ChatGPT can’t count, and that’s a problem

How an ultra-useful research tool is being held back

Ariel Lubonja
Jun 5, 2023

Many scientific studies consist of counting the occurrence of an event, phenomenon or occurrence. With all the training data it has seen, ChatGPT or GPT-4 would make an excellent source.

Unfortunately, it seems the model is not allowed to answer those questions. Google’s Bard can do it though.

My intuition is OpenAI is trying to prevent the model from revealing their sources of data, but that is just speculation.

Even GPT-4 refuses to do it

Funnily enough, if you ask the model if it can count, of course it can!

Just not stuff from its training data



Ariel Lubonja

I am a PhD student in Computer Science at Johns Hopkins University. Area: High Performance Computing, Graph Machine Learning