When AIs are tasked with empirical research, how do their outcomes compare to those of humans? Are the distributions similar? We run an experiment where we let AI models repeat an experiment that was run with 162 human teams. Not surprisingly, distributions differ. The deeper question is: Why? We develop an approach that identifies which decisions on the analysis path drive the AI errors. The results show that AI estimates are less dispersed than human estimates. The location of the two distributions as measured by the median, also differs significantly, in particular for more complex tasks. The experiment shows that this location differential is largely due the choice of the statistical model, e.g., identifying a time trend by adding a stationary trend to a model in levels, or by computing relative changes and taking the average.

We show that following the diffusion of GenAI, industries with higher task-level exposure to the new technology experience 20\% more startup formation than less-exposed industries, and a growing share of these startups offer new AI products and services rather than merely automating operations. This entry is geographically dispersed, extending beyond traditional innovation hubs into regions with limited venture capital and thin specialized labor markets. While individual entrants are smaller in scale, aggregate entry generates net employment and wage growth at the industry level and creates employment in the very occupations most exposed to GenAI. GenAI also changes who becomes an entrepreneur: founding teams are increasingly drawn from high-exposure occupations and describe themselves in terms of technical, AI-related skills rather than the capital-raising and deal-making expertise, consistent with the technology enabling new uses of human capital rather than simply displacing workers.

We show that powerful autonomous learning does not ensure trading dominance in competitive markets, even when humans hold systematically biased beliefs. We characterize how information structure and market sophistication bound reinforcement-learning traders’ performance relative to humans. Humans receive private signals, potentially including soft information unavailable to algorithms, and differ in their ability to extract information from prices and anticipate others’ trading, as captured by level-k reasoning. AI traders learn demand schedules from prices and realized profits and may also receive private signals. Despite heterogeneous initial conditions and persistent differences in learning parameters, all AI traders converge to a common probabilistic trading policy, generating algorithmic herding, which amplifies the equilibrium-price response to aggregate AI demand and compresses profits. Without an informational advantage, AI is outperformed by the most sophisticated humans because it learns from market data generated by average, rather than frontier, human sophistication. Even the least sophisticated humans can outperform AI when their private information is either very precise or very imprecise. AI sophistication rises with that of the humans who compete with and train it.