Report: ‘AI safety’ needs to mean safety from authoritarian abuse

An ASPI report documents how the Chinese party-state is integrating large language models and generative AI into censorship, mass surveillance, policing, courts and prisons—extending monitoring to minority languages and automating political filtering. The paper warns this state-centered AI architecture prioritises regime security and could drive global adoption of opaque, politically filtered models, prompting calls from democracies for procurement rules, disclosure mandates, export restrictions and human-rights due diligence. For investors, the findings underscore regulatory and reputational risks for AI vendors, potential restrictions on surveillance and opinion-management technology exports, and the strategic importance of transparency and compliance in AI supply chains.

Analysis

Market structure is rotating toward vendors that sell censorship, surveillance and auditability: expect recurring-revenue winners in cybersecurity and AI-audit tooling (CrowdStrike CRWD, Palo Alto PANW, Fortinet FTNT, Splunk SPLK) and defense/infra spend (NVIDIA NVDA via datacenter GPUs; onshore foundry investment beneficiaries). Losers are advertising-dependent Chinese platforms (BIDU, BABA, KWEB constituents) and Western platforms that rely on opaque LLM deployments or Chinese distribution. Pricing power will concentrate in specialized SaaS/security vendors with gross margins >60% and in semi suppliers where GPU scarcity gives 10–30% premium pricing in stressed export scenarios.

Tail risks include rapid export-control escalation or coordinated democratic regulation banning opaque political filters — a 0–12 month shock could remove 5–15% of GPU-addressable TAM for China-facing vendors and trigger >20% re-rating for exposed names. Immediate (0–30d) risk is sentiment-driven volatility; short-term (3–6 months) is regulatory drafting and export-control announcements; long-term (1–3 years) is ecosystem bifurcation and supply-chain realignment. Hidden dependencies: Chinese AI capability still depends on foreign GPUs, memory and datacenter tools, while Western moderation norms depend on commercial incentives that may shift absent regulation.

Trade implications: favor +security and +AI-audit exposure, overweight CRWD/PANW (target portfolio allocation 2–4% each) and selectively own NVDA (1–2%) with downside protection; underweight or short China-centric ad/consumer platforms (BIDU/BABA/KWEB) by 1–3% given policy and censorship demand risk. Use options to hedge regime-change risk: buy 3–6 month 10% OTM puts on NVDA sized at 30–50% of notional if export-control language tightens; consider 3-month straddles on KWEB for governance-driven volatility. Monitor legislative calendars: US/EU proposals within 60–120 days are binary catalysts.

Contrarian angles: the market may over-penalize NVDA if democracies stop short of draconian export bans — if no ban within 90 days, NVDA could reprice up 15–25% as GPU demand stays strong. Conversely, the market underestimates growth in private/third-party AI-audit startups and open-source explainability tools which could capture 10–20% of compliance budgets by 2026; look for takeover targets. Historical parallel: 2018–20 Huawei supply shocks show swift capex rerouting and domestic substitution over 12–36 months; expect similar cadence here.

Key Decisions for Investors

• Establish a 2–3% long position in CrowdStrike (CRWD) within the next 10 trading days, target 12–18 month upside 20–35%, set a tactical stop-loss at -12% to reflect policy/earnings risk; rationale: immediate secular demand for AI-enabled security and censorship-resilience.
• Initiate a 1–2% long position in NVIDIA (NVDA) but buy 3-month 10% OTM puts sized at 50% of the long notional as insurance; if US/EU announce tighter GPU export controls within 60 days, increase put hedge to 100% notional and trim NVDA exposure by 50%.
• Enter a pair trade: long 2% CRWD and short 2% Baidu (BIDU) over a 3–9 month horizon to express security/censorship premium vs China internet regulatory risk; rebalance if PRC policy signals ease or U.S. bans are lifted.
• Reduce exposure to China ad/consumer internet ETFs (KWEB) or individual names (BABA, BIDU) by 50% within 30 days and redeploy proceeds (+3% portfolio) into onshore semiconductor ETF SMH or SOXX to play supply-chain reorientation; if KWEB falls >20% on regulatory headlines, consider buying 3-month straddles to capture volatility.