AI's Benefits and Harms Are Not Symmetrically Distributed or Reversible

The accelerationist case for AI development rests on a compelling premise: given that advanced AI could in principle solve arbitrarily hard problems in medicine, physics, and materials science, the expected value of pushing forward is effectively unbounded. This argument deserves serious engagement. But it contains a structural flaw that is rarely articulated with precision. The upside and downside distributions are not symmetric in their causal relationship. Beneficial breakthroughs in one domain — say, oncology — do not mitigate catastrophic failures in another, such as autonomous weapons or algorithmic financial contagion. The upsides are domain-specific and additive; the downsides are systemic and potentially terminal. A single catastrophic loss-of-control scenario can foreclose the entire space of future benefits.

This asymmetry is compounded by a distributional mismatch. The positive tail of AI outcomes accrues disproportionately to a narrow set of actors: the firms and individuals who own frontier model capabilities and the infrastructure surrounding them. The negative tail — surveillance architectures, Labor Displacement, erosion of epistemic autonomy, ecological Externalities — is socialized across populations and ecosystems that have no seat at the development table. The expected-value framing beloved by techno-optimists implicitly aggregates across these populations as though gains to capital holders offset losses to everyone else.

The honest version of the abundance argument must therefore answer three questions simultaneously: abundance for whom, secured against which tail risks, and with what mechanism ensuring that catastrophic outcomes do not arrive on a faster timeline than beneficial ones. The asymmetry is not merely probabilistic — a matter of likelihood estimates — but structural, embedded in the causal and distributional architecture of the technology itself.