Learn more about the host Laurence Gill at www.laurencegill.com.

In Episode 12, Laurence Gill takes that number apart. A Stanford research team called BetterBench built a 46-point audit covering benchmark design, reproducibility, and documentation, then scored 24 widely-cited tests against it. MMLU came in at 5.5. GPQA, a far less publicized test, scored double that. The reasons are specific: ambiguous question phrasing that swings scores when a comma moves, a reproducibility gap across most published benchmarks, and a quiet contamination problem where models may have already seen the answer key buried somewhere in their training data.

Laurence walks through how these tests actually work, why Goodhart’s Law explains the industry’s race to game them, and how newer benchmarks like GPQA and ARC-AGI are trying to close the gap. It closes with five questions to run through before any benchmark score is allowed to inform a real decision and one open question about what happens when AI starts writing the tests that grade other AI.

Over 90 percent of U.S. employers now use AI to screen job applicants. And over 60 percent of the Fortune 100 runs that screening through the same vendor model. A 2026 study from Stanford, Chapman, and Northeastern Universities — the largest independent research ever conducted on deployed hiring algorithms, reveals what that concentration is actually doing to real people at scale.

Researchers analyzed 3.4 million applicants submitting 4 million applications across 156 employers and 11 market sectors. What they found is not a conventional bias problem. It's an architectural one. More than a quarter of all applications submitted by Black applicants landed in positions where the algorithm was actively producing adverse impact. 29,000 additional Asian applications would have moved forward in a fair system. And to statistically guarantee one interview, candidates in an algorithmic monoculture now need to submit 25 applications, two and a half times the number required in a human-driven system.

This episode breaks down how algorithmic monoculture works, why prior vendor studies missed the discrimination, what the disaggregated data reveals, and what a governance framework capable of addressing it actually looks like. Essential listening for every leader whose organization relies on AI in hiring or whose team members are navigating this market right now.

Every AI tool your team uses today runs on infrastructure most leaders have never been taught to think about. It’s called an API — and once you understand what it is, your entire mental model of what your team can actually do with AI right now is going to shift.

In this episode, Laurence breaks down the mechanism that connects your organization to world-class AI — no technical background required. You’ll learn what an API actually is, why the “menu contract” framing is the one that matters for decision-makers, and how a small team with the right knowledge can now access the same AI models powering enterprise products without a data science department or a six-figure infrastructure budget.

This episode covers:

— What an API is and why stability in that contract is everything

— The real reason your team can access world-class AI today — and what that means for what’s possible right now

— How to think about the major AI API providers — OpenAI, Anthropic, IBM Watson, Google Cloud, and SiliconFlow — and the decision logic for matching the right tool to your specific constraints

— What Hyrum’s Law is, why it applies directly to AI, and the governance question every leadership team needs to answer before building workflows on top of an AI API

If you have approved an AI tool for your team without understanding what’s running underneath it — this is the episode.

AI Literacy for Leaders is a podcast for executives, directors, and managers navigating real AI decisions without a technical background. New episodes weekly.