China doesn't care about AI safety? Two different meanings of 'safe'

Caveat: I have not been back to China in several years, so most of this piece is second- or third-hand.

In the US, the default view is that China does not care much about AI safety right now. When I talk with people in the Chinese industry, they say the opposite: everyone is doing AI safety — just quietly — and Chinese people trust the government more; when the government wants to regulate something, top-down can land fast.

First we need to align: what “AI safety” means

In Berkeley / Anthropic / MATS circles, AI safety usually means:

Loss of control, deception, recursive self-improvement
Existential risk, superalignment
Independent evaluators, RSPs, pause advocacy
Civil-society pressure on labs and governments

In mainstream Chinese public discourse, AI 安全 more often means:

Content compliance (no incitement, no threats to social stability)
Deepfakes and fraud (top public concern in surveys)
Data security, algorithm filing, generative-AI service standards
“Science and technology ethics” review — human welfare, controllability, trustworthiness

Both can be serious. Different objective. One optimizes technical loss of control and goal alignment; the other optimizes state control and social stability. So the precise claim is not “China doesn’t care about safety.” It is: you may not be talking about the same kind of safe — at least not yet, not in public or industry narratives that treat Western alignment as a priority.

What China is already doing

When people in industry say “we are already doing it,” that is largely true if they mean regulatory AI safety. China has one of the world’s most complete pre-deployment review stacks for GenAI: algorithm recommendation rules, deep-synthesis rules, interim measures for generative AI services, national standard GB/T 45654-2025. By end of 2025, 748 GenAI services had completed filing (more in A map of AI governance).

Recent moves include, among others:

“Qinglang: Rectifying Chaos in AI Applications” campaign (Cyberspace Administration of China, from April 2026, ~four months): two phases — first source issues (model filing, safety review, training data, data poisoning, synthetic-content labeling); then content harms (false information, “digital slop,” impersonation, harms to minors, online astroturfing). Phase-one public numbers: 14,000+ non-compliant AI products handled, 6M+ illegal/violating items cleaned.
Interim Measures for the Administration of Anthropomorphic AI Interactive Services (CAC and four other ministries, published April 2026, effective 15 July 2026): rules for services that simulate personality and sustained emotional interaction — clear “this is AI” labeling, minor protection, safety assessment and algorithm filing, intervention for extreme distress. Product/content regulation, not alignment evals.
CnAISDA launched Feb 2025; April 2025 Politburo study session on AI cited “unprecedented risks and challenges”; May 2026: ten ministries issued trial measures on AI science and technology ethics review.

On “don’t let non-compliant models/apps go live,” China may be the only major power with mandatory GenAI pre-deployment review — more aggressive than the US. The standard is compliance and controllability, not “passed a deceptive-alignment eval.”

The public really isn’t panicking

The data backs this, and the gap is large.

Dimension	China	United States
Net benefits outweigh harms	80–83% (Ipsos 2024)	39%
Concern vs excitement	High adoption and enthusiasm; only 5.3% think they’ll be unaffected	~50% more concerned (Pew 2025)
Top worries	Deepfakes/fraud 71.3%	Jobs, misinformation, privacy
Who should act	Government and researchers	More diffuse; low trust in firms (Gallup: 31% trust business with AI)

This is part of a developed vs developing optimism gap: China, Indonesia, and Thailand are most bullish; the US, Netherlands, and Canada are most cautious. Edelman 2024: institutional trust in developing countries 63%, developed 49%.

“Cultural factors” here is not just “trust government”:

Technology nationalism — “AI China moment”; falling behind is the default fear, not runaway superintelligence
Different anxieties — fraud and jobs, not loss of control
Media ecosystem — weak x-risk narrative; post-DeepSeek discourse is pride and competition, not pause letters

Globally, only about 3% of AI researchers name existential risk as a top open-ended concern in surveys. China sits at an extreme of that spectrum — not an exception to it.

Is top-down really fast?

Industry people often say: when the government wants to act, it lands fast. Empirically, yes — depending on what is being regulated.

Actually fast: content and deployment. Filing, takedowns, rectification orders, account bans, Qinglang campaigns — enforcement is stronger than US federal practice. The anthropomorphic-interaction measures go from publication to effect in a few months. A special campaign can publish numbers like “14,000+ products handled” within months. That is not empty talk.

Not that fast / easy to misread:

CSET on GB/T 45654-2025: detailed standard, but enforcement uncertain under cutthroat domestic competition — ship-first-file-later is common.
Inter-ministry friction and policy whiplash: top-down can also be slow and messy.
Carnegie’s CnAISDA analysis is blunt: its primary function is international representation, not independent frontier-model testing like UK/US AISIs; CCP backing reflects global-participation ambitions at least as much as deeply held alignment concern.
China has emergency-stop authority, but triggers are government-defined disasters, not METR-style autonomy evaluations.

So “top-down is fast” holds as a content-governance empirical claim. Reading it as “alignment / loss-of-control is already handled” swaps categories.

What I think is still missing

If you switch the ruler to Western AI safety — independent research, reproducible evals, public pressure on frontier capability — the gaps are obvious, and they are not just “has the government issued a document.”

For researchers:

Tiny independent alignment / x-risk community; almost no civil-society AI safety orgs, mostly CAICT, Tsinghua CISS, and other state-adjacent research.
The topic is easy to frame geopolitically: “helping America slow China down.” Open work on deceptive alignment, RSI, loss of control has narrow space and high narrative cost.
Missing an ecosystem of citable external evaluators of frontier closed/open models — METR, Apollo, independent red teams. CnAISDA is not an AISI.

For entrepreneurs:

Heavy compliance load (filing, labeling, content review, new anthropomorphic rules), but product incentives sit almost entirely on growth and applications. Few can survive on an “alignment product.”
Day-to-day safety work is content filters, refusals, data compliance — necessary, but a different job from “will the model strategically deceive.”
Funding and narrative: a safety-as-mission startup story is hard to tell in China. It is hard in the US too, but at least there is an EA/philanthropy side channel. That channel barely exists in China.

Elites are warming (Andrew Yao, Xue Lan, Fu Ying and others enter international frontier-risk conversations via CnAISDA). That is still not the same as Reddit-style public panic, or independent-lab research infrastructure.

One table to keep straight

                    Western alignment safety          China's regulatory safety
                    ──────────────────────          ─────────────────────────
Public salience      Medium (jobs > x-risk)          Low (deepfakes > x-risk)
Civil society        Many (METR, FLI…)               Almost none
Government action    Weak (US deregulating)          Strong (filing, review, campaigns)
Actually optimizing  Technical loss of control       Social stability + industrial competition
2025+ shift            Industry RSP self-reg           CnAISDA, Qinglang, anthropomorphic rules, ethics review

Closing

I want more US–China discussion — first align on what “AI safety” means, then research together. Two countries that hold most of the world’s frontier compute cannot cooperate or build trust if they are not even using the same word.