Abstract
Katja Grace argues that calling AI development an “arms race” conflates two different things: the observable behavior of labs racing to build and deploy AI quickly, and the claim that this racing reflects each competitor’s genuinely best strategic move. Using a chess-scored-as-checkers analogy and a “Trojan horse race” metaphor, she argues that going fast may not actually be good for the racers themselves, which matters because “we have no choice but to race” is routinely used to justify both accepting high costs to others and giving up on coordination.
The core distinction
- Grace separates behavior (labs racing) from incentive structure (whether racing is actually the strategically correct move) — the fact that people are visibly racing doesn’t by itself establish that they are in a genuine race, in the sense of a competition where speed is the winning strategy.
- Her summary of the gap: “many people building AI very fast is different from being in a game where going very fast is the best selfish strategic move.”
The chess/checkers analogy
- Grace illustrates the distinction with an analogy: imagine chess champions competing intensely at chess, while all the prizes and recognition on offer are actually awarded for checkers performance.
- They would genuinely be racing — at chess — but the competition they’re actually being judged and rewarded on isn’t a chess tournament at all, so their strategy of playing fast chess would be badly mistaken if what they wanted was tournament success.
- The parallel to AI: labs may be racing to build and deploy as fast as possible, while the “prizes” that matter (long-term safety, market position, avoiding catastrophic mistakes) may not actually be won by raw speed the way the racing behavior assumes.
Why the distinction matters
Grace argues this distinction is not merely semantic, because “we’re in an arms race” is used to justify two specific and costly moves:
- Moving fast despite high costs imposed on others.
- Giving up on coordination or cooperation attempts, on the assumption that competitors are immovably locked into racing by the incentive structure itself.
- Both justifications depend on there actually being a race — a structure where fast is optimal — not merely on racing behavior being observed.
The Trojan horse race metaphor
- Grace proposes replacing “arms race” with “Trojan horse race” as a more accurate framing: various groups rapidly pull large wooden horses through their own gates, each uncertain whether their horse holds treasure or hidden enemy agents, driven by fear that rivals will get theirs through first.
- The question she poses: “Is it enough to know that other cities are pulling horses through their gates?” — i.e., does the mere fact that competitors are racing to deploy justify following suit, when nobody actually knows whether what’s inside will help or harm them?
- The metaphor is meant to capture that competitive urgency toward deployment doesn’t by itself tell you anything about whether deployment is good for you.
Implications for coordination
- If the real incentive structure is closer to chess-scored-as-checkers than to a genuine arms race, then coordination and voluntarily refusing to race remain live options — rather than being ruled out by an inevitability that doesn’t actually hold.
- The framing pushes back against treating “everyone must race” as a fixed fact about the strategic situation, since that framing itself can become self-fulfilling if left unexamined.