There’s a meaningful difference between asking a chatbot to draft something for you and using it to create something with you.
This is the operating model: how to keep the speed of generative AI without accepting a more brittle legal mind.
In metallurgy, you do not reject homogenization. You use it, then you design microstructure: grain size, precipitates, composites. You introduce controlled heterogeneity because that is where toughness comes from. The legal analogue is not “less AI.” It is AI as a furnace and a sparring partner—not as the finished part.
Rule 1: Discussion Engine, Not Author
Thomson Reuters puts the distinction in language boards can adopt as policy: use AI as a discussion engine—a sparring partner that pressure-tests ideas and pushes past the first plausible answer—rather than as a drafting appliance. Efficiency-only deployment leaves the creative upside on the table and delivers the dulling effect with nothing to show for it (Thomson Reuters Institute).
For legal work, “sparring” means prompts of this family—not “write the memo”:
- Attack this thesis. What facts would make it fail?
- Give me the supervisor’s uncharitable reading.
- What theory is missing because it is unfashionable?
- Where is this ‘market standard’ clause misaligned with these incentives?
- Steelman the other side; then tell me why our position still holds.
The human remains the author of the core proposition. The model is cross-examination. That preserves the cognitive friction that forms an original viewpoint (Tech Research Online). It is also the only posture that is consistent with non-delegable professional responsibility.
Never use a model to generate the core legal thesis on a material matter. Use it to try to kill the thesis. If the thesis survives, you have a position. If it does not, you found the issue before the board, the court, or the supervisor did.
Rule 2: Stage and Expertise, Not Blanket Rollout
INFORMS Information Systems Research work on generative AI in creative process finds a double edge. In ideation, AI helps across expertise levels by reducing fixation. In implementation, novices may still benefit while experts can lose efficiency without gaining creativity, because the tool fights trained professional routines (Hou et al., ISR).
Map that onto a legal function:
| Stage | Example | AI posture |
|---|---|---|
| Ideation | Issue-spotting, alternative theories, “what could this be?” | Allowed, even encouraged—then human ranking |
| Implementation (junior) | First outline, public-source summary, checklist against a known framework | Allowed with mandatory review |
| Implementation (expert) | Opinion, filing, negotiation strategy, board recommendation | Human-led; model as critic, not drafter of the operative text |
| Decision | What we tell the board / the regulator / the court | No model authorship of the decision itself |
The Goldilocks finding belongs here as a usage cap: moderate collaboration beats both neglect and saturation. “We use AI for everything” is not a maturity statement. It is often a Goldilocks violation.
Rule 3: Pilot Mindset, Not Autopilot
HBR’s workslop research is a quality-culture paper wearing a technology headline. Low-effort, high-polish artefacts shift work downstream and tax the people who still have to think. The organisational antidote is a commitment to task quality and to human–AI dynamics that are actually collaborative (HBR).
Pilot mindset, in one sentence: the lawyer (or legal director) is accountable for the destination; the model is crew. Autopilot mindset: the fluent output is the work. Pilots still use instruments. They do not confuse the instrument reading with the decision to land.
Practical marks of pilot mindset in a legal team:
- The signer can explain the recommendation with the screen off.
- Alternatives that were considered are written down, including at least one the model did not offer.
- AI use is disclosed internally on material work (not as theatre—as a quality signal).
- Seniors refuse workslop the way they refuse an unverified citation.
- Time saved on drafting is partly reinvested in challenge, not entirely returned to the matter mill.
A One-Page Framework Legal Board Members Can Adopt
Ask management to implement, and then evidence, the following. It is short enough for a policy appendix.
1) Classify the work
- Green: brainstorm, public guidance, outlines, devil’s advocacy.
- Amber: internal drafts that will be fully rewritten by an accountable lawyer.
- Red: opinions, filings, privileged strategy, board decisions, regulator-facing positions—model may critique, not originate.
2) Require a human thesis
On amber and red, a thesis paragraph exists before the first generative pass.
3) Require a diversity check on material matters
A short map of distinct legal characterisations. If they all sound like the same model, start over.
4) Separate panel independence from convenience
Outside counsel do not begin from the internal AI draft unless the board (or GC) has accepted that they are not independent on that issue.
5) Reform what you reward
If you only measure cycle time and polish, you will buy homogenization at a discount. Reward the uncomfortable memo that changed the decision. Reward the junior who killed a fluent but wrong framing. Thomson Reuters is right that the highest-leverage fix is upstream of the tool: what the institution selects for.
6) Train for friction, not for prompts
Literacy under the EU AI Act is not “how to write a clever prompt.” For legal staff it is: when offloading has started, why fluency is not accuracy, and how to use a model as cross-examination.
Questions for the Next Legal / Risk Committee
- Do we have a green / amber / red map for legal AI use, with named owners?
- Can we show, on the last three material matters, a thesis that existed before the tool, and a theory the tool did not suggest?
- How do we refuse workslop without humiliating people into hiding their AI use?
- Is any panel firm instructed to disagree, not to tidy?
- If we gained 40% drafting speed, where did the recovered time go—calendar, or judgment?
Uniformity Is Not Strength
A legal board that hears one statistically likely answer, faster, from a more similar set of advisors, has not automatically improved oversight. It has homogenized the melt.
Use the engine. Design the microstructure. Keep the messy middle. Make the model argue with you. Stay the pilot.
That is how a legal member of a board captures productivity without paying the hidden creativity tax in the only currency that matters: the quality of independent judgment.
If you want this framework as a board policy appendix, a legal-function playbook, or a half-day session for directors and GCs, I teach and advise on exactly this. Contact me.
Relevant Sources
- Organizations Are Misdiagnosing What’s Killing Their Innovation — Thomson Reuters Institute — https://www.thomsonreuters.com/en/institute/articles/feature-misdiagnosing-whats-killing-innovation
- The Double-Edged Roles of Generative AI in the Creative Process — Hou et al., Information Systems Research — https://doi.org/10.1287/isre.2024.0937
- AI-Generated “Workslop” Is Destroying Productivity — Harvard Business Review — https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity
- Unlocking Creativity with AI: The Goldilocks Effect — Huang (2025) — https://pubmed.ncbi.nlm.nih.gov/41051839/
- Is AI Making You Less Creative? — Tech Research Online — https://techresearchonline.com/blog/is-ai-making-you-less-creative/
- Article 4: AI literacy — EU AI Act Service Desk — https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-4