Law Isn’t War, but the Same Rules of Trust Apply

Whilst having a poorly 11-month-old who wanted holding all night, I found this paper and had a read. It looks at how the military is designing human-AI teams that adapt under pressure. What struck me is how relevant the ideas are to legal work. Law may not be war, yet the dynamics are familiar: shifting information, uneven risk, and the need to combine human judgment with machine scale.

Autonomy is a dial, not a switch

A lot legal AI tools treat "human in the loop" as an on/off choice. The military model shows autonomy should be adjustable, rising or falling with confidence, shared awareness, and uncertainty. If error rates creep up or explanations weaken, autonomy should step back. If performance is stable and the lawyer is comfortable, it can move forward.

The paper defines six variables that shape the balance:

Once you see these six as dials, you can design legal workflows that adapt under stress instead of relying on static rules.

Applying this to legal work

Due diligence and disclosure reviews

Start with low authority. The AI suggests clusters, flags possible issues, and generates draft observations, but none are final. As lawyers provide rationale for corrections, accuracy improves, trust rises, and shared awareness builds. At that point the AI can take on low-risk tagging. If a late tranche of documents arrives or the quality of inputs drops, the system should clamp its autonomy, widen sampling, and demand stronger explanations until the uncertainty clears.

Playbooked contract reviews

The assistant proposes redlines with confidence scores and cites the playbook rule used. Lawyers respond with reasons for acceptance or rejection, for example "accepted due to clause 12.3 context." Those rationales teach the system what really matters in practice. Routine clauses may eventually be handled with high autonomy, while complex indemnities always require human oversight. Autonomy scales with risk rather than convenience.

Litigation triage and case strategy

Shared awareness is built first: issues, facts, procedure, and deadlines. The AI produces several strategic options with confidence levels and trade-offs. The lawyer chooses one, records the reasoning, and the system improves over time. If the case shifts due to new evidence or rulings, autonomy reduces until confidence and alignment recover.

Building co-learning into legal platforms

• Instrument the six dials. Track accuracy to estimate AI competence, log rationales to calibrate trust, maintain coverage maps to measure shared awareness, and monitor workload to protect human capacity.
• Set autonomy policies. For example: if accuracy is above 92% and explanations are consistently rated useful, allow auto-tagging of low-risk clauses. If uncertainty spikes, clamp authority and raise sampling.
• Require explanations both ways. Lawyers should record why they accept or reject an output, linked to clauses or precedents. That feedback is how the system learns.
• Propose options, not single answers. Multiple candidates with trade-offs force human choice and capture valuable rationale.
• Scale oversight with authority. As autonomy rises, increase audit, sampling, and explanation depth.

Metrics that matter

• Time taken for autonomy to stabilise on a matter
• The proportion of explanations rated as useful by reviewers
• How quickly trust recovers after an error or data shock
• Error types caused by over-reliance or under-reliance
• Frequency of cognitive load breaches and how the system responded

These are better measures than simple “accuracy scores" because they capture the behaviour of the team under pressure.

Does this slow things down?

At first glance, a framework built around clamps, feedback loops, and shifting autonomy might look like it slows progress. Law already has a reputation for caution, so why build in even more brakes?

The answer is that legal work has never been about speed at any cost. It is about defensibility. A diligence process that misses a liability, a contract that accepts a redline without context, or a regulatory response that exposes sensitive data because these aren’t small errors, they are liabilities that damage trust and carry real cost. In that world, slowing down at the right moments is a feature, not a flaw.

Agentic AI still fits. When conditions are stable and performance is high, autonomy can rise and the system can run at speed. That delivers the efficiency law firms and clients want. The difference is that when uncertainty creeps in, the agent adapts rather than charging ahead blindly.

The point of this model isn’t to hobble agentic AI. It is to make it resilient enough for the realities of practice. Law moves quickly in bursts, but always under the eye of clients, regulators, and courts. Defensibility under scrutiny matters more than raw throughput, and this framework makes sure AI respects that balance.

Treat human-AI legal work as a dynamic system. Instrument the dials, wire in the clamps, and build workflows that flex when uncertainty hits.

That is how legal AI earns real trust, and why this military framework may be exactly what law needs to balance pace with defensibility.