The dominant framework for measuring AI progress — pitting models against individual humans on discrete tasks like chess, math, or essay writing — is broken, and the field needs a replacement, according to a detailed analysis published by MIT Technology Review on March 31, 2026.

For decades, AI evaluation has relied on a deceptively simple question: can the machine beat the human? The appeal is obvious. Head-to-head comparisons produce clean numbers, legible headlines, and the kind of milestone moments — a computer beating a chess grandmaster, a model passing a bar exam — that signal genuine progress. But researchers and critics increasingly argue this framing obscures more than it reveals.

The Problem With Beating Humans at Isolated Tasks

The core issue is that real-world AI deployment looks nothing like a controlled benchmark. When a model answers a standardised math problem correctly, it tells us little about whether that same model can reliably assist a student across a full semester of coursework, handle ambiguous or incomplete inputs, or avoid confidently wrong answers when the stakes are high. The human-vs-AI comparison on isolated problems, critics say, creates a false equivalence between narrow task performance and genuine capability.

Benchmarks also suffer from a structural vulnerability: contamination. As AI models are trained on ever-larger sweeps of internet data, the probability that training sets include the exact questions used in evaluation grows substantially. A model that has, in effect, seen the answers is not demonstrating reasoning — it is demonstrating recall. This makes high benchmark scores increasingly difficult to interpret with confidence.

An AI that aces a standardised test but fails unpredictably in deployment is not a capable AI — it is a well-rehearsed one.

How the Metrics Game Distorts Development Priorities

The pressure to post competitive benchmark numbers has downstream consequences for how AI systems are actually built. When a specific benchmark becomes the accepted measure of a capability, developers face strong incentives to optimize for that benchmark specifically — a dynamic researchers call Goodhart's Law in practice: once a measure becomes a target, it ceases to be a good measure. The result is that leaderboard rankings may reflect benchmark-specific tuning as much as genuine underlying ability.

This matters because benchmarks don't just describe the field — they shape it. Funding decisions, hiring, product launches, and regulatory discussions all reference benchmark performance as a proxy for real-world readiness. If those proxies are unreliable, decisions built on them carry hidden risks.

What Better Evaluation Would Look Like

Researchers pushing for reform generally agree on several principles, even if consensus on specifics remains elusive. First, evaluations should test end-to-end task completion in realistic conditions rather than isolated subtasks. Instead of asking whether a model can solve a single coding problem, evaluators might ask whether it can complete a full software project with realistic constraints, ambiguous requirements, and error correction.

Second, evaluations should incorporate adversarial and edge-case testing — deliberately probing the boundaries of model performance rather than measuring only central-tendency success rates. A model's failure modes are often more informative than its average performance.

Third, comparison baselines should shift from individual humans to teams and workflows. In practice, AI tools augment human work rather than replace individual humans in isolation. Evaluating whether a doctor-plus-AI system outperforms a doctor alone, for instance, is a more meaningful question than whether the AI alone beats the doctor alone.

Finally, there is a growing call for independent, third-party evaluation infrastructure — bodies that develop and administer benchmarks without direct ties to the companies whose products are being assessed. Currently, a significant proportion of benchmark results cited in AI research and marketing are self-reported by the developers themselves, a conflict of interest that independent researchers have repeatedly flagged.

The Regulatory Dimension

The stakes of getting evaluation right extend beyond academic credibility. Policymakers in the European Union, the United Kingdom, and the United States are actively developing frameworks for AI oversight, many of which rely on performance thresholds and capability assessments as triggers for regulatory requirements. If the benchmarks underpinning those thresholds are unreliable, the regulatory scaffolding built on them may provide less protection than intended.

The EU AI Act, for example, classifies AI systems partly on the basis of their capabilities and the domains in which they operate. Establishing those classifications accurately requires evaluation methods that are robust, reproducible, and resistant to gaming — qualities that current benchmarks frequently lack, according to independent researchers.

What This Means

For anyone building with, buying, or regulating AI systems, benchmark scores alone are an insufficient guide to real-world performance — and the field's leading voices are now saying so plainly enough that the pressure for better alternatives is unlikely to subside.