibrated to each. They run from purely mechanical retrieval at one end to professional judgment at the other:
| Mode | Description | Typical Examples | Governing Rule |
|---|---|---|---|
| 1. Retrieve / Extract | Mechanical location and extraction of information from defined sources. GenAI locates material that already exists and returns it according to instructions. Purely mechanical—no new content is created. | Identifying cases; pulling quotations; listing citations; compiling procedural histories; extracting dates or regulator statements | Permitted with verification. Output must link back to identifiable sources. Lawyer remains responsible for accuracy. |
| 2. Organize / Map | Structuring retrieved information into organized formats. GenAI arranges dates, events, or holdings in logical order (e.g., chronological timelines, comparison grids). Purely mechanical—no new content is created. | Creating timelines; tables; holdings grids; cross-references; issue matrices | Permitted. Review for completeness and correct categorization. No alteration of substantive meaning. |
| 3. Summarize / Synthesize | Condensing source material into key points, themes, or narratives. Introduces selection risk: the model chooses what to emphasize, include, or omit. | Case summaries; factual narratives; investigation synopses; background decks | Permitted if source-grounded and reviewed. Independently read and verify for accuracy any authority cited. No new facts; no evaluative conclusions; verify nuance and omissions. |
| 4. Generate Candidate Material | Exploratory generation of options to expand possibilities and accelerate iteration. Output exists to suggest, not to resolve issues or supply text for verbatim adoption. | Draft outlines; issue lists; argument trees; alternative framings; counterarguments; developing and writing arguments for briefs or other written work product | Internal use only. Clearly provisional. Must be substantively rewritten by a lawyer before external use. Never filing-ready. |
| 5. Edit / Rewrite for Clarity | Refining existing lawyer-authored text for readability and structure without altering substance. GenAI tightens, restructures, or simplifies language. | Tightening prose; restructuring paragraphs; plain-language rewrites; formatting consistency | Permitted for internal drafts. Lawyer must confirm no shift in meaning or introduction of new assertions. |
| 6. Critique / Stress-Test | Adversarial “red team” function: surfacing blind spots, weaknesses, adverse authority, and counterarguments before regulators or opposing counsel do. Confined to critique rather than judgment, it strengthens work by expanding perspectives considered. | Identifying weaknesses; missing issues; adverse authority; regulator or opposing counsel perspectives | Encouraged as a supplement to lawyer judgment. Treat as issue-spotting aid, not authoritative evaluation. |
| 7. Evaluate or Decide | Expressions of professional responsibility requiring contextual judgment: determining materiality, assessing settlement value, choosing litigation strategy. | Materiality determinations; disclosure judgments; litigation strategy; negotiation posture; probability assessments | Not permitted as a substitute for professional judgment. This is a categorical prohibition. AI may inform background analysis but may not make the call. |
Modes 1 and 2 sit at the safer end: mechanical work that can be encouraged with verification. The middle modes carry rising risk as the AI starts making choices about emphasis, wording, and argument, so each calls for a lawyer’s review. By Mode 7, the boundary is no longer a matter of degree. Determining materiality, assessing settlement value, deciding strategy: these are the lawyer’s call, full stop. AI can inform the background analysis, but it cannot make the decision. That is a categorical prohibition, not a control to be calibrated.
What makes the framework useful is that it gives educators and policymakers, at firms and law schools alike, something to govern besides yes or no. A rule that asks which cognitive function is being handed off can permit the retrieval that saves hours while drawing a hard line at the judgment that defines the profession. Policies stuck on the permit-or-prohibit axis can’t do that. As more firms and schools draft and revise their AI rules, this is the question worth putting at the center.
Simon also published a shorter version of this paper for the Thomson Reuters Institute, worth a look if you want the framework without the full empirical treatment.
Disclosure: This post was developed by the author with organizational and drafting assistance from Claude AI. All content was reviewed and refined for accuracy.