Once the basics are automatic, the failures change character. The beginner's enemy is the hallucinated citation and the unauthorized commitment, both catchable with a review list. The expert's enemy is subtler: the draft that is locally correct everywhere and globally wrong, the conflict that only appears across two jurisdictions, the defined term that drifts meaning across forty pages without ever being misused in a single sentence. These do not show up in a quick read, and they are exactly where expensive mistakes live.
This piece is for practitioners who already ground their inputs, stage their prompts, and verify their citations. It covers the edge cases that survive a competent basic process, the places where grounding stops helping, and the techniques that experienced drafters use when the easy defenses run out. None of this replaces counsel; it makes the work you bring to counsel sharper and the failures you hand them rarer.
Multi-Jurisdiction Conflicts
A document that must satisfy two regimes at once is where models quietly fail, because they tend to satisfy whichever regime they thought of last.
The Failure Pattern
- The model applies the stricter rule in one clause and the laxer rule in another, producing internal inconsistency that reads fine clause by clause.
- It defaults to the most common jurisdiction in its training data, not the one you specified, when the prompt is ambiguous.
Techniques That Help
- Draft to the strictest applicable standard explicitly, then relax deliberately where allowed, rather than asking the model to reconcile regimes itself.
- Prompt for a jurisdiction-by-clause table separately, so conflicts become visible as a structure rather than hiding in prose.
Defined-Term Drift
In long documents, a term can be used correctly in every sentence and still drift in aggregate meaning. This is the failure that survives the consistency checks in A Working Review List for AI-Drafted Legal and Compliance Text.
Why It Happens
- The model's attention to a definition fades over a long generation, so later usage subtly broadens or narrows the term.
- Defined terms that overlap in plain meaning ("Service" versus "Services") invite quiet conflation.
Techniques That Help
- Generate the definitions section last, after the body, so the definitions are pinned to actual usage rather than guessed up front.
- Run a dedicated pass that extracts every use of each defined term and checks the meaning holds, separate from the general read.
The Limits of Grounding
Grounding is the strongest basic defense, and experts have to know where it stops working.
Where Grounding Fails
- When your reference material is itself outdated, the model faithfully reproduces a stale obligation, and grounding makes the error more confident, not less.
- When the document requires synthesis across sources that conflict, grounding gives the model material but not the judgment to reconcile it.
Techniques That Help
- Date and version your grounding material, and treat reference freshness as a first-class input, not a static asset.
- For genuine synthesis, use the model to surface the conflict and a human to resolve it, rather than asking the model to paper over it.
Adversarial Review of Your Own Drafts
Experienced drafters stop trusting their own prompts and start attacking the output. This is where the trade-offs in Speed Versus Defensibility When AI Drafts Compliance Language get operationalized at the expert level.
Pressure-Testing Techniques
- Prompt a fresh session to argue against the draft as opposing counsel would, then triage what it finds.
- Ask specifically "what commitment does this create that the business may not have intended," which surfaces buried exposure.
- Re-run a high-stakes section with a different model and diff the outputs; divergence marks the spots that need human judgment.
Managing Long-Document Coherence
Length is its own adversary. A model's attention to early constraints fades over a long generation, so documents past a certain size develop coherence problems that have nothing to do with any single clause being wrong.
Where Long Documents Break
- Constraints set at the top of a long prompt weaken by the bottom of a long output, so later sections drift from the rules earlier sections obeyed.
- Cross-references multiply, and the model loses track of which section actually contains the referenced content.
- Tone and register shift subtly across sections, leaving the document feeling stitched rather than authored.
Techniques That Help
- Generate long documents section by section with the constraints re-stated for each, rather than in one sweep that dilutes them.
- Build the cross-reference map yourself and have the model fill against it, instead of trusting it to invent and track references.
- Reserve a dedicated coherence pass that reads only for global consistency, separate from any clause-level review.
The discipline here is the long-document version of the staged thinking in The DRAFT Method: Structuring Prompts for Regulated Writing, applied so that length does not erode the constraints the method establishes.
Handling Regulatory Change Mid-Stream
Experts also have to manage the case where the underlying regulation shifts while a document or template is in active use. This is where stale grounding becomes a live operational risk rather than a theoretical one.
The Failure Pattern
- A template grounded against last year's rule keeps producing confident, outdated drafts long after the rule changed.
- The model gives no signal that its reference material is stale; it reproduces the old obligation as authoritatively as a current one.
- Volume amplifies the harm, because a single stale template can defect every draft it touches.
Techniques That Help
- Date and version grounding material and tie a review cadence to the regulatory calendar, not to convenience.
- Treat a regulatory change as a trigger to re-validate every template that touches the affected regime.
- Watch your citation verification failure rate by document type, as described in Signals That Tell You AI Compliance Drafts Are Holding Up, since a stale template often shows up there first.
Knowing When to Stop Prompting
The mark of an expert is recognizing the document the model should not be drafting at all, the territory mapped in What AI-Assisted Compliance Drafting Saves, and What It Costs.
Signals to Hand Off
- The document is novel enough that no grounding material exists, so the model is improvising on regulated ground.
- The exposure is high and the reversibility is low; the cost of being wrong dwarfs any drafting savings.
- The synthesis required is genuinely contested among experts; a model cannot adjudicate what counsel debates.
Frequently Asked Questions
Why do multi-jurisdiction documents fail even with good grounding?
Because grounding supplies material but not the reconciliation logic. The model tends to satisfy one regime per clause and does not hold both in view across the document, so it produces locally correct clauses that conflict globally. Draft to the strictest standard explicitly instead.
How is defined-term drift different from a consistency error?
A consistency error misuses a term in a specific sentence and is catchable by a normal check. Drift keeps every sentence locally correct while the term's aggregate meaning shifts across a long document. It survives sentence-level review and needs a dedicated term-by-term pass.
Can grounding ever make a draft worse?
Yes, when the reference material is stale. The model faithfully reproduces an outdated obligation with full confidence, and the grounding makes the error harder to spot. Treat reference freshness as an input you verify, not a static asset you trust.
Is using a second model to check the first actually useful?
For high-stakes sections, yes. Divergence between two models on the same prompt reliably marks the places where the answer is uncertain and human judgment is required. It is a cheap way to find where to spend expensive attention.
When should an expert refuse to use AI for a document?
When no grounding material exists, exposure is high and irreversible, or the required synthesis is genuinely contested among experts. In those cases the model improvises on regulated ground and the drafting savings are dwarfed by the risk.
Does adversarial self-review replace counsel?
No. It sharpens the draft and reduces the easy failures you hand to counsel, which makes their review faster and more focused. The judgment about what the business intended and what the regulation requires still belongs to a qualified person.
Key Takeaways
- Expert failures are global, not local: drafts that read correctly clause by clause but conflict in aggregate.
- Multi-jurisdiction documents need an explicit strictest-standard approach, because models do not reconcile regimes on their own.
- Defined-term drift survives sentence-level review and requires a dedicated term-by-term pass and definitions written last.
- Grounding fails when reference material is stale or when genuine synthesis across conflicting sources is required.
- The expert skill is knowing which documents the model should not draft at all, and handing those off early.