Generative AI has transformed contract drafting. Something that once took a few hours can now take seconds. Documents arrive for review at a volume and pace that the unaided reviewer cannot match, assembled from firm precedent, AI-generated amendments, counterparty redlines of uncertain origin, and human edits.  A document reads cleanly, the definitions are tidy, the cross-references, on the surface, work.

While contract drafting has transformed, contract review is conducted largely as it was a decade ago: a lawyer, a screen, and the time to read. Undertaking contract review has always generally been framed by some sort of tension whether it be negotiation with a counterparty, a second set of eyes review, senior lawyers supervising junior lawyers, and the signposts of human intervention. With the advent of legal technologies, it has become increasingly difficult to identify and highlight any such tension. A lawyer now must take personal accountability for every clause in a document, despite having spent no time writing any of them.

The Issue

Under the new reality of unclear provenance of documents, particularly in a world where generative AI continues to creep into contract preparation, unaided review risks becoming an exercise which is no longer fit for purpose. Review runs the risk of being no longer able to carry the accountability owed to clients, regulators, the courts, and the profession itself.

We call the resulting gap comprehension debt. The distance between an existing document and how much the lawyer dealing with it understands. It shows up as an indemnity that does not map to the risk allocation everyone agreed on a call two weeks ago, or a defined term that gradually drifts in meaning across 200 pages.

The question it raises is whether, in the age of generative AI in document production, regulatory bodies in the legal profession should clarify how best to carry out obligations to meet the added pressures review of legal documents now requires.

‍
The Argument for Augmented Review

Professional obligations already in force - competence, supervision, the duty not to mislead, the ability to justify a judgment - point toward a specific practical requirement: technology built to challenge documents should be used alongside technology that generates them, particularly when a document's provenance is unclear.

Existing standards of competence and supervision, should be properly applied to how documents are produced today. There must be a layer of independent, augmented review that most firms have not yet built in.

The Four Properties of Credible Assisted Review

Not all "AI review" is equal, and the differences matter enormously for accountability. A review tool worth relying on has to satisfy four properties:

1. Structural separation of generation and review, demonstrable on inspection. 

The technology checking a contract must be genuinely independent of the technology that helped draft it - independent in how it reasons - and a lawyer must be able to show that separation to a client, regulator, or risk committee. This is important because a generative system reviewing output produced by the same reasoning architecture tends to reproduce the very failure modes it's meant to catch. Its known tendency toward sycophancy inclines it to ratify text rather than test it.

^{Relevant SRA Principles: Principle 2: Uphold public trust in the profession, Principle 7: Act in the client's best interests, Para 3.5(b): Supervise others' work effectively.}

2. Per-output traceability. 

For each material finding, there should be a record of what was checked, against what reference, and where human judgment entered - detailed enough for the lawyer to reconstruct and stand behind the reasoning, not merely to point back at a prompt.

^{Relevant SRA Principles: Para 3.2: Provide a proper standard of service to clients, Para 1.4: Don't mislead clients, the court, or others, including by omission or through your client's conduct, Para 7.2: Be able to show, after the fact, why your decisions and actions complied with your obligations.}

3. Reproducibility of findings. 

Running the same check on the same document should produce the same result, so that a flagged issue is a fact about the document, not an artefact of a particular run. General-purpose LLMs are probabilistic and won't reliably return identical results; deterministic approaches offer real certainty here in a way generative ones structurally cannot.

^{Relevant SRA Principles: Para 3.2: Provide a proper standard of service to clients, Para 7.2: Be able to show, after the fact, why your decisions and actions complied with your obligations}

4. Structured and visible human oversight points. 

A credible tool identifies, rather than blends away, the points where the system has reached its limits and human judgment must take over, in a form the supervising lawyer can act on.

^{Relevant SRA Principles: Principle 7: Act in the client's best interests, Para 3.5(b): Supervise others' work effectively.}

Why the urgency?

Courts and regulators are already responding to failures of exactly this kind, and the direction of travel is unmistakable.

In R (Ayinde) v London Borough of Haringey [2025] EWHC 1383 (Admin), the Divisional Court articulated the duty of independent verification of AI-assisted material and expressly invited regulators to consider further measures. In Bandla v SRA [2025] EWHC 1167 (Admin), a solicitor who filed 25 non-existent AI-generated authorities was ordered to pay over £24,000 in costs. Courts have begun locating responsibility at the level of systems rather than individuals: in Ndaryiyumvire v Birmingham City University (November 2025), the court made a wasted costs order against the firm, not the named solicitor, treating the filing of fabricated authorities as, in substance, a failure of management and verification controls.

By July 2026, more than 1,700 judicial decisions identifying AI-generated errors in filings had been recorded internationally. Errors of this kind have passed both partner-level and secondary review at leading US firms, a sign that the problem isn't confined to solo practitioners or under-resourced teams. 

Regulatory and professional bodies across jurisdictions are converging on the same themes, even where they've stopped short of new rules. The Civil Justice Council's February 2026 consultation on AI in court documents proceeded from the premise that no new rules are needed for advocacy documents, provided a legal representative takes professional responsibility for them. Recent SRA compliance and supervision guidance — the latter expressly extending supervision expectations to AI use — restates existing standards rather than adding to them. The EU AI Act's transparency obligations take effect on 2 August 2026, with the fuller high-risk regime, including its Article 14 human oversight requirements, deferred to December 2027 under the Digital Omnibus, leaving professional judgment and firm-level discipline, not statute, as the binding constraint in the interim. ABA Formal Opinion 512, Bar Council and Law Society guidance, and the CCBE's October 2025 guide all converge on the same four themes: auditability, traceability, human oversight, and demonstrable independence. 

Regulators are also beginning to treat architecture itself as a substantive question, not an implementation detail. Recent UK authorisations of hybrid and deterministic legaltech providers - with generative AI confined to a bounded role - suggest that how a system reasons, not just what it outputs, is becoming part of what "adequate" review means.

What the industry needs to do

Taken together, this points to a few concrete asks for the industry - firms, in-house teams, technology providers, and the bodies that set standards for all three:

• Treat review architecture as a professional judgment. Firms should be able to answer, concretely, whether the tool checking a document is independent of the tool (or model) that helped draft it.
• Build traceability into the review process itself, so that a finding can be reconstructed and defended after the fact.
• Favor reproducibility where it's available. Where deterministic checks can do the job, they offer a form of certainty that probabilistic, general-purpose models cannot currently match.
• Make human oversight points visible. The places where judgment must take over should be flagged as such, not smoothed into a single confident-sounding output.
• Treat existing competence and supervision obligations as already requiring this, rather than waiting for bespoke AI rules before building the discipline.

None of this is an argument against generative AI in legal work. It's an argument that drafting and review need to remain separable - architecturally and procedurally, if the profession's existing accountability to clients, courts, and regulators is going to survive contact with how documents are produced now.

Definely protects the sovereignty, sanity, and certainty of the lawyer. The world’s leading lawyers and in-house teams trust Definely to read, analyse and proof complex contracts. Get in touch to find out more.

‍