Scikit-Learn Co-Founder Calls For Statistical Harness In AI Data Science

Can AI become a real data scientist?

I spoke with @GaelVaroquaux , co-founder of @scikit_learn and Probabl, about why AI systems need more than code generation.

They need statistical judgment: uncertainty, leakage detection, evaluation, bias checks, and methodological discipline.

An AI agent can know the full scikit-learn API and still do bad data science.

That is why Gaël describes the need for a “statistical harness”: a way to connect AI systems to tools while keeping the analysis trustworthy.

We also discussed Skore, human-in-the-loop science, why real-world data is harder than mathematics, and what creativity means for humans and machines.

New episode of Creative Difference.