During the 1980s and 90s, Nobel laureate Robert Solow famously observed that you could see the computer age everywhere except in productivity statistics. Now a similar phenomenon is unfolding with AI — and it might be even worse. Seminalysis has coined a term for this blind spot: Dark Output. The concept describes AI-enabled economic value that exists but fails to register in GDP, prices, labor statistics, or industry accounts.

Two Flavors of Invisible Value

Dark Output splits into two distinct categories. Substitution dark output covers work previously done by humans that AI now handles — tasks with roughly $1.5 trillion in exposed labor costs where current-generation AI has credible displacement potential, according to Seminalysis' Dark Output Monitor. New dark output is trickier: it's work that nobody would have paid a human to do at prevailing wages, but which becomes economically viable once AI makes it essentially free.

The Legal Services Ghost

Consider a basic legal document. In theory, GDP should record the same inflation-adjusted value whether a lawyer or AI drafts it. But service sector accounting doesn't work that way — there's no "unit" of legal services, just receipts and surveys. When AI takes over drafting, receipts vanish as costs get absorbed into token spend. Meanwhile, government surveys show lawyers charging higher average prices because the simplest documents now go to AI instead of junior associates.

Why Services Can't Count What They Produce

Manufacturing automation gave economists something tangible: machinists who got better at making screws reported higher output, lower costs, and better margins. Real GDP captured this correctly because factories could count physical units. But there's no standard unit for legal services, no metric ton of literature reviews, no barrel of consulting.

Token Spend vs. Economic Truth

Anthropic's March 2026 Economic Index reveals that 37% of all tokens are being consumed in computers and mathematics occupations — yet contribution to GDP from software investment hasn't broken from its pre-AI trend. Incoming Fed Chairman Kevin Warsh acknowledged the problem directly in December 2025, stating that current data is "backward looking" and that policymakers risk being "late" to recognize non-inflationary growth from AI productivity.

The Measurement Gap That Matters

The $1.5 trillion figure isn't a claim that this much labor has vanished — it's exposed labor cost sitting inside categories where AI displacement economics are favorable. Seminalysis tracks evidence across tiers ranging from benchmark performance (weak signal) to actual business deployment claims, court cases defending AI work, and insurer underwriting of AI-generated output (strongest signal).

Key Takeaways

  • Dark Output describes AI-created value that vanishes from economic statistics while costs remain visible
  • Substitution effects (~$1.5T exposed) get counted as declines; new work creates value invisibly beyond token spend
  • Service sectors lack units of measurement — unlike screws, there's no way to count legal documents or literature reviews
  • The mismatch between heavy AI usage and flat GDP contribution in tech occupations suggests massive invisible productivity gains

The Bottom Line

We're essentially trying to measure a revolution with tools built for an industrial economy. Until statistical frameworks catch up — and Seminalysis argues this could take years, similar to the 2013 GDP revision that added $3.6 trillion for R&D and IP — we'll be flying blind on one of the most significant economic transformations in modern history.