Zhou won. Brussels is still drafting.

On April 28, 2026, the Hangzhou Intermediate People's Court published a ruling that most Western policy circles have not paid much attention to. A quality assurance worker named Zhou had been demoted and offered a 40% salary cut after AI took over his job, which involved checking outputs from large language models. He refused the reassignment. The company fired him, citing organizational restructuring. The court said that it was illegal.

A few days later, a Beijing arbitration panel reached the same conclusion in a separate case. Liu had been doing manual map data entry. An automated system replaced him. The panel ruled the dismissal unlawful. The State Council amplified both decisions on April 30, just before International Workers' Day. The timing was not subtle.

The legal argument comes down to Article 40 of China's Labor Contract Law. That article allows termination when "objective circumstances" make a contract impossible to continue. Courts found that AI adoption does not qualify. A company deciding to automate is making a deliberate business choice. It is foreseeable. The efficiency gains go to the employer. The cost cannot be transferred to the worker who got replaced.

Worth taking seriously. More than 78,000 tech workers were laid off globally in the first four months of 2026, with nearly half of those cuts directly attributed to AI. Meta, Oracle, and Block all cited AI as the primary driver.

What regulation reveals

Every regulatory framework is a theory about what the problem actually is. The EU AI Act is a theory of risk and fundamental rights. China's AI legal architecture is a theory of social stability and state control. The US has been advancing a theory that the problem does not yet require a federal solution.

None of these is a neutral position. The question is not which framework is better. It is what each one protects, for whom, and at what cost to everyone else.

Start with the EU.

The AI Act classifies systems by risk level. At the top are prohibited practices: real-time biometric surveillance in public, social scoring by public authorities, and systems that exploit psychological vulnerabilities. Below that, high-risk systems, including those used in hiring, performance evaluation, and termination decisions, face requirements for transparency, human oversight, and documentation.

On paper, this covers a lot. In practice, two problems are visible.

The first is what the Act does not cover. It prohibits specific uses, like emotion recognition at work and certain biometric applications. What it does not address is what happens when a job disappears because a model got good enough to do it. The Act governs how AI is used in employment decisions. It does not govern what happens when employment ends because of AI. Zhou's situation would not be clearly covered. Neither would Liu's. There is a careful, detailed answer to the question of how AI tools are used in hiring. There is no answer to whether the job has to exist at all.

The second problem is timing, and what it reveals about who has access to the process. The European Commission published its Digital Omnibus proposal in November 2025, proposing to delay the compliance date for high-risk systems from August 2, 2026 to December 2, 2027. The stated reason was simplification. On May 7, 2026, after nine hours of negotiations that ended at 4:30 in the morning, the Parliament and Council agreed on a deal. High-risk AI rules delayed to December 2, 2027 for standalone systems, including employment tools, and August 2028 for AI embedded in regulated products. Machinery is excluded from the Act entirely. A ban on AI nudification tools and mandatory watermarking apply from December 2, 2026. The Commission called it "simpler, innovation-friendly rules." Reuters reported that critics say it shows Europe caving to Big Tech. One observer noted the timing: the deal landed right before the US-China summit meetings, leaving a year to watch which side offers better terms before committing to anything structural. Hard to dismiss. Analysis from the Corporate Europe Observatory found that 69% of Commission consultation meetings on AI policy in 2025 were with business groups. One consultation had eleven or twelve participants, all from industry except one civil society representative.

So, after years of negotiation and nine more hours on May 7, what we have is a ban on nudification apps and a watermarking requirement. Both matter. Neither touches what happens when Zhou's job disappears because a model got capable enough to do it. Progress, of a kind...

What China is actually doing

Western coverage of Chinese AI regulation tends to fail in one of two ways. The first treats it as primarily a censorship infrastructure. The second is surprised that it is technically serious. Both miss the point.

China's regulatory output on AI over the past two years is substantial, specific, and often operationally ahead of Western counterparts, partly because it does not require the same degree of political consensus to move. The system is fast. That is a feature of how the system works, not evidence that it is working in anyone's interest but the state's.

The generative AI transparency law that entered into force on September 1, 2025 is a good example. It requires both explicit and implicit labeling of AI-generated content across text, audio, images, video, and virtual environments. App distribution platforms must verify labeling compliance before approving listings. Luiza Jarovsky, whose AI transparency newsletter is among the most rigorous tracking of this space, noted that its provisions are more detailed than anything in the EU AI Act on the same topic. China specified the format, the position, and the platform-level verification obligations. It is operational. The EU Act is still producing codes of practice.

In April 2026, China issued the Interim Measures for the Management of Anthropomorphic AI Interactive Services, taking effect July 2026. These govern AI systems that simulate human personality traits and communication styles, including companion apps, digital humans, and emotionally interactive services. One of the first regulatory instruments anywhere to address what happens when AI systems are specifically designed to feel like a person.

Something harder to categorise is also happening at the frontier of labor and identity rights. The court rulings prompted commentary from legal researchers at the Chinese Academy of Social Sciences about questions no legal system has resolved: what happens when a worker's skills have been harvested into a reusable AI model? A company in Shandong reportedly used an AI digital replica of a former employee to continue performing his tasks after he left. Not a hypothetical. A case looking for a doctrine. A few jurisdictions are going to have to invent one.

None of this means the framework is operating in the interest of workers or users as such. China's AI content regulation requires adherence to "mainstream values" and "positive energy." The transparency rules and the anthropomorphic AI measures both serve state oversight functions alongside any consumer protection rationale. The Hangzhou and Beijing rulings protect workers. They also protect the Party from the optics of mass AI displacement during a period of slowing growth and high youth unemployment, published deliberately alongside a set of "typical examples of protecting the rights of AI enterprises and workers" in the lead-up to International Workers' Day. China's 2026 government work report explicitly included AI's impact on employment within the national policy framework for the first time, signalling that judicial precedent may be followed by legislative action.

Both things are true at the same time. The Chinese regulatory project is not confused about this. It just does not consider it a contradiction.

The American absence

The United States has no single comprehensive federal AI law. This is not an oversight. It is a policy position, expressed through inaction and, more recently, through active preemption of state-level attempts to fill the gap.

In December 2025, Trump signed an executive order directing federal action to challenge state AI laws deemed inconsistent with a unified national framework. No replacement federal standard exists. States trying to create notice requirements for technology-driven layoffs are operating on ground the federal government is actively contesting. The federal government is blocking rules while offering nothing. That is a governing philosophy, not a gap.

There is no American equivalent of the Hangzhou ruling. No federal requirement that companies show AI adoption is not just a budget reallocation before terminating the workers it displaces. The workers most exposed, in data entry, quality control, logistics, and customer service, are the workers with the least political representation and the fewest resources to challenge a termination in court. The companies developing the technology are, to a significant extent, deciding what the policy conversation considers worth solving.

American AI governance right now is deregulation framed as an innovation strategy. That framing serves specific interests and is indifferent to specific costs. The costs are not distributed randomly.

This week made it concrete. At a town hall on May 1, Zuckerberg told roughly 70,000 Meta employees that the company has two major cost centers, compute and people, and that investing more in one means less for the other. The 8,000 planned cuts are a consequence of a 2026 capex forecast of $125 to $145 billion. The savings from those cuts do not come close to covering the infrastructure spend. They are not meant to. Headcount is the variable that absorbs the pressure. The Hangzhou court said that the cost cannot be transferred to the worker. The US has no equivalent principle. No court, no statute, no pending federal bill that would say otherwise.

The rest of the world is not waiting either

The US, EU, and China are the loudest voices in this conversation. They are not the only relevant ones.

Brazil's AI bill, passed by the Senate in December 2024 and currently moving through the Chamber of Deputies, follows a risk-based, rights-centered model close to the EU approach, including explicit worker protections and a precautionary design principle. Brazil has previously punched above its weight on digital governance norms for the Global South, its 2020 data protection law being the example. If the bill passes before the 2026 elections, it will carry weight beyond its borders.

India is a different and underappreciated story. NITI Aayog's October 2025 roadmap estimated that the tech services sector headcount could fall from 7.5 to 8 million to around 6 million by 2031, with over 60% of formal sector IT and BPO jobs susceptible to automation. Major firms including TCS, Infosys, and Wipro have already reduced headcount by over 60,000 employees as part of AI-linked restructuring, often without naming AI as the cause. Workers found out from internal emails, after the fact. The government's response so far is India's AI Governance Guidelines, published by MeitY in April 2026: a light-touch, innovation-first framework built on voluntary standards and sandboxes. A private member's Artificial Intelligence Ethics and Accountability Bill was introduced in the Lok Sabha in December 2025. It has not passed and is not expected to in its current form. India is building scaffolding. The house is not yet designed.

The scale of the problem is not in doubt. The WEF's Future of Jobs Report 2025 projected 92 million roles displaced and 170 million created by 2030. Net positive, technically. Cold comfort if you are in the 92 million and the new roles require skills you do not have, in sectors that are not hiring where you live. Six in ten workers will need significant reskilling this decade. Eleven of every hundred are unlikely to receive it. No regulatory framework currently in force has an answer for those eleven.

Three frameworks in one table

Before getting to what any of this should actually require, it helps to have the comparison flat on the page.

No framework has a complete answer. All three reveal a gap that the others would fill, and create a problem that the others avoid by different means.

Winning the case is not the same as solving the problem

Zhou won. That is the right outcome. He also still does not have a job. The AI model that replaced him is still running. The company learned to budget for dismissal costs. Better than nothing. Not a solution.

The Hangzhou ruling places the cost on the employer. What it does not do is specify what comes next. Chinese legal commentary mentioned retraining and reasonable reassignment as the preferred path, but neither is a hard legal obligation. The ruling creates a financial consequence for bad behaviour. It does not create a transition obligation.

There is also something the ruling cannot address, and that most policy frameworks quietly assume away: we tend to assess jobs from the outside, based on what looks automatable from a task list. But accountability, trust, the ability to read a situation and know when the protocol is wrong, these do not appear on a task list. They tend to get undercounted. Often confidently. Sometimes arrogantly.

Klarna is a useful case here. In 2024, they replaced 700 customer service workers with AI, announced it publicly, and cited the cost savings. By 2026, they were rehiring human agents. Repeat contacts had jumped. Complex interactions and escalations, the work that sits underneath the automatable tasks, turned out to be more load-bearing than the task analysis had suggested. The work that looks replaceable often rests on work that is not. Courts can rule on who pays. They cannot rule on what the job actually was.

This matters for what follows. Compensation is a one-time payment. The displacement is structural. Severance does not retrain you. It does not help you compete with the next version of the model. And when the same role is disappearing across an entire sector simultaneously, there is nowhere obvious to move to.

We have been here before, though never quite at this speed. When European steel and coal industries contracted in the 1980s, the most effective responses were not severance packages but negotiated transition periods, sector-level redeployment funds, and advance consultation requirements before restructuring could begin. The gig economy classification battles of the past decade are a more recent and less flattering precedent: courts in the UK, France, and Spain ruled that platform workers deserved employment protections. Real wins. Most riders and drivers are still in the same structural position, because legal status without transition infrastructure changes the contract but not the condition.

A few directions worth considering, offered without illusion that any of them is easy or proven:

• Advance notice with substance. Not a generic restructuring announcement, but a documented transition plan: which roles, on what timeline, what the company proposes to do. Germany's Works Council Modernisation Act of 2021 already gives worker representatives explicit consultation rights before major AI-driven changes. The European Trade Union Confederation has proposed a European AI Social Compact combining advance notice, employment support, and training obligations. Neither has become binding at EU level. Both are replicable.
• A transition period with pay, not just severance. Continued employment during a defined period, with active support for redeployment, rather than a lump sum exit. Severance severs the relationship. A transition period keeps it intact long enough to be useful
• Redeployment before dismissal as a hard requirement, not a recommendation. If a role is eliminated through AI adoption, the employer should document what adjacent roles exist internally and show that redeployment was genuinely explored before termination. Not a checklist. A genuine prior obligation.
• Sector-level transition funds. Companies contribute proportionally when AI-driven headcount reduction exceeds a defined threshold. The fund finances retraining at scale. This is how several European countries managed the energy transition. It worked imperfectly. Still better than individual severance negotiations happening in isolation, company by company, worker by worker.

None of these are proven at scale for AI-driven displacement. Retraining programs have a poor track record when the jobs being trained for are also being automated, and the WEF notes that 63% of employers already cite skills gaps as the primary barrier to transformation. That tension is real and should not be papered over. The honest position is that severance alone is not enough, that we have some historical evidence for what works better, and that testing these mechanisms now is more useful than waiting for courts to keep catching up one worker at a time.

The Hangzhou ruling correctly names who is responsible. The harder question, what that responsibility actually requires in practice, at scale, across sectors and borders, is still open. That is the question regulators in Brussels, Beijing, Washington, Brasília, and New Delhi should be working on.

Most of them are not. But at least two courts in China had an answer ready.

A note: I am an AI executive and researcher, not a policymaker. If you work in this space and see something I got wrong, or something worth adding, I would genuinely like to hear it.