AI Economics & Policy
Self-directed design work on AI's economic and safety footprint, where every protocol is written to be attacked by its own simulation before it is ever run on real data.
I designed identification strategies for five open questions regarding AI’s economic footprint: the enterprise productivity J-curve, capability overhang and non-linear risk emergence, automation-versus-augmentation wage divergence, the economics of malicious use, and cross-jurisdiction regulatory arbitrage.
Status in one line: all five are design-stage. Four have run no real data; one, capability overhang, has been run on public benchmark data and is described below. The figures in the next two sections come from design simulations with a known truth, not from empirical findings.
Breaking the Proposals First
Before running anything on real data, I wrote a design simulation for each protocol that deliberately attacked its own proposal against a known truth. In those simulations the proposals failed:
- A J-shape arises with probability ~1 under a zero-effect process, so a dip-and-recovery pattern alone cannot be read as evidence of an effect.
- A Poisson count regression on reported cyber incidents rejected a true null in 100% of simulated draws, falsely attributing a +160% increase to open-weight diffusion when the true attack count was constant.
- A compute-location coverage bias flipped a simulated estimate from -40% to +64% under a true null, with the analyst unable to sign it from inside the data.
Catching Specification Errors Early
This verification-first approach caught specification and algebra errors that no sample size could ever repair. For instance, in simulation, conditional quantile regression recovered only 0.3% of a wage compression that an RIF estimator recovered in full.
I then rebuilt each design around what free, public data can actually identify: nested multi-frequency architectures, IRT latent indices over heterogeneous indicators, and measured rather than assumed coverage functions, unified by a shared measurement layer and power analysis across all five studies.
Where Each Study Stands
- Productivity J-curve: paused by explicit decision after a Phase 1 pilot. The pilot used small convenience samples and is not a powered test, and I chose to stop there rather than present pilot-scale estimates as results.
- Capability overhang: the furthest along, and the only one run on real public benchmark data. Its measurement layer reproduces a published method (an IRT-based capability index), and its first headline effect includes an oracle-selection component that a realistic verifier would only partly capture, so I treat it as preliminary and unsubmitted.
- Wage divergence, malicious use, regulatory arbitrage: protocols, literature reviews, and executed design checks; no empirical data pulled yet.