Nordhaus argues the learning-curve coefficient is statistically unidentified and biased upward — a zero-learning process still fits a coefficient of 0.2

In "The Perils of the Learning Model for Modeling Endogenous Technological Change" (NBER w14638, January 2009), William D. Nordhaus argues that the coefficient at the heart of Wright's law / the experience curve cannot, in the standard specification, be separated from ordinary exogenous technological progress. Because cumulative production and time both trend upward together, a regression attributing cost decline to learning-by-doing absorbs unrelated progress into the learning term: "the estimated learning coefficient will generally be biased upwards." His numerical demonstration is the sharp edge — in a constructed case with zero true learning, the fitted parameter still comes out positive: "the empirical learning coefficient is 0.2 even though the actual learning coefficient is zero."

The empirical instability is of the same order. Across 34 industries, "only 4 have estimated empirical learning coefficients in the plausible range between 0 and 0.5," and the correlation between two reasonable alternative specifications is 0.009 — effectively no agreement about which technologies learn fast. The forecasting consequence Nordhaus draws is that model-selected "high-learning" technologies (he names solar and wind) can have their future costs "underestimated by a factor of two."

This complicates the forecasting confidence carried by Sahal's identity and the monotonic-decline mechanism in claim-cheaper-extraction-disruptions-fall-monotonically-not-hold-then-collapse: the shape of a past cost curve can be fit cleanly whether or not endogenous learning was actually driving it, so a fitted curve is weak evidence that a collapse will continue. Nordhaus's critique is one economist's argued position within his own model, not settled field consensus; it is recorded here as his claim, verbatim from a Tier 1 primary, rather than as a proven fact about the world.