China, the UK, and the divergence problem: one product, three rulebooks
The short answer
One product is never one legal thing
The same code, shipped once, lands in as many rulebooks as it has markets, and those rulebooks were written independently for different purposes. Governing a global feature means holding several rulebooks in view at once, not picking the one you know best.
What you will be able to do
- Describe the core AI-content obligations of three divergent regimes: China's mandatory labeling of AI-generated content (the Cyberspace Administration of China measures plus standard GB 45438-2025), the European Union's transparency obligations (EU AI Act Article 50), and the United Kingdom's principles-based, no-statute approach.
- Distinguish the three underlying philosophies (state-directed control, rights-and-risk regulation, and pro-innovation principles) and explain why the same technology produces such different rules in each.
- Analyze one of your own product features across all three rulebooks, classifying each obligation as an agreement, a divergence, or a true conflict between the regimes.
- Locate a true conflict, a case where satisfying one regime's rule would breach another's, and explain why "build for the strictest" fails there.
- Choose among the four responses to divergence (comply to the highest common standard, geo-fence the feature by region, feature-flag behavior per jurisdiction, or withdraw from a market) and justify the choice for a specific obligation.
- Connect this analysis to the EU AI Act work you already did in Module 5 and to the cross-border shipping decision you will make in Topic 6.6, so the divergence map is a working input and not an essay.
- Read each obligation three ways, by its goal, its required form, and the payload that form carries, so a single row's agreement, divergence, and hidden conflict all become visible.
- Sequence the analysis as a repeatable procedure rather than an ad hoc scan, so the step that would have caught a collision is never skipped under deadline pressure.
- Defend a divergence decision against a challenge that you either over-complied (and lost a market you could have kept) or under-complied (and shipped an illegal feature somewhere).
The ideas, one by one
Obligations do not simply add up
Every obligation-pair is an agreement (one action serves both), a divergence (more work, no contradiction), or a true conflict (one regime's required act breaches another's rule). Sorting each obligation into one of the three is the analytic core; leaving an obligation unsorted is how a conflict hides.
"Build for the strictest" fails at true conflicts
It is a fine default for agreements and divergences and actively dangerous at a conflict, because the strictest rule in one country can be the forbidden act in another. There is no single strictest country; strictness is per-obligation, so the highest common standard is assembled clause by clause.
China's rulebook is different in kind, not just degree
The CAC Labeling Measures and GB 45438-2025 (in force September 2025) are prescriptive, name the responsible provider by design, and express state-directed information control. Reading them as "the EU but stricter" hides the places where their mandated form carries something a rights regime constrains.
The EU protects a right and arrives on a date
Article 50 transparency (machine-readable marking, deepfake disclosure) applies from 2 August 2026, an outcome-based duty softened by a 2026 Code of Practice. It is enforceable and dated, and its style (legislate the outcome, code the method as guidance) is the mirror image of China's (code the method as law).
An absence is not a safe harbor
The United Kingdom has no AI-specific labeling statute, but consumer-protection, data-protection, defamation, and sector law still reach AI content, and the DSIT five principles signal direction. The correct UK reading is "no AI statute, check the underlying law," never "nothing required."
Divergence runs across time, not only space
China's regime is live; the EU's switches on in 2026; the United Kingdom has no dated duty. A feature can be lawful in the EU today and unlawful on a known future date with no code change, so every cell needs a "from when."
The four responses must match the relationship
Comply to the highest common standard, geo-fence, feature-flag per jurisdiction, or withdraw and redesign. Agreements and simple divergences take response 1 or 3; a true conflict can take 2, 3, or 4 but never 1, because comply-to-highest assumes the standards can coexist, which a conflict denies.
Transparency is not yet interoperable
China's implicit metadata and the EU's machine-readable marking both aim at detectable synthetic media but are not guaranteed to read each other. Convergence on machine-readable provenance is emerging, not achieved; never promise a client one global label.
The map is an instrument, not an essay
It exists to feed a real decision, which markets you ship into and with what behavior, made concrete in the cross-border decision of Topic 6.6. Built before the ship decision, it prevents an illegal launch; built after, it is just documentation of one you already made.
The signature failures are over- and under-compliance
Over-complying pushes one regime's mandated act into a market that constrains it or needlessly shrinks a market; under-complying reads an absence or a future date as "fine." A defensible map is attacked for both before a regulator does it for you.
Read every row three ways
Ask what outcome each regime wants (goal), what exact form it demands (form), and what that form carries or requires (payload). Agreements hide in the goal, divergences in the form, and conflicts in the payload; reading only the goal makes everything look green and ships the one incompatible act.
Run the analysis as a fixed procedure
Name the feature precisely, write each regime's core question before its rules, fill cells with dated specific obligations, read each row three ways, inspect the prescriptive-versus-rights pair hardest, and assign responses only after classifying. The step you skip is where the surprise lives.
A rulebook can conscript third parties
China obligates content-distribution platforms to detect and reinforce labels, so your distributor is part of your compliance system, not a neutral pipe. Your map of who must do what includes actors you do not control, which is a structural difference the UK's principles regime does not create.
Learn the philosophy, not just the current clause
Clause numbers and dates move; the reason each regime regulates (state control, individual rights, pro-growth restraint) is stable and predicts the shape of the next rule. An analyst grounded in the three "whys" survives a currency change that strands anyone who memorized only the text.
Adding a market is combinatorial, not additive
A fourth regime does not just add a column; it adds obligation-pairs with every existing regime, any of which can be a new conflict. Re-run the full classification when you expand, and never assume the new market resembles the ones you know.
Listen to it
The same lesson, talked through by two hosts as episode 44 of the podcast.
Real cases
These examples show the divergence problem in real regimes and real obligations, with the agreement, divergence, or conflict named explicitly. Where an event is the centerpiece of another topic, it is named only in passing and pointed to its owner.
Example 1: The anchor, AI-content labeling across the three regimes (2025 to 2026). A company shipping an image-generation feature globally meets three different answers to one question. China (CAC Labeling Measures and GB 45438-2025, in force September 2025) demands a visible label and machine-readable metadata carrying a provider code, to a national standard. The EU (AI Act Article 50, from August 2026) demands a machine-readable mark and, for realistic person-fakes, a deepfake disclosure. The United Kingdom demands no AI-specific label but leaves consumer and data-protection law in force. Read as a map: China and the EU agree on the goal of detectable synthetic media (agreement on the outcome), diverge on the form (a prescriptive national metadata standard versus an outcome plus code of practice), and carry a candidate conflict where the mandated Chinese metadata contents meet EU data-protection limits. The United Kingdom diverges by having almost no statutory obligation, which is a divergence to be read against underlying law, not a blank. This single feature contains all three relationships at once, which is why it anchors the topic.
Example 2: China's deep-synthesis and generative-AI stack as the pattern. The 2025 labeling measures did not appear from nowhere. They sit on the Interim Measures for the Management of Generative AI Services (in force August 2023) and earlier deep-synthesis rules, all run through the CAC. The pattern to learn is that China regulates AI content through prescriptive, centrally administered mandates layered over time, with security-review and registration expectations. An analyst reading the China column should expect specification, not principle, and should expect the specification to name the responsible provider, because attribution to an accountable entity is the design goal.
Example 3: The EU's Article 50 and its Code of Practice. In 2026 the European Commission published a Code of Practice on Transparency of AI-Generated Content to help providers meet Article 50. This is the EU's characteristic move: an outcome-based legal duty (mark it, detectably and interoperably) softened into practice by a voluntary code that fills in the how. Contrast this directly with China, which put the how into a mandatory national standard from the start. Same goal, opposite regulatory style: the EU legislates the outcome and codes the method as guidance; China codes the method as law. That contrast is a textbook divergence in style even where the regimes agree on the aim.
Example 4: The United Kingdom reading an absence. A UK-based studio using generative AI to produce marketing images faces no AI-content-labeling statute. A weak analyst writes "no obligation." A strong analyst writes that the Advertising and consumer-protection regime still forbids materially misleading content, that data-protection law applies if a real person's likeness or data is used, and that the DSIT Blueprint's principles (transparency, fairness, accountability) signal the direction of travel even without a statute. The United Kingdom is the case that teaches you to regulate a product against the law that exists, not only the law with "AI" in its title. (The broader question of what voluntary and principle-level instruments ever actually buy you is owned by (see Topic 6.5).)
Example 5: Divergence across time, not only space. Consider a synthetic-voice feature launched in early 2026. In China it must already be labeled to the 2025 standard. In the EU, if the system was on the market before 2 August 2026, the marking obligation does not bite until the 2 December 2026 grace date, then it does. In the United Kingdom, no dated statutory duty applies. The same feature has three different compliance timelines, and a map that records only "what is required" without "from when" will tell a team it is fine in the EU right up to the morning the duty switches on. Timing belongs in every cell.
Example 6: A near-conflict that resolves to a feature flag. A global platform wants one content-labeling implementation. It discovers that China requires a visible on-image label sized to a national standard, while the EU requires machine-readability and the United Kingdom requires nothing. Building the Chinese visible label for every user worldwide would degrade the product for hundreds of millions of users who do not need it, and building only the EU mark would leave China unlawful. The clean resolution is response 3, a per-jurisdiction feature flag: Chinese users get the national-standard visible-plus-metadata label, EU users get the machine-readable mark and deepfake disclosure, UK users get the platform's voluntary baseline. The cost is code complexity and a new duty to prove which behavior shipped where, which becomes evidence in the cross-border decision of (see Topic 6.6).
Example 7: The chatbot disclosure obligation, not just content labeling. Divergence is not limited to generated media; it reaches how a system announces itself. A customer-service chatbot that talks to people faces the EU AI Act's Article 50(1) duty that a person be informed they are interacting with an AI system unless it is obvious. China's generative-AI and deep-synthesis rules similarly push toward disclosure and registration of such services. The United Kingdom imposes no AI-specific disclosure statute, though consumer-protection and fairness principles point the same way. Read as a map, all three lean toward "tell the user it is a bot," which makes this largely an agreement in aim, with divergence in whether it is a hard legal duty (EU) or a principle plus general law (UK). The lesson: run the same three-way reading on the interaction obligations, not only the content-marking ones, because a global assistant can be compliant on its output labels and non-compliant on its self-disclosure.
Example 8: The grace-period trap. The EU AI Act gives systems already on the market before 2 August 2026 a grace period on the Article 50 marking obligation until 2 December 2026. A team reads "grace period" as "we are fine" and forgets to schedule the work, so on 2 December 2026 an unchanged, still-shipping feature becomes non-compliant overnight. The correct reading of a grace period is not relief but a deadline: it tells you exactly when the obligation bites. That is a gift to a planner and a trap to anyone who mistakes "not yet" for "not ever." This is divergence across time made concrete, and it belongs on the map as a dated task, not a green cell.
Example 9: The distributor becomes part of your compliance system. Under China's regime, content-distribution platforms must detect and reinforce labels, which means the platform that carries your AI-generated content is legally obligated to police it. A company shipping into China therefore cannot treat its distribution partners as neutral pipes. Their labeling duties interact with yours, and a mismatch (your content labeled one way, the platform expecting another) is a compliance failure that surfaces at the platform, not only at your product. The analytic lesson is that a rulebook can conscript third parties into enforcement, so your map of "who must do what" includes actors you do not control. This is a structural difference from the UK, where no such platform-policing duty exists by AI statute.
Example 10: When two regimes agree so well that the risk is complacency. Sometimes all three columns point the same way, and the danger flips from conflict to false comfort. For clearly harmful synthetic content (non-consensual intimate imagery, child sexual abuse material), China's rules, the EU (including the 2026 Digital Omnibus addition of an Article 5 prohibition on such material), and UK criminal law all prohibit it. An analyst might mark the whole row green and relax. The residual work is not "which regime is stricter" but proving your system actually prevents the prohibited output everywhere, because three agreeing prohibitions still require one working control, and a green row is a claim you must be able to defend, not a task you can skip. Agreement reduces conflict risk; it does not reduce the duty to build and evidence the control.
Example 11: The fourth column the three-column habit misses. A team maps a generative feature across China, the European Union, and the United Kingdom, marks Asia outside China as "no AI statute," and ships. South Korea's AI Basic Act, in effect since 22 January 2026, requires operators to tell users in advance that a product or service runs on AI and to label generative AI output, with stronger notice where synthetic sound, image, or video is hard to distinguish from the real thing, and it obliges a foreign provider above the Enforcement Decree's revenue or Korean user thresholds to appoint a domestic representative. None of that appears in a map whose only Asian column is China. The lesson is not that Korea is a fourth archetype every analyst must memorize; it is that a blank column asserts a verified absence, and this one was never verified. Note also the representative duty, which is a different species from a labeling rule: it is a condition on the provider rather than on the output, and a map built only from content obligations will not have a row for it.
Where people go wrong
- "Build for the strictest regime and you are covered everywhere." This works for agreements and simple divergences and fails exactly where it matters, at a true conflict, because the strictest rule in one country can be the forbidden act in another. There is also no single strictest country; strictness is per-obligation. China is strictest on label form, the EU on enforceability and fines, and the United Kingdom on nothing by statute. You assemble the highest common standard obligation by obligation, not by copying one country.
- "Compliance requirements just add up." They add up only when they are compatible. Divergence means more work without contradiction; true conflict means two requirements cannot both be met in the same artifact for the same user. Treating a conflict as if it were addable is the specific error that ships an illegal feature into a market you were not thinking about.
- "The United Kingdom has no AI law, so there is no obligation there." The United Kingdom has no comprehensive AI statute and no AI-specific labeling mandate, but existing consumer-protection, data-protection, defamation, and sector law still reach AI content. An absence of AI-specific law is not an absence of law. The correct UK cell reads "no AI statute; check the underlying law," never "nothing required."
- "China's rules are just a stricter version of the EU's." They rest on a different foundation. China's regime is state-directed information control expressed as prescriptive technical mandates that name the responsible provider; the EU's is rights-and-risk regulation expressed as outcome duties a person or regulator enforces. Same technology, different question being answered. Reading China as "EU plus more" hides the places where its prescriptive form carries something the EU constrains.
- "Transparency is transparency; one label satisfies the world." China's implicit metadata scheme and the EU's machine-readable marking both aim at detectable synthetic media but are not guaranteed to be interoperable. A label built to one standard is not automatically readable or sufficient under the other. Convergence on machine-readable provenance is emerging, not achieved; do not promise a single global label.
- "If it is lawful today it is lawful tomorrow." Divergence runs across time as well as space. China's regime is live now; the EU's Article 50 transparency duties apply from 2 August 2026 with a marking grace period to 2 December 2026; the United Kingdom has no dated statutory duty. A feature can be fully lawful in the EU today and unlawful on a known future date with no code change. Every cell needs a "from when," not only a "what."
- "Geo-fencing is easy; just block the country." Region detection is imperfect, because users travel, use virtual private networks, and share accounts, and enforcing a real market boundary is engineering work with its own failure modes. Geo-fencing is a legitimate response to divergence, but it is not free and it is not simple, and a map that assumes clean borders will be wrong at the edges.
- "Principles are the same as requirements." The United Kingdom's five principles (safety, transparency, fairness, accountability, contestability) describe direction, not specification. You cannot hand a principle to an auditor as proof of compliance the way you hand a met clause. Reading a principles regime well means translating its direction into concrete actions you choose, while knowing the enforceable floor comes from the underlying law, not the principle. (see Topic 6.5)
- "The divergence map is documentation, so it can wait." The map is an analytic instrument that feeds a real decision: which markets you ship into, with which behavior, made concrete in the cross-border decision of Topic 6.6. A map produced after the ship decision is an essay; a map produced before it is the thing that stops you shipping an illegal feature. (see Topic 6.6)
- "Read the goal and you have read the rule." At the goal level (transparency, detectability) the three regimes usually agree, which is exactly why goal-only reading is dangerous: it makes everything look green and hides the divergences of form and the conflicts of payload that live below the goal. You have not read a row until you have read it three ways: goal, exact form, and what the form carries or requires.
- "When all three regimes agree, we can relax." Agreement removes conflict risk, not the duty to build and prove the control. Three regimes all prohibiting a harmful output still require one working safeguard and evidence that it holds; a green row is a claim you must be able to defend under attack, not a task you get to skip. Complacency at agreement is its own failure mode.
- "A conflict is the vendor's problem if we bought the model." The obligations in all three regimes reach the provider and often the deployer and distributor; buying a model does not transfer the labeling, disclosure, or metadata duty away from you when your product ships the output under your name. Outsourcing the model does not outsource the divergence analysis. (see Topic 6.2)
Questions people ask
- What is divergence problem?
- The situation in which one product is governed by several independently written rulebooks at once, whose obligations do not simply add up because they were designed for different purposes and can point in different or opposite directions. The analytic task is to sort every obligation into agreement, divergence, or true conflict.
- What is agreement (between rulebooks)?
- A relationship in which two regimes want the same outcome and a single action satisfies both. The easy majority of cross-border obligations; noted and moved past.
- What is divergence (between rulebooks)?
- A relationship in which two regimes want different but compatible things, so you must do more than one thing, but the things do not contradict. Most cross-border complexity lives here; it means more work, not a breach.
- What is true conflict (between rulebooks)?
- A relationship in which doing what one regime requires would breach what another requires or forbids, for the same artifact and user. Rare, but the reason the whole analysis exists, because it is where "comply to the strictest" is not just costly but wrong.
- What is Cyberspace Administration of China (CAC)?
- China's central internet regulator, the lead agency (with MIIT, MPS, and NRTA) behind the 2025 Measures for Labeling AI-Generated Synthetic Content and the generative-AI and deep-synthesis rules that precede them. More on Cyberspace Administration of China (CAC)
Keep going
This lesson builds Framework crosswalking without false equivalence, and that page shows the roles that hire for it. Every Certified AI Governance Professional (CAIGP) lesson.
Watch the lesson and prove it
Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.
Watch the lesson and prove it →