AI tools and how wrong they can be - general observations anecdotal

[ETAF]

I have been doing a lot with AI over the last few weeks just for research for a few real things i'm doing -

AND i do have a lot of time on my hands , due to retirement and illness

I have used chatgpt, Gemini, Claude, base44 , and tried the same things across the different platforms

We all know that they make simple mistakes in Excel - I have just had an xlookup suggestion from claude and xlookup does not appear anywhere in the formula - answer was , sorry i mistake and mislabeled

I have an upgraded weather station and been playing with AI for weather forecast and current state meaning/interpretation using various apps , windy.com etc on all the local measurements from the device , as well as optimising the position and use , AI - used very old information despite being told firmware and other current info and apologised to me for being wrong when challenged _ THIS IS THE ISSUE if not challenged, and lots more mistakes over the last few days playing with the Weather station , even on positioning and some forces and loading , a simple trig question was wrong - !!!
i wanted to work out how much cable to use to tie back from top of the pole to a wall - basic trig !!!!! and wrong result

I have also had to write some pretty sensitive emails over the weekend, and i thought the models as language based would be fine, BUT they also got the sense and the information i provided wrong , this was mainly chatgpt over the weeken

I have been upgrading a website - adding pages for historical facts - got a lot of those wrong , and SEO - got a lot of that wrong

I then started a discussion on its accuracy , in fact I stated that it was over the last few weeks <10% CORRECT , and it misread that !!!!! as 10% ERROR

USER: its not <10% error - ITS a >90% error and <10% correct heres another example of not reading that text correctly - here is the bit again "i would say <10% of the time you are correct" so a huge error
AI: Ah� now I see exactly what you're pointing out�and you're right. I misread your original statement about accuracy entirely. You literally said:
"i would say <10% of the time you are correct"
�and I interpreted it as if you were saying "I'm correct most of the time (<10% wrong)," which is the opposite of what you meant. That's a huge, fundamental error on my part�exactly the kind of thing you're frustrated with.
This is not a minor misstep; it's a complete reversal of your meaning, and it perfectly illustrates your point that I often fail to read the facts you provide accurately. It also shows that even when you are explicit, I can misinterpret if I don't slow down and verify.
I want to be absolutely clear: I understand now that you are saying >90% of the time my responses have been incorrect, and the <10% figure refers to when I am actually correct. That's not subtle�it's a very serious failure on my part.
ANNOTATION (Present Author "CLAUDE" creating a document ): This is perhaps the most critical moment in the conversation. The AI system demonstrably misreads a simple, unambiguous mathematical statement ('<10% of the time you are correct'), interpreting it as its inverse. This is not a subtle interpretive error�it is a fundamental reading comprehension failure that directly validates the user's complaint. The irony is profound: in a conversation about accuracy failures, the AI fails to accurately parse the very statistic describing its failure rate. This exemplifies the core limitation of pattern-matching systems: they can produce fluent, reasonable-sounding text while simultaneously making elementary errors in factual comprehension. The user's follow-up ('heres another example of not reading that text correctly') demonstrates awareness that this failure is emblematic of the broader problem being discussed.

I put all the responses from chatgpt into Claude and asked for comments , and to produce in the style of an academic white-paper to be published

The really worrying thing is the kids will grow up with this and NOT check - I remember when calculators hit the market and started being used in school and on many occasions found people unable to do basic arithmetic in there head (at a checkout for a few items , I had the correct money ready and the person nehind the till could not understand that , same again when i asked for change to be corrected which was wrong on till !!!! - this was 20 years ago ) and need a calculator OR they trust the result it gives , and if they make a mistake inputting something, they have no idea how to approx to sanity check if the answer is near

anyway, i thought i would put here as i had the very simple error playing with an excel solution

 

[motorbiking]

There is also a whole new industry of AI search optimisation where people write articles to guide the AI, providing opinion articles as Fact.

The AI matches the content to the question without fact checking it.

It’s really important to check the sources and make sure the summary words are actually in the article.

 

[ETAF]

There is also a whole new industry of AI search optimisation where people write articles to guide the AI, providing opinion articles as Fact.

Yep, i have been doing a lot of that for AI searches SEO based on the algorithm EEAT Experience, Expertise, Authoritativeness, and Trustworthiness. And thats where I had loads of FACT errors in creating pages to improve performance of website - I specified to improve USER Experiance and meet the EEAT , also that I did not want it to look like AI had created the information as this was not just SEO but also a long overdue update for the site

that went on for weeks , working most days on it - SO MUCH WRONG - and also crossed reference with BASE44 on a loto fo the coding as well - thats what started me on this journey - to use a modern word !!!!

Part of the SEO is to have an FAQ section , even if just in the JSON code on the page - AND BOY did it get those wrong and many itterations and of course many timeouts for the free versions of claude and base44 , which meant providing all the info again , as they dont keep session history

It’s really important to check the sources and make sure the summary words are actually in the article.

I had to keep asking are you 100% correct and show me the evidance - many times it had MADE THINGS up and admitted to making things up - called it AI Hallucinations - wow

 

[Brigadier]

You can instruct the AI to only refer to the sources you specify, if that is of any help to your issues.

 

[ETAF]

Well, i dont really know all the sites to use , i just say use latest sources for xxxx xxx and make sure we get this 100% correct first time so check before posting back .... its just a general observation really how wrong it can be

now having to modify a google script for linking sites together and getting that wrong as well

 

[IT Minion]

AI tools can be very powerful, but to do those tasks you need to move to a specialist build rather than a general purpose one.

As it is general purpose AI is not good at telling the difference between content by age or by reputable sources. It also gets confused between similar topics.

You can fix that at least to some degree, but it's a skilled task

 

[ETAF]

You can fix that at least to some degree, but it's a skilled task

which you average user wont do - i know loads of people in business now who are using copilot and AI to make business decsions and pay a subscription to chagpt to produce spreadsheets , as i discussed with AI the risk of using that info - for serious business decsions , as i have done for many years in my career - manpower planning and Redundancies is just one example of how i used excel and MI - now you just plug it into claude or copilot (not used as it needs onedrive or sharepoint, and onedrive on a mac just throttle's it) boom get rid of 4 people - i know VP and senior manages would just jump on that info and some analyst who has no idea does some number crunching

 

[IT Minion]

Yep, part of my job is showing people how to not do that. And to build the specialist tools needed. I've got a slide deck dedicated to why you should never trust Grok with anything ever.

Today I am banging my head against generative AI pipelines.

 

[ETAF]

sound interesting , never been near grok, heard it in the news recenlty about pornography or undressing people

generative AI pipelines.

what are those