If ten people reach a competent legal answer faster while exploring fewer genuinely different answers, productivity has increased while something else has disappeared.
That “something else” is the diversity of legal thought a board actually needs: dissenting theories, uncomfortable facts, and the outlier position that turns out to be right.
The Metallurgy Point Boards Should Steal
Metallurgists homogenize alloys to remove harmful segregation. Uniform composition is useful. It is not the same thing as strength.
Strength comes from microstructure: grain boundaries, precipitates, controlled defects, and the friction of phases that do not perfectly agree. A perfectly uniform piece of metal can be easier to machine and easier to break. Toughness, the ability to absorb stress without fracturing, depends on heterogeneity you can still control.
Legal judgment works the same way. Standardized, fluent, “market-standard” output is not a robust legal position. A board that hears one statistically likely answer from every advisor is not a more informed board. It is a more uniform one.
What an LLM Optimizes For (Plain Language)
A large language model does not “reason” in the sense a lawyer or a director reasons. It predicts the next token that is most likely given its training data, the prompt, and the conversation so far. That is a statistical operation across overlapping knowledge bases: statutes, judgments, textbooks, blogs, prior memos, and a great deal of confident-sounding commentary that is not law.
Two consequences follow, and both matter at board level.
First, conventional answers become extraordinarily cheap. If the question is common—“What should a GDPR Article 32 programme look like?” or “What are typical DORA ICT third-party clauses?” the model will converge on the modal answer in its training distribution. That answer will often be competent. It will also be the answer everyone else is getting.
Second, convergence is not a bug. When many people query the same family of models with overlapping prompts, they are sampling the same distribution. The outputs will rhyme. They will use the same risk language, the same three residual risks, the same “balanced” recommendation. Fluency hides the collapse in variance.
This is the mechanism: faster answers, fewer answers.
The Productivity Number That Misleads Legal Functions
Research on generative AI and creative work keeps finding the same trade-off. After adoption, creators can post roughly 50% more work, while the rate of genuinely frontier-expanding ideas per artefact falls by about 2.1 percentage points. Volume rises; per-piece novelty drops. The absolute count of new ideas can still rise because you produce more items—until the organization starts treating volume as quality (BU Questrom synthesis).
Translate that into a legal function.
- More first drafts of board papers, DPIAs, litigation hold notices, and vendor assessments.
- Faster turnaround on “what does the law say” questions.
- A shrinking set of distinct legal theories, deal structures, and challenge questions reaching the board.
A GC who reports “we are 50% faster” is not lying. The legal member of the board should still ask: faster at producing what, and how many genuinely different positions did we consider?
Why This Is a Board-Legal Problem, Not a Tech Problem
Legal members of boards sit at a specific intersection: you are accountable for the quality of legal judgment the board relies on, not for the word count of the papers. Three duties collide with statistical homogenization.
Independent judgment. Directors must exercise their own judgment, not rubber-stamp a fluent consensus. If management, in-house counsel, and outside counsel have all “sense-checked” the same model, the papers will agree with each other for a reason that has nothing to do with the facts.
Duty of care in high-impact decisions. Privilege, verification, and hallucination risk are already well documented. Homogenization is the quieter cousin: the advice is not fabricated; it is merely the same advice. You can verify every citation and still miss the unconventional defence, the awkward jurisdiction, or the structure that does not appear in the training mean.
Challenge culture. A legal director’s value in the boardroom is often the question nobody wanted to put in the pack. Models are trained to be helpful and complete. Helpful, complete text is the enemy of the incomplete, irritating question that should have been asked.
Uniform legal output feels like control. In metallurgy terms, you have homogenized the melt. You have not tested toughness.
What “Fewer Answers” Looks Like in Practice
Watch for these patterns in board packs and legal reporting:
- Risk sections that could be swapped between two unrelated agenda items without anyone noticing.
- “Options” that are three phrasings of the same recommendation.
- Outside counsel memos that read like the in-house draft, which reads like last quarter’s paper, which reads like a chatbot.
- Regulatory strategy that matches the industry’s public talking points because those talking points dominate the training data.
- No recorded dissent, not because the room agreed, but because the first plausible answer arrived too quickly to leave time for a second.
None of this requires the model to be wrong. The conventional answer is often right. The governance failure is that the distribution of answers has collapsed, so the board cannot see the tail—the rare, fact-specific, legally material outlier.
Questions Legal Board Members Should Ask This Quarter
- Where do we use AI to produce legal or governance text (board papers, policies, opinions, filings), and who is accountable for the thesis, not the polish?
- How many distinct legal positions were considered on the last material matter—and can we show them, not just the winner?
- Do our outside firms use the same models as we do, on the same templates, against the same “market standard” prompts?
- What would look like quality if we stopped counting pages and turnaround time?
Want this translated into board questions, legal-function controls, and a challenge culture that survives AI drafting? I deliver board-level courses and consult on AI governance and legal risk. Contact me.
Relevant Sources
- How AI Works Under the Hood: Statistics, Not Magic — Goldmanmalka — https://goldmanmalka.com/how-ai-works-statistics-not-magic/
- Generative AI and the Future of Creativity: Trade-Offs, Risks, and Opportunities — Insights@Questrom (Boston University) — https://insights.bu.edu/generative-ai-and-the-future-of-creativity-trade-offs-risks-and-opportunities/
- Generative AI Profile (NIST AI 600-1) — NIST — https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence
- AI RMF 1.0 (NIST AI 100-1) — NIST — https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10
- Homogenizing Effect of Large Language Models on Creative Diversity — Moon, Green & Kushlev, Computers in Human Behavior: Artificial Humans (2025) — https://doi.org/10.1016/j.chbah.2025.100207