The market reaction to Claude Legal Co‑Worker says more about hype and expectations than about what legal and regulatory AI can reliably do today. This post looks at three things: why markets panicked, why hallucinations remain a critical unsolved problem, and why operational reality is being glossed over. It concludes with why discipline, not demos, will decide who actually wins.

The Market Blinked

Anthropic’s Claude Legal Co‑Worker announcement triggered an immediate selloff across the sector. Thomson Reuters and LegalZoom were both down roughly 20% yesterday alone. That tells you how forward‑looking markets are. It does not tell you how ready the technology is. The demo video has been public for more than two months, yet only now did the narrative fully land. Investors moved fast. Practitioners should move carefully.

Hallucinations Are the Real Story

Even the Claude video quietly concedes hallucinations.

In legal and regulatory workflows, that concession is not minor. Across the industry, we consistently see hallucination rates exceeding 50% when large language models are applied without hard constraints. A system that sounds confident but cannot reliably trace outputs back to specific rules, regulations, or sources is not a co‑worker. It is a risk.

This is not a model limitation. It is a design choice.

The No‑Code Illusion

Many vendors now position no‑code legal agents as the answer. In practice, these systems often sacrifice consistency, determinism, and auditability. I have personally tried to break them and in many cases, it is not difficult. When answers change materially based on phrasing or context, you do not have reliability. You have probabilistic guesswork dressed up as productivity.

The Platform Reality

There is also a practical detail being overlooked. Most lawyers use Windows. Claude Legal Co‑Worker is Mac‑only today, with a Microsoft version not expected until mid‑2026. In regulated enterprises, that alone can delay adoption by years.

The Marketing Irony

The demo uses a marketing review as an example, which is oddly fitting. Some of the claims around legal AI today feel like marketing themselves. At times, the promises outpace what can be safely delivered in regulated environments.

A More Grounded Approach

At Surveill, we take a different view. Legal and regulatory AI does not succeed by ignoring uncertainty. It succeeds by engineering around it.

Rule traceability, consistent outputs, and explainability matter more than eloquence. AI should support professionals, not replace judgment with confident‑sounding guesses.

The future of legal AI is real. But it will be built on discipline, not hype.