AI Dark Output: The Visible Cost of Invisible Output
Malcolm Spittler
· SemiAnalysis
· May 29, 2026 at 20:02
· ⏱ 26 min read
| Read on Substack ↗
Summary
The article argues that AI is generating substantial real economic value ("dark output") that remains invisible to GDP and inflation statistics, mirroring the computer revolution's productivity paradox but on a far larger scale — especially in services. This measurement failure risks misreading the AI boom as a bubble, even as visible costs (data centers, tokens, jobs) pile up, and warns that policy and investment decisions based on flawed data will be systematically wrong.
•Robert Solow's 1987 quip 'You can see the computer age everywhere but in the productivity statistics' is the historical parallel; the article claims AI's measurement problem is much larger.
•A 2013 methodology revision added R&D and IP to GDP accounting, boosting 1990s output by ~$3.6T — nearly 30% of 2000 GDP — showing how statistical revisions can change perceptions.
•The article identifies roughly $1.5T in tasks that current-gen AI could substantially augment or automate (the 'Dark Output Monitor' exposed labor estimate at Tier 4+ evidence).
•An illustrative legal will dropped from ~$400 in 1990 to ~$0.50 in 2026 via AI, a >99% cost decrease that GDP accounts cannot capture because service sector statistics lack quantity units.
•According to the Anthropic Economic Index (March 2026), 37% of tokens are used in computers and mathematics, yet software investment GDP has not broken from its pre-AI trend.
•The article's 'Evidence Ladder' ranks market signals from Tier 1 (benchmarks) to Tier 6 (insurer underwriting risk); most current evidence is Tier 3-4 (public claims and production use), not full replacement.
•The Feminist Economics appendix notes that 16.4 billion hours of unpaid care work worth $11 trillion annually are invisible to GDP — a precedent for AI-generated output also falling outside the production boundary.
•The article warns that without better measurement, AI's costs (power, water, jobs) remain hyper-visible while its output disappears, creating a systematic bias toward calling it a bubble.