Anthropic's Alarming 2028 AI Geopolitics Paper: More Briefing Than Safety Research
Original: Anthropic just published a pretty alarming 2028 AI scenario paper and it's not about AGI safety in the usual sense View original →
The Core Argument
Anthropic's new paper describes two possible worlds by 2028: one where the US maintains its AI lead, and one where it loses it. The paper reads less like AI safety research and more like a geopolitical intelligence briefing — which is why a Reddit summary of it earned 560+ upvotes on r/artificial.
The Compute Gap
The paper argues that the US currently holds a real advantage through its control of the semiconductor supply chain — NVIDIA, TSMC, ASML — that China cannot yet replicate. Export controls have made this gap tangible. But Chinese labs are closing it via two methods:
Chip smuggling: PRC labs are training on export-controlled chips they shouldn't have. A Supermicro co-founder was charged with diverting $2.5B worth of servers to China.
Distillation attacks: Creating thousands of fake accounts on US AI platforms, harvesting model outputs at scale, and using them to train competing models — effectively free-riding on billions in US R&D investment.
Two Scenarios for 2028
Scenario 1 (good): US closes loopholes, compute gap widens to 11×, US models stay 12–24 months ahead, democracies set global AI norms.
Scenario 2 (bad): US fails to act, China reaches near-parity, floods global markets with cheaper models, CCP shapes AI governance norms and exports AI-enabled surveillance tools to authoritarian governments.
Anthropic's Ask
The paper explicitly calls distillation attacks 'industrial espionage' and advocates for legislation to criminalize them. Whether this is genuine policy analysis or sophisticated lobbying is the central debate in the comments.
Related Articles
Anthropic published a policy paper urging democratic countries to maintain their AI advantage over China before a critical window closes in 2028, framing advanced AI as a geopolitical asset.
Claude Corps is a $150m program placing 1,000 early-career fellows into at least 400 nonprofits for 12 months. The bet is that AI adoption needs paid capacity inside civic organizations, not only better models.
The HN debate centered less on downtime and more on whether a narrow jailbreak concern is enough to halt access to commercial frontier models.