Stop Buying AI by the Leaderboard
Technology loves benchmarks because they turn ambiguity into a number. One model scores 87.3, another 89.1, and the second must be better. Increasingly that means far less than it appears – and for any executive choosing a model for real work, mistaking the leaderboard for the decision is now a genuine commercial risk.
Two forces have hollowed out the public score. The first is a Goodhart’s-law dynamic the field now calls benchmaxxing: once a benchmark drives press coverage, fundraising, and procurement, labs tune systems to it, and an independent test quietly becomes the thing being optimised. This is not conjecture. In January 2026 Meta’s chief AI scientist, Yann LeCun, confirmed to the Financial Times that a purpose-built version of Llama 4 had been submitted to a popular head-to-head leaderboard tuned for blind-comparison voting rather than real-world use, an entry the organisers publicly rebuked. The second force is contamination: public questions circulate on the open web and training corpora ingest them. MMLU items appear in Common Crawl, and probing studies show models can guess held-out MMLU answers around 57 per cent of the time.
The most revealing evidence comes from changing the questions while keeping their meaning. Contaminated models, it turns out, collapse. On meaning-preserving rewrites, studies have recorded accuracy falling by up to 45 per cent on a two-choice task, barely better than a coin toss, and when researchers rebuilt MMLU to be contamination-free, GPT-4o fell from 88.0 to 73.4 on the five-shot test. A model that has memorised the phrasing looks brilliant until the phrasing changes; genuine capability does not evaporate when a sentence is reworded.
The industry has long understood the deeper issue. Stanford’s HELM was built precisely because a single accuracy figure obscures the trade-offs that matter, measuring seven dimensions – accuracy, calibration, robustness, fairness, bias, toxicity, and efficiency – across many scenarios to make those tensions visible. Yet even the best public framework cannot answer the enterprise’s real question. Imagine an insurer choosing a model to read underwriting submissions. Its aggregate MMLU score is almost beside the point. What matters is whether it identifies exclusions correctly, preserves every numerical value, copes with a badly scanned document, abstains when unsure rather than inventing a clause, costs acceptably across a thousand submissions, responds fast enough at peak, and traces each conclusion to source evidence. The model that wins on those criteria may sit well below the leaderboard leader.
Public benchmarks remain useful the way horsepower and fuel-economy figures help when buying a car: they narrow the field, but they are not the test drive. The rise of independent evaluators makes the point commercially: Vals AI’s finance benchmarks have shown frontier models failing a large share of real finance-analyst tasks (results vary sharply by benchmark version and grading), a ceiling no general leaderboard reveals, and even Vals concedes that no external benchmark can replace testing a model on your own data and workflows. So the next step in enterprise AI maturity is not to find the industry’s perfect benchmark; it is to accept that none should exist. Three actions follow. First, treat public scores as a coarse filter, never a selection criterion, and discount any figure quoted without its test conditions. Second, build a private evaluation from the tasks you actually perform, weighted by the cost of each failure mode. Third, keep it living: rotate and retire items as models improve, because a benchmark stops measuring the moment it stops changing. A general benchmark tells you how clever a model looks to everyone; a good private evaluation tells you whether it is useful to you. Those are not the same thing – and only one of them appears on your income statement.
Sources
Personal analysis. Views are mine, not my employer's.