Benefit — is the world getting value, even where investors aren't?
Loading…
The value society gets back — time saved, work improved — set against the same capital cost. The most important number in the debate, and the one nobody else tracks as a live series. A bounded estimate, gross of harms; the harms are shown separately below.
Social return (Cs)
How it is built
Per-task productivity effects from randomised experiments × measured AI usage by occupation × wages × adoption, integrated over a horizon and set against the user cost. Published as a range with acknowledged biases in both directions — the experimental tasks were chosen because AI was expected to help, so this is not a floor.
Experiment
Measured effect
Domain
Noy & Zhang 2023 (Science)
−40% time, +18% quality
professional writing
Brynjolfsson, Li & Raymond 2025 (QJE)
+14% issues/hour
customer support
Dell'Acqua et al. 2023/26
+12% tasks, +25% speed (inside frontier)
consulting
Cui et al. 2026
+26% tasks
software development
Usage weights: Anthropic Economic Index. Wages: BLS OEWS. Adoption: Census BTOS. The productivity coefficient ρ is the AEI-occupation-share-weighted mean of these four trials (0.26–0.30; mid held at 0.27) — derivation in data/curated/rct_effects.csv.
The minus side — harms, tracked not netted
The benefit number is published before subtracting harms. Two indicators stand beside it rather than inside it.