Anthropic just published "When AI builds itself." I read it twice—first as a technologist, then as a cross-border lawyer. Both reads were uncomfortable.

The facts on the table:

→ As of May 2026, over 80% of code in Anthropic's own codebase is written by Claude.
→ The duration of tasks AI can reliably complete is doubling roughly every 4 months. March 2024: ~4 minutes. One year later: ~90 minutes. Today (Opus 4.6): ~12 hours.
→ April 2026: Claude shipped a single commit fixing 800+ issues, reducing a class of API errors to 1/1000 of prior rates. Engineers estimated the human-equivalent workload at ~4 years.

A 4-minute AI is a tool. A 12-hour AI is a colleague.

But the real protagonist of the piece isn't the numbers. It's a concept: Recursive Self-Improvement (RSI).

A few months ago, the frontier conversation was about multi-agent systems—orchestrating AI engineers to work together. I run one myself for my own legal content and compliance pipelines. The frontier has already moved: AI designing, training, and iterating its own successors. For the first time in history, technology may push its own accelerator.

Anthropic is candid: subtle alignment flaws in today's models could compound across AI-built generations. They call alignment under RSI "the thing we're least certain we can solve."

Now the lawyer's question almost no one is asking: what kind of legal document is this, really?

My read: it serves two simultaneous functions.

Risk disclosure. Putting internal velocity data into the public record is, functionally, a formal warning to society.
Liability positioning. Under tort and compliance frameworks, an actor who—facing material uncertainty—publicly and specifically warns and invites third-party governance is shaping how future courts will assess foreseeability and duty of care. This isn't speculation about motive. It's what any prudent institution does when staring at a cliff.

For Chinese companies going global, AI vendors, and anyone moving cross-border data: governance and verification will become commercial friction points before the technology does. GDPR Art. 22, the EU AI Act, China's PIPL, Brazil's LGPD Art. 6—these are floors, not ceilings.

Build your own disclosure-logging-human-review chain now. After the incident, you can't manufacture evidence that you exercised due care.

Others shout "AI will end the world." Lawyers should stay sober: the danger isn't that AI is too fast. It's that rules, evidence, and human judgment are too slow.

How is your team documenting AI decisions today? Genuinely curious.

#AIGovernance #CrossBorderLaw #AICompliance #AIAct #PIPL

———
Mike Wu, China-licensed Attorney · Cross-border practice
4 languages (CN/EN/ES/JA)
Site: wzklaw.com/contact.php?source=linkedin · Book consultation: wzklaw.com/contact.php?source=linkedin