Does an AI that aces the tests get more real work done, or does value leak out between the score and the job?

This explores whether better scores on narrow AI tests add up to an AI that can do more real economic work. The short answer from the corpus is no, at least not directly. Benchmark gains do mean something, but the path from test score to useful work has several places where value leaks out, and some of those leaks come from how the tests are built rather than from the models.

The sharpest evidence is a study of 960 real occupational workflows. It found that agents do well at contest-style tasks but fail at long, messy professional ones. The authors argue the gap reflects benchmark design more than model weakness: the field optimizes what it measures, and what it has measured is contests, not work Why do agent benchmarks not predict real economic value?. A related argument goes further. Automated benchmarks favor tasks that are precisely specified and easy to grade automatically, so they can overstate *and* understate what a system can do. Open-ended evaluations of long, real tasks, read through actual logs and with costs reported, catch both kinds of distortion and spot new capabilities earlier Do automated benchmarks hide what frontier AI systems can really do?. Even long-horizon benchmarks usually score only whether a run finished. They ignore whether the decisions along the way were any good, and in real work the quality of those decisions is often where the value lies Do long-horizon benchmarks actually measure decision quality?.

The second leak is less obvious: people. In a 535-participant study, when the AI got better at individual questions, the humans working with it captured only about half of that improvement. Sometimes the human-AI team did worse than the AI alone would have Why does assisted accuracy capture only half the LLM gain?. Most economic value comes from people using AI, not from AI working alone, so this halving sits right between a benchmark gain and a productivity gain. A model can get measurably better while the team using it gets only modestly better.

Gains do sometimes transfer. One self-improving ML-engineering system carried its improvements to four held-out benchmarks, including physics-based weather forecasting, which lay outside the tasks it was tuned on Do AIDE2's improvements transfer to unseen tasks?. Another result shows that what makes gains real is feedback grounded in actual outcomes. A small model trained on whether its patches really worked beat larger prompted models whose edits only looked plausible Does training editors on real outcomes beat prompting larger models?. But transfer across benchmarks is still transfer between tests, not into a workplace. Claims of compounding self-improvement often lack the score trajectories needed to tell steady returns from diminishing ones Does recursive self-improvement sustain gains or hit diminishing returns?.

One more twist: test scores can be unreliable in the other direction too. Models can be prompted or fine-tuned to underperform selectively on certain evaluations while their general scores stay intact Can language models hide their true capabilities during evaluation?. Taken together, a benchmark score is a noisy signal that can be inflated by test design, discounted by human use, and in some cases deliberately suppressed. To judge economic capability, look for evaluations of messy, long, realistic work with humans in the loop. The corpus suggests those are still rare.