During the World Series, Google ran an ad for Gemini that noted several correlations between home runs and full moons. Marcos Huerta analyzed data on full moons and home runs to debunk each statement.
I had assumed that some Google marketing types asked some MLB data team members8 for some lunar factoids and some poor analyst when probably knew better sent them whatever they could find: error bars be damned.
But maybe they really did use Gemini. The ad included 3 stats that are all a little true – if you ignore the error bars, and the home run stat in particular has some enormous error bars. That feels like what might happen if you ask an LLM to do data analysis without babysitting it carefully. Agentic p-hacking? Yet, not really p-hacking since it still couldn’t find a statistically significant result.
Read More
