Chapter 10AI Compliance course

Chapter 10. World II: China and the international layer, a divergence map

Story phase: horizon. Wrapping up the survey of the world — a third philosophy (China) and the supranational layers. A broadening of the view, not an operational guide by country.

The situation at Kompas

Kompas looks further — Asia, Latin America, global corporate clients. At this scale the question changes: no longer "which law in country X," but "by what logic are the regimes built at all, and where is the industry heading." The team needs not an encyclopedia of fifty jurisdictions (it will age out by the end of the quarter), but a map of the axes along which regimes diverge — so it can make expansion decisions rather than drown in national checklists.

This chapter closes the survey of the world with a third philosophy of regulation (China) and the supranational layers that stitch the scattered regimes together.

What the rule says

China — a third philosophy: content control and state security, not protection of the individual's rights (as in the EU) and not market freedom with ex-post lawsuits (as in the US). Regulation is targeted and fast, by types of application:

Rules on algorithmic recommendations (regulation of recommender systems).
Rules on deep synthesis (deepfakes and synthetic content — registration, labeling).
Measures on generative AI (a security assessment before public launch, labeling of AI content, a requirement of conformity with "core socialist values").

For Kompas this means: even innocuous HR scoring in China lives in a regime where control over data and content and state security come first, not the individual rights of the candidate. The requirements may be not "stricter/softer" than the European ones, but orthogonal to them.

The international layer — what makes the mosaic at least a little compatible:

OECD AI Principles — a common baseline that many regimes reference (including the very definition of "AI system" in the AI Act, aligned with the OECD).
Council of Europe AI Convention — the first binding international treaty on AI, tied to human rights, democracy and the rule of law; a frame for signatory states.
ISO/IEC 42001 and NIST AI RMF — bridge standards: a single certifiable management framework recognized across jurisdictions lowers the cost of a multi-regime entry (see ch. 8).

Other jurisdictions (statuses move, verify as of the date): Canada (AIDA within Bill C-27 — stalled at the dissolution of parliament), Brazil (PL 2338 — risk-based in the EU spirit), South Korea (a framework AI Basic Act, entry ~2026), Japan (a soft, guidance-based framework approach).

The axes of divergence — the thing the chapter is for: ex-ante vs ex-post; risk regime vs sectoral; individual rights vs state content control; law vs principles vs standards. Any national regime decomposes along these axes.

How it lands on the product

The practical conclusion for Kompas: keep the EU AI Act as the floor + deltas by market, rather than launching N parallel compliance projects. The European baseline closes most of the requirements of rights-oriented regimes; on top of it you maintain a list of differences (the US hiring specifics from ch. 9, China's content control, local data requirements).

ISO/IEC 42001 is the common denominator that lowers the cost of entering each new jurisdiction: one governance framework, many regimes.

The artifact is a divergence map along the axes, which will become the input to the cross-jurisdictional matrix of ch. 11.

Where it breaks

Statuses go stale fast. A bill is not a law; Canada showed that an act can die at the dissolution of parliament, Colorado — that it can be cut back on the go. The course is not a source of dates; before a real entry the status is verified as of the current day (the fact discipline from _index).

Orthogonality, not a scale. China and similar regimes can't be laid out on a "stricter/softer" axis: their requirements for content and data may be simply incompatible with Western ones. "Passed the EU" here neither helps nor hinders — it's a different axis.

A "map" ≠ a country guide. The temptation to pass off a survey map as an operational instruction for a specific jurisdiction. No: the map is for the expansion decision; the details of a specific entry — with a local lawyer.

What to do as engineer/product

Keep one baseline (EU + ISO 42001) and maintain a list of deltas for target markets instead of parallel regimes — that way the cost of each new jurisdiction falls rather than multiplies.

Before a real entry into a jurisdiction — verify the current status of the acts, don't rely on a snapshot from the course or a consultant's report from last year.

Provocation

A global AI product has no single "compliance" — it has a matrix of regimes that diverge not by strictness but by philosophy. Trying to make one document "for the whole world" is like writing one passport for all borders: the EU floor plus explicit deltas by market works, while a universal paper works nowhere, because it pretends that Brussels, Washington and Beijing regulate the same thing.

Shipping an AI product under regulatory risk?

A read of your product against the EU AI Act: risk class, role in the value chain, obligations and dossier — as a design constraint on the way in, not a lawyer's check at the end.

Email me

The transition engine

Next Move Engine — the system that takes a team to an autonomous delivery loop.

Next Move Engine →