AI benchmarks are a 'dark art.' Moritz Hardt has a fix.

Moritz Hardt has a message for anyone who has ever picked an AI model because it topped a leaderboard: the ruler you are using is broken. Hardt, author of a new book on AI benchmarks, argues that the standardized tests that dominate artificial intelligence are fundamentally flawed, but he insists they are fixable. His core claim is that benchmarks are not just measuring sticks; they are also rulers in the sense of commanding the field's behavior, shaping what researchers build and what companies buy. That dual role is exactly why a broken benchmark is not a minor technical annoyance. It is a structural problem that distorts the entire AI economy.

The stakes are immediate. Benchmarks are how the industry decides which model is best for coding, for reasoning, for customer service, for almost every commercial use case. If those tests are broken, then every procurement decision, every performance claim, and every safety comparison built on them inherits the flaw. Hardt's book arrives as AI spending and deployment accelerate, making the question of how we measure progress not academic but commercial. A company that chooses a vendor based on a flawed benchmark may be building its entire AI strategy on a mirage, while competitors who dig deeper can capture the same capability at lower risk and lower cost.

The history runs deep. In 1987, the Defense Advanced Research Projects Agency (DARPA) issued its first benchmark on automatic speech-to-text, setting a template that the field has followed ever since: create a standardized test, run models against it, rank the results. That template has produced enormous progress, but Hardt argues it has also created perverse incentives. When a benchmark becomes a target, researchers and companies optimize for the test itself, not for the underlying capability the test was meant to measure. The result is a field that measures what is easy to measure rather than what matters, and a market that rewards models that game the test rather than models that genuinely understand.

The problem is not that benchmarks are useless, Hardt argues, but that they are treated as definitive when they are actually brittle. A model can ace a test and still fail in the messy, unpredictable conditions of real deployment. This is not a hypothetical concern. The history of AI is full of models that looked impressive on standardized evaluations and then stumbled when confronted with noisy, ambiguous, real-world inputs. For executives, the implication is concrete: if you are choosing between AI vendors based on benchmark scores, you may be comparing artifacts of the test design rather than real-world capability. If you are setting internal performance targets for AI features, those targets may be driving your team to optimize for a leaderboard instead of for your customers. And if you are relying on benchmarks to justify AI safety or compliance decisions, you are building on a foundation that Hardt says is not yet scientific.

Hardt's proposed fix is to make benchmarking a science, not a dark art. That means treating benchmarks as hypotheses about model behavior that need rigorous validation, rather than as permanent verdicts. It means designing tests that are harder to game, that measure capabilities in context, and that are updated as models and applications evolve. The book is a call for the field to apply the same standards to its measurement tools that it applies to its models. Just as a good scientist would never trust a single experiment without replication, a good AI buyer should never trust a single benchmark score without probing the assumptions behind it.

The timing matters. AI benchmarks have become a battleground for market share, with vendors touting scores on popular tests as proof of superiority. Regulators and enterprise buyers increasingly ask for evidence of model performance, and benchmarks are the default answer. If Hardt is right, that default answer is currently unreliable, which means the companies that take benchmarking seriously as a discipline, rather than as a marketing tool, will have a real edge. They will be the ones who can separate signal from noise, who can identify which models actually work in their specific environment, and who can avoid the costly mistake of betting on a model that looks great on paper and fails in production.

The strategic takeaway for leaders is simple: do not let a benchmark score make your decision for you. Use benchmarks as one input, probe the assumptions behind them, and demand evidence that a model's performance holds up in your environment. The companies that treat AI measurement as a science will not only make better procurement decisions; they will also build internal evaluation capabilities that become a competitive moat. In a market where every vendor claims to be the best, the ability to rigorously verify those claims is a genuine advantage. Hardt's book is a reminder that the tools we use to measure AI are just as important as the AI itself, and that the dark art of benchmarking is ready for a scientific reckoning.