AI Mode Is Not Good At Math
Paid LLMs have gotten quite good at computation, but AI Mode isn't designed to do that
In general, I think AI Mode is a good starting point for a lot of questions. It’s not as precise, knowledgeable, or nuanced as a lot of paid tools, but it has the advantage that it is fast, easily integrated with web activities, and free.
At the same time, there’s a bunch of stuff it’s just not designed to do, and knowing its limitations is crucial. Currently one of those limitations is math. Contrary to popular belief, the “reasoning” LLMs you get with paid accounts have gotten really good at math. Much better than you probably are! But that sort of computation is expensive in an LLM and not available (currently, October 2025) in a lot of the quick and cheap versions, of which AI Mode is one.
This video walks through an example you can use to show students that.
Incidentally, YouTube now has an ask a question mode for YouTube videos, and I tried it out on this one, asking what I meant by AI Mode is “bad at math”. It did really well! Answer below, and maybe I’ll use these feature for these posts more regularly. 1
The video explains that AI mode is generally “bad at math” because it doesn’t use external tools for precise calculations, unlike more advanced or “expensive” large language models (LLMs) (7:12, 7:21).
Instead of calculating, AI mode tends to:
Be hedgy: It gives vague or approximate answers rather than precise dates or numbers (3:26, 7:14).
Provide links: If it’s not confident it can give a high-quality, precise answer, it often defaults to providing external links to websites that might have the information (4:29, 6:57).
Show code examples: If you push it to “use tools” (like JavaScript or Python), it might explain how you can use those tools for calculation, rather than using them itself (6:23).
Essentially, AI mode is not designed for the kind of precise, tool-based mathematical reasoning that other dedicated tools or more advanced LLMs can perform (7:34, 7:52).
Also, all my videos of this sort are in this playlist.
And at least for the moment, most of what I am focusing on in these fits into this framework:
Thanks!
For the main body of the post you should know I don’t tend to use AI because twenty years as a blogger has made it possible for me to type out a blog post in about as much time as it takes to watch a thinking model generate it. You think I’m joking, but there are times I do feel like my blogging is sort of LLM-based because I am just watching my fingers type words with no idea what is going to come out, until I hit the end of the post. Like this. :)