In the second of special series recorded live at the PSE-CEPR Policy Forum 2025, we are asking, how good is AI at doing real-world job task? And how can we measure their capability without resorting to technical benchmarks that may not mean much in the workplace?

Since we all became aware of large language models, LLMs scientists have been attempting to evaluate how good they are at performing expert tasks. The results of those tests can show us whether LLMs  can be useful complements to our work, or even replacements for us, as many fear. But setting or grading a test to decide whether an LLM can do a problem-solving job task, rather than solve an abstract problem, isn't easy to do.  Maria del Rio-Chanona, a computer scientist at UCL, tells Tim Phillips about her innovative work-in-progress, in which she asks an LLM to set a tricky workplace exam, then tells another LLM to take the test – which a third LLM evaluates.