The UK government integrates its Consult AI agent for drafting legislation, pioneering the era of algorithmic governance.
What "Consult AI" Actually Does
The UK government's Consult AI agent applies large language models to one of the slowest, most labour-intensive parts of governing: turning policy intent into the precise wording of a bill. Legislative drafting is not ordinary writing. Every clause has to fit inside an existing body of statute, avoid contradicting other laws, and survive the adversarial reading that courts and opposing counsel will later give it. An agent aimed at this task is less a general chatbot than a specialised tool that proposes clause language, cross-references related provisions, and flags where new text might collide with what is already on the books.
The word "consult" points to the intended posture. Rather than replacing the human drafter, the system is positioned as something a lawyer or official queries during the process — a way to accelerate first drafts, surface precedent, and check consistency, while a person retains the pen.
Why Legislation Is a Hard Target for Automation
Statutory language is unforgiving in ways ordinary prose is not. A misplaced modifier can widen or narrow a power far beyond what was intended, and ambiguity that a reader would gloss over becomes the exact seam a dispute is fought along. That makes drafting a good fit for machine assistance in some respects — it is pattern-heavy, repetitive, and built on reusable structures — and a dangerous one in others, because the cost of a plausible-sounding error is high and may not surface until the law is tested.
The practical challenges an agent in this domain has to handle include:
- Keeping generated clauses internally consistent and consistent with the wider statute book.
- Distinguishing settled legal meaning from everyday word usage, since terms of art carry fixed definitions.
- Producing text a human expert can audit line by line, rather than an opaque output that must be trusted wholesale.
- Handling amendments, which edit existing law by reference and demand exact pointers to the passages they change.
The Shift Toward Algorithmic Governance
Using an agent inside the drafting process moves automation from the periphery of government — form-filling, casework triage, public-facing chat — into the core act of making rules. That raises questions that are political as much as technical. Who is accountable when drafted text carries an error: the official, the agency, or the tool's builder? How is the model's influence recorded so that scrutiny bodies and the public can see where machine suggestions shaped a law? And how do you prevent the convenience of accepting a generated clause from quietly narrowing the range of options a drafter considers?
These are not reasons to avoid the technology, but they are reasons to build guardrails around it. Provenance tracking, mandatory human sign-off, and clear disclosure of where an agent was used are the kinds of controls that keep automation legible rather than hidden.
What to Watch as This Matures
The useful test is not whether an AI can produce clause-shaped text — models do that easily — but whether it reduces genuine drafting errors while keeping a human firmly in control. Signals worth watching are how the system is evaluated against expert-written baselines, whether its suggestions are logged and reviewable, and whether it is confined to assistance rather than allowed to originate law unsupervised.
If drafting agents deliver, the gain is throughput and consistency on a task that has always bottlenecked on scarce specialist time. The failure mode to guard against is subtler: automation that looks authoritative, gets waved through, and bakes mistakes into text that then binds everyone.