The New Statesman's April investigation confirmed what many suspected: large language models have already written passages in UK acts of parliament, and AI systems analysed departmental bids in the June 2025 Spending Review that allocated £2bn to AI while cutting 16% elsewhere. For mid-market legal, insurance, financial services and accountancy firms, this is not an abstract story about government modernisation. It is a structural problem. When the rules that govern your conduct are being drafted by systems with no traceable reasoning, no accountability for errors, and no way to challenge their logic, your compliance obligations become opaque. The FCA's Consumer Duty (PS22/9) requires firms to understand the tools they use. The SRA Code demands transparency. Yet the legislative framework itself—the source of your regulatory obligations—is now partly generated by the same black-box systems regulators are asking you to audit and explain.

This story reveals a dangerous asymmetry at the heart of UK regulation. Firms are being held to standards of AI governance that the government itself does not meet. The EU AI Act imposes strict conformity assessment and documentation requirements on high-risk AI systems. The UK's approach has been lighter-touch, but that lightness is now weaponised: it allows government to deploy LLMs in policy-making without the kind of impact assessment, audit trail, or human review that regulated firms must apply to their own tools. What we are seeing is not innovation; it is regulatory arbitrage. Government gets to move fast and break things. Regulated firms cannot. The firms that adapted first to this reality—those using structured RAG (retrieval-augmented generation) tools rather than off-the-shelf LLMs, those building explicit human checkpoints into document drafting workflows, those treating AI outputs as drafts requiring senior review rather than publishable work—are positioning themselves correctly. Those still treating Copilot or comparable general-purpose tools as productivity multipliers without institutional control are not.

Trovix's position is that government use of LLMs in legislative drafting should trigger immediate questions from regulated firms about how they interpret new law, train staff on unclear requirements, and manage the legal risk created by ambiguous rules. The honest assessment: no AI system can reliably disambiguate legislative intent when that intent was never crystallised by human reasoning in the first place. This is why we built Trovix Watch to monitor regulatory change with human analysis at the core, not AI pattern-matching alone. Competitors like Harvey or Luminance excel at contract analysis and precedent extraction—genuinely valuable for the 80% of work where the rules are stable and clear. They are not designed for the problem of emerging legal ambiguity created when the source documents themselves are partly AI-generated with no human redaction trail. That problem requires a different architecture: one that flags changes, surfaces contradictions with prior interpretation, and makes explicit when new rules lack clear human authorship. Tools like Trovix Aria are built for that because they assume adversarial conditions—meaning they work backwards from regulatory obligation to evidence, not forwards from training data to guess.

What you should do now: First, audit how your firm currently adapts to new regulation. Are you relying on AI summaries of legal change, or is a qualified person reading source documents with scepticism built in? Second, establish a policy on AI-generated legislative material. When you encounter a new rule, statute or guidance, ask: who wrote this? Can you identify the human author's reasoning? If you cannot, treat it as provisional and build in additional compliance buffer. Third, resist the false confidence that general-purpose LLMs create. Microsoft Copilot is efficient; it is not reliable for regulatory interpretation under uncertainty. Fourth, ensure that your matter intake, document handling and compliance workflows—whether powered by Trovix Brief or another tool—have explicit human sign-off, particularly when AI is involved in understanding regulatory requirements. The government has shown that AI in policy-making creates ambiguity. Do not compound that by automating your own compliance response.

Source: New Statesman