Listen on
Transcript
Old tests can become too easy. In 2026, AI evaluation is shifting towards real tasks, errors and judgement. Better testing asks what happens outside the exam.
Full transcript of Briefing 05, 0:10, published 5 August 2026.
Key points
- Benchmarks lose meaning once systems are optimised against them
- Real-task evaluation measures work, not recall
- How a system fails matters as much as how often
- The useful question is performance outside the test set
Why old benchmarks stop working
A benchmark is useful while it is hard. Once a class of systems has been trained, tuned and selected against the same fixed set of questions, a high score stops being evidence of capability and starts being evidence of familiarity. The test has not become easier; it has become known.
This is not a failure of anyone's integrity. It is what happens to any measure that becomes a target, and it is the reason evaluation has to keep moving.
What real-task evaluation asks instead
Newer evaluation tends to look less like an examination and more like a trial run: multi-step tasks with genuine ambiguity, incomplete information and no single marked answer. The interesting output is not a percentage but a description of behaviour — where the system hesitated, what it assumed, how it failed.
For anyone deciding whether to rely on a system, that behavioural picture is far more informative than a leaderboard position. It tells you what the failure looks like, which determines whether it can be caught.
Frequently asked questions
Why are AI benchmarks becoming less useful?
Because systems are optimised against well-known test sets. A high score then reflects familiarity with the test rather than general capability in the task it was meant to represent.
What is real-task evaluation?
Testing a system on realistic, multi-step work with ambiguity and incomplete information, and examining how it behaves and fails rather than only whether it scores well.
What should a buyer ask about evaluation?
Not the headline score, but what the failures looked like, whether the test material could have been seen during training, and how performance held up on tasks resembling their own.