Chat GPT is usually wrong on financial matters

Fortune said:
I think AI is fine for this type of financial 'coaching', but it's not to be trusted where there's an analysis of numbers required.
Click to expand...
I disagree based on extensive recent experience using ChatGPT to analyse some household budget and medium/long term early retirement scenarios. As ever I did double check things and correct it if it got something wrong but, by and large, it was useful and accurate.
 
Dr Strangelove said:
I’d say that’s pretty specific
Click to expand...
I'd disagree. It's the question that is specific.

Most of the advice there relatively objective, and doesn't really analyse the numbers. A quick glance at the numbers and I can see some errors (the tax relief jumps out). They don't materially affect the advice given (as it's not based on the numbers) but are wrong all the same. I note it also states that rental property investments are liquid.
 
ClubMan said:
I disagree based on extensive recent experience
Click to expand...
Let's look at the example posted.


"Eliminating a €200k mortgage at 3.5% would save approximately €7,000 annually in interest, totaling €140,000 over 20 years."

Without even thinking hard about it at all, the €7k figure could only be correct for an interest only mortgage. Some quick calculations give a saving of c€78k, not €140k.

"This aligns with the “debt avalanche” method, targeting high-interest obligations first."
Hardly true of a mortgage.

"Tax Relief: Contributing €30k to a pension would yield €12,000 in tax savings (40% relief on €50k income), effectively costing €18,000 net."

Standard rate cut off is €44k, meaning not all of this contribution will get tax relief at 40%. No mention of backdating to the precious year.

"With a €100k pension at age 40, the user is below the recommended benchmark (1x annual salary by 30, 3x by 40)."
This is just rubbish
 
Fibbernacci said:
Did you sense any pattern in the type of questions it got wrong or right?
Click to expand...
I didn't see any particular pattern. Numbers were sometimes right, sometimes wrong. It was almost as if it was someone not great with numbers doing the calculations (I suppose that's likely the case in the days it was trained on). Sometimes it hallucinated regulations that didn't exist, or just got them backwards (something allowed wasn't allowed, etc).
 
I did a test myself there with a specific question related to my industry and it gave a wrong answer. When I enquired why, it said that it doesn't parse text every time and uses 'generative' mode to make predictions. I then asked if I had specifically worded my question to force it to parse the text, would it have done that and it said it would have switched to 'parsing' mode. If that's true I wonder if the user could possibly increase accuracy of the responses by wording the questions in a particular way.

(but still not take financial advice from it lol)
 
Fibbernacci said:
If that's true I wonder if the user could possibly increase accuracy of the responses by wording the questions in a particular way.
Click to expand...
I’ve been using LLMs since late 2022.

You can materially improve the quality of the replies by changing how you do with the prompts. All I can advise is for people to play around a little bit with it. The more precise your prompt often the better quality the response. But if you badger an LLM it will hallucinate a false response in many cases.
 
Dr Strangelove said:
if you badger an LLM it will hallucinate a false response in many cases.
Click to expand...

That's quite the statement...

Been trying to get copilot to return documents I know are there. But it's like trying to get a cat to find a book in a library. I can't help but think it's easier to find it myself.

Going to try some coding stuff at some point.
 
Last edited:
It's pretty well known from the research that ChatGPT is quite willing to just make stuff up if it doesn't 'know' the answer to a question, or part of a question (ie if it hasn't been trained on the relevant information). It actually 'likes' to be corrected as this new knowledge (whether correct or not) is then fed into the database.

The problem is the amount of wrong or ambiguous information that has been used for the training. It's easy enough to spot this if you ask it questions about any area you are expert in already yourself. Or ask it questions about a very narrow field and see how it generates false information, or just throws in partially correct information (e.g. a person with the same name; a similarly named scheme from a different jurisdiction) as fact.

I find it completely useless as presently constituted. It will undoubtedly get better over time, as long as the big ethical questions around copyright etc can be sorted out.
 
ChatGPTs own 'opinion'...

Q: Is ChatGPT reliable for financial matters?

"ChatGPT can provide general information and insights related to financial matters, such as investment principles, budgeting tips, or explanations of financial concepts. However, it is not a substitute for professional financial advice. For personalized financial guidance, it's best to consult a certified financial advisor or expert who can consider your specific circumstances. Always verify information from multiple credible sources before making financial decisions."
 
A better test would be to give 5 financial advisors a few questions to answer and see how many of them are right. I would imagine that there would be great variety in the answers.

Then give the questions to Chat GPT and see if any of them are actually wrong.
 
I have occasionally asked all the big AI models questions about Irish deposit products. Often asking just out of curiosity as to what the quality of the response would be like. And the quality of responses is notably improving lately especially in Gemini. Very impressed with some of the replies.

Meanwhile, Gemini quoted askaboutmoney (as well as other sources) when I asked a non-deposit financial question to it. Interesting that Askaboutmoney posts are being used to feed the Gemini model.
 
Lightning said:
Interesting that Askaboutmoney posts are being used to feed the Gemini model.
Click to expand...
Isn't it the case that most or all such publicly accessible AIs/LLMs are being trained on the complete visible (i.e. non deep/dark) web contents? Which is part of the reason that some of the results can be of variable/questionable quality - because there's a lot of rubbish on the web. E.g. the "AI recommends that people eat at least one small rock daily" example that came about because of a parodic article on The Onion...
www.bbc.com

Google AI search tells users to glue pizza and eat rocks

Google has defended the answers given by AI Overview, describing them as "isolated examples".
www.bbc.com www.bbc.com
 
ClubMan said:
Isn't it the case that most or all such publicly accessible AIs/LLMs are being trained on the complete visible (i.e. non deep/dark) web contents?
Click to expand...

Yeah, pretty much. LLMs are trained on massive datasets derived from publicly available online sources like websites, books, and code repositories. Which includes a huge amount of websites but not all websites. Reddit, the FT and some other large websites get paid for inclusion in the LLM datasets.
 
Back
Top