Quite often — it’s all about how you ask. I’ve gotten into the habit of always asking at the end to list the sources and links it’s drawing from and the method it used for the calculations.

The assignment was: 10 companies that were added to the S&P 500 and their performance one year before and two years after.

In the end, after several follow-up questions, it gave me only two companies, saying it couldn't find precise information for the others.

I did it using ChatGPT.

It looks up some things really well, but sometimes it's a bit stuck and falters, and it tends to steer you in another direction, like it doesn't want to admit it couldn't find your request.

Otherwise, I often use it for a small analysis, to find specific details, and sometimes for an opportunistic take — after all, you don't understand everything, and it's great for clarifying the pitfalls of certain businesses.

Lately I’ve been using it quite often, and it usually helps me understand the business. Otherwise, I look up the financials and results myself.

Something like you — I often use AI to do initial research, for example comparing several companies in a sector or mapping the main players, etc. Then I often paste transcripts/articles there and ask follow-up questions about details or ask for summaries of those articles; sometimes a simple piece of information is scattered across two different pages.

If LLM models didn't have problems with hallucinations, I'd gladly use them and trust them more, but for now you need to verify and check they aren't spouting nonsense (this applies mainly when they're working with data they've found themselves, less so when you give them a source). But I have to say that even over months I've seen major improvements, and it probably won't be long before the error rate is minimal.