As artificial intelligence becomes embedded in everyday legal drafting, a quiet risk is emerging: what I call machine-introduced legal drift. It occurs when an AI tool subtly alters a clause such as by changing a verb, a qualifier, or a limitation, and neither lawyer nor client notices. The clause still looks right. It reads well. But its meaning has shifted.
That shift in meaning may not become obvious until months or years after the contract has been executed, when the parties look to the contract to determine their responsibilities or to resolve a dispute. The temptation to ask an AI system for a quick solution is growing day by day, and many firms and businesses are adopting ‘enterprise wide’ systems built around AI, and with the same limitations.
What Legal Drift Looks Like
Legal drift isn’t about obvious mistakes. It’s about plausible edits that carry real consequences. Those edits are typically subtle and embedded within otherwise well-written clauses. Examples might include:
- Adding “subject to the supplier’s discretion” where no discretion was intended.
- Inserting references to non-existent standards, “as required by law”, that don’t actually exist.
- Carrying over regulatory terms across jurisdictions without considering differences and their implications.
Each change seems minor. Together, they can shift risk, dilute obligations, or even break compliance with applicable law.
Why It Happens
AI models optimise for plausibility, not legal fidelity. They generate text that looks right. When a prompt is vague “tighten this clause” or “simplify this agreement” the model may restructure entire risk allocations in the process. Even trained systems that draw on firm precedents can misfire when they “borrow” similar but contextually incompatible clauses.
The Red Flags
From an outcome perspective, you need to read the product, carefully and critically and with distrust. Watch for drift in clauses that determine:
- Liability and indemnities
- Termination rights
- Warranties and representations
- Regulatory obligations
- Governing law and jurisdiction
- Privacy and data protection
These are the levers of risk, and the ones most prone to AI-driven variation.
Building a No-Drift Workflow
To manage this, lawyers need layered controls that go beyond simple proofreading:
- Guardrail the generator. Use locked clause libraries with metadata (jurisdiction, version, risk tier). Block free-form generation for critical provisions.
- Enforce schema-based outputs. Every AI draft should identify the clause source, the jurisdiction, and how it differs from the baseline.
- Run automated diff and risk gating. Compare output to the gold-standard template and flag verb, number, or modal changes (shall vs should).
- Use dual-model verification. Have a second AI or rule-based tool identify contradictions or weakened obligations.
- Apply policy-aware validation. Check the draft against current regulatory instruments.
- Target human review. Reviewers shouldn’t simply read the contract; they should see the changed clauses with context and rationale.
- Maintain provenance and audit. Record which model generated each clause, from which precedent, and when. Store this in a transparent audit file.
- Freeze critical sections. For client-facing documents, lock high-risk clauses; allow AI to suggest, not insert. Provide a plain-English appendix showing what actually changed.
From a process perspective, the key is to step back or to involve fresh eyes. Whatever is at risk with a deadline will almost certainly pale in comparison to the consequences of agreeing to a contract that contains long-term traps and trip-wires.
A Simple Reviewer Checklist
Before finalising any AI-assisted draft, confirm that:
- Liability caps, indemnities, and warranties are unchanged.
- Governing law, jurisdiction, and regulatory references are intact.
- Definitions and cross-references still align.
- No hallucinated standards have appeared.
The Way Forward
AI can accelerate and enhance legal drafting, but only when it operates within strict boundaries. Legal drift is insidious precisely because it is subtle: it hides in the margins of the plausible. The solution is not to abandon AI, but to engineer systems that expose and control its changes, allowing lawyers to focus on judgement, not syntax.