What's this about?
Mollick reports that Prinz has been running legal benchmarks against AI systems: the latest models are very good, and the cheaper low-reasoning models are not. His question: how do you make sure a vendor selling you an AI solution is incentivised to serve the good — and therefore expensive — model?
Why is this a structural problem rather than a contractual one?
Because the vendor's margin is the difference between what you pay and what inference costs, so every quality downgrade is pure profit and is invisible to you. You cannot tell from the output which model produced it, model routing is a legitimate engineering practice, and "we optimise model selection per query" is both true and unfalsifiable. Unlike most quality questions, there is no artifact to inspect — the same interface returns a worse answer and nothing indicates why.
Why does this matter to you?
- It is the exact mechanism that produced the lawyer's memo in the previous entry.
- The only real defences are contractual model specification, or your own benchmark run periodically against the vendor.
- This will be the dominant AI procurement question within a year and almost nobody is asking it yet.