Wright's Law works for individual technologies. This paper asks whether it works for the entire world economy. It does — and the gap between what we measure and what Paris demands is the climate problem in one number.
Wright's Law works for solar panels, batteries, and DNA sequencers. This paper asks whether it works for the whole world economy treated as a single technology converting resource inputs into GDP. Yes, it does. Energy intensity falls 18.3% per doubling of cumulative GDP. Carbon falls 22.2%. Materials — the inconvenient one — fall a statistically insignificant 6.5%.
Then we add the bill the planet pays in unpriced damages: climate impacts, biodiversity loss, health costs from pollution, depletion of natural capital. The "honest" learning rate falls to about 12.4%. Roughly 38% of apparent global economic progress is cost-shifting, not learning.
To stay on a 1.5°C path, the carbon learning rate needs to be 52%. We are at 22%. Carbon dioxide removal, learning at 12–20%, can buy us back about 17 percentage points of that gap. The remaining 13 percentage points have to come from political will, which has no experience curve.
This is a remarkable result. The figure sits comfortably inside the 10–25% range observed across individual technologies (solar, batteries, semiconductors). The world economy — the composite super-system of every technology, institution, and market — is learning at roughly the same rate as a single well-behaved technology.
Why this should not be obvious: the global economy has no outside. A solar panel manufacturer can externalise pollution to its surroundings. The world economy cannot. Yet the aggregate rate is still in the typical Wright's Law band. That tells us the same mechanisms — scale economies, knowledge spillovers, substitution, institutional adaptation — operate at the macro level, even without a clean external accounting.
Carbon is decoupling from output faster than energy is. Some of this is real (renewable substitution, efficiency, structural shift toward services). Some of it is the offshoring of heavy industry to economies whose emissions get counted under a different national flag. Both effects are present in the data and the paper does not try to separate them — the aggregate is what matters for the planetary constraint.
22.2% is a real achievement. It is also nowhere near enough, which is finding 5.
This is the inconvenient finding. The dematerialisation thesis — that economies can grow indefinitely while shrinking their physical footprint — does not survive the data. Materials extraction has roughly tracked GDP since 1970, with concrete, steel, sand, and metals leading the way. Every story about green growth has to reckon with this.
The paper is careful here: 6.5% is a positive point estimate, just not significantly different from zero. It is not evidence of no learning. It is the absence of clear evidence for learning, which is what you should take into a policy argument.
Two learning rates are estimated in parallel. The conventional one uses market GDP, the way it appears in national accounts. The inclusive one charges the economy for the damages it offloads to the biosphere and to future generations. The difference between them — about 38% — is the share of measured economic progress that is bookkeeping, not real efficiency improvement.
Useful framing for board rooms: the world's economy looks like a fast-learning technology when you look at the income statement. It looks like a slow-learning technology when you also look at the balance sheet of the planet. Both readings are true. They just answer different questions.
52% is historically unprecedented. No technology in the global record — including solar PV, batteries, or DNA sequencing — has sustained a 52% learning rate at a system level for thirty years. So the conventional path requires either a one-time discontinuity (a paradigm-speed shift, like the post-1980 collapse in semiconductor cost) or persistent failure to comply.
35% is different. Several individual technologies have sustained rates near 35% over multi-decade windows. The role of removal in the Paris arithmetic is to drag the required system rate from "historically unprecedented" to "demonstrated but difficult." That single shift is what makes Paris compliance plausible at all. It is also why the case for scaling removal is independent of any judgement about its current cost.
The Paris 1.5°C carbon budget translates to a required system-level learning rate. That number is 52% without removal, ~35% with it. Treat removal not as charity but as the variable that makes a real target plausible.
Conventional GDP-deflator decoupling overstates progress by roughly 38%. If you are running scenario models, run both rates — the conventional and the inclusive — and report the gap explicitly.
The argument for scaling removal does not depend on current cost. It depends on the gap between 22% and 52%. Removal is the variable that moves the required system rate from historically unprecedented to demonstrated-but-difficult.
Three numbers, one URL. Carbon: 22%. Required: 52%. With removal: 35%. Use them. Each finding has its own anchor for citation.