The clock that measured a policy backfire

For a year, the argument that US chip controls handed China the open-source lead has been an op-ed staple — RAND said it, the US-China review commission said it, Perplexity's CEO said it. Nobody had put a number on it. A new arXiv paper from five academics, led by the University of Chicago's James Evans, does: it treats four dated US policy shocks (the 2022 CHIPS Act, two rounds of advanced-computing controls, and a 2024 export-rule revision) as natural experiments and counts what happened to forking activity on open language-model repositories in the weeks after each one.

Controls meant to slow China's ecosystem coincided with its developers forking open models at eleven times the US rate — the containment measure and the acceleration land in the same weeks.

The authors report that every control event added about 0.143 forks per repository-week among China-linked developers, versus 0.012 among US ones — an eleven-fold gap. Read plainly: the measures meant to slow China's AI ecosystem coincided with China-linked developers copying open models to build on at eleven times the American pace. The interpretation the paper draws is that Beijing's ecosystem started treating open weights as insurance against being cut off upstream.

Two things a careful reader should hold. The China-versus-US split is not verified nationality — it is a clock: any GitHub event logged before 17:00 UTC is counted as China-associated. That proxy is clever at the scale of millions of events and a genuine soft spot, and the precise coefficients live only in the paper's own text, uncorroborated elsewhere. And a fork is a cheap gesture — a copy, not a contribution or a shipped product. What the paper measures is a surge in the first rung of engagement, timed to policy. The quieter finding sits underneath it: even as that forking surged, Chinese-origin models stayed thin in US patents — heavy open-source uptake, light formal capture.

The value here is not the thesis, which is saturated, but the evidence class. It arrives as the US debates whether to extend export controls to open-source software itself — and hands both sides a falsifiable number to argue over.