Summary of Erik Brynjolfsson, “The Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence,” Stanford Digital Economy Lab, forthcoming in Dædalus, Spring 2022 (arXiv:2201.04200v1).

Core argument

Brynjolfsson argues that Alan Turing’s “imitation game” set a misleading north star for the AI field. By making human-like intelligence (HLAI) the implicit goal, researchers, executives, and policymakers have been biased toward building AI that substitutes for human labor rather than AI that augments it — a bias that is economically and politically dangerous.

Two paths for AI

  • Automation (substitution): Machines replace workers task-for-task. Gains accrue mostly to capital owners, entrepreneurs, and platform architects. Workers’ marginal value drops, weakening wage bargaining power and political leverage.
  • Augmentation (complementation): Machines extend human capabilities, creating new products, services, and tasks. More of the surplus flows to workers (a bulldozer operator earns far more than a shoveler), and overall value created is larger because it expands the production frontier rather than just redistributing it.

Historically, augmentation — not automation — drove the >10× rise in real wages since 1820.

The “Turing Trap”

The trap is the combined pull of three misaligned incentives toward automation over augmentation:

  1. Technologists use “matches a human” as the benchmark because it is measurable.
  2. Business executives find plug-and-play labor substitution easier than redesigning processes around human+AI teams.
  3. Policy (e.g., U.S. tax code favoring capital over labor) tilts the build-vs-hire math toward replacing workers.

Why it matters

If the trap closes, the result is not just inequality — it is a concentration of economic and political power in whoever controls the AI, since workers lose the leverage that comes from being needed. Augmentation-first AI keeps that leverage distributed.

Recommendations

  • Researchers should chase problems that are hard for machines and easy for humans (and vice versa), not Turing-style mimicry.
  • Businesses should reinvent processes around human+AI complementarity, not swap-in automation.
  • Policy should rebalance tax treatment of labor vs. capital, and heavily subsidize worker training (Brynjolfsson cites a ~$9 of human-capital spend needed per $1 of ML spend).

One-line takeaway

Human-like AI is a seductive but narrow target; human-complementing AI generates more total value and distributes it more broadly — and the choice between them is one we are actively making right now through research priorities, business decisions, and tax policy.