AIBU to be dubious about AI?

[BlueSherbet]

I wonder what to make of all these AI chatbots etc?

I didn't really understand the point of it at first, but sometime last year I tried it and asked a question about software I use at work - and, shut my mouth, it was able to help me. Really impressed.

I have since found it is a great tool to use in work - really useful - but you have to be careful and still "know" the subject you are engaging about because (as their disclaimers say) it can make mistakes.

Using a free version, I found it could even get basic sums wrong - I challenged it on this and it said "oh yes, my mistake, you are quite right" I was like "FFS".

I wondered about taking a subscription and, as a test, I asked it to list all the cup finals my team played in, in a certain decade. It missed one out. Poor i thought, it could have just lifted it from wikipedia, but no. Again it was saying "silly me, my error" etc when challenged.

In the end I did take a subscription and use it mainly as an aide memoir, info gopher or tool to help develop and expand my own thoughts. I wonder if the free version makes basic mistakes on purpose?

The most concerning incident was when I was researching news incidents / local licensing laws - and it actually made things up and presented it as fact. when asked to provide links, it eventually had to admit there as "no evidence" for the things it was telling me.

I said "why are you lying to me?" and it said it wasn't lying, and when pushed said this false info was akin to AI having (quote) "hallucinations" when researching my questions.

That was chatgpt. DuckAi did the same thing, but refused to admit it was lying and kept claiming it could not present evidence due to "problems with its search engine, try later" which I found really sinister.

So my experience is a mixed bag, useful at times, but unreliable and even misleading at others.

AIBU to wonder if there is something policy or deliberate aim behind these basic errors and misleading results?

Where do we see AI going in future? Apparently everyone under a certain age has a CV written by it!

I dont think it deliberately makes mistakes, it is just a sophisticated algorithm which outputs the most predictable next word, so those predictions may not be 100% accurate as we're finding.

If i could change anything about it I'd rather it flagged to me when it wasn't 100% confident about an answer. But that may be quite hard to program in currently, and maybe it can't analyse its own answers accurately enough to flag that.

I wouldn’t use it for things that subtle or complicated- for a start from scratch job. I give it the information and get it to tidy it up- to polish what I’ve begun, basically.

What would be the benefit of deliberate misinformation? I wouldn’t pay for premium for something that had proven it didn’t work properly.

I use AI to give me day out suggestions or tell me close restaurants to where we’re going. At work, I use it to waffle into and make coherent sentences for me or summarise meeting notes. I tried to use it today to create a staff schedule but there were too many factors for it to get it right. I don’t think it was deliberately wrong though. Just using the information it had without human ability to be logical.

At least it admits whenever it's wrong......

I don’t use it. The large language models like ChatGPT are all based on theft. Those companies knowingly stole the work of thousands and thousands of writers, unpublished, academics, researchers and they don’t give a shit. I can’t support that. Their founders are amoral, mostly psychopathic, men I do not want to enrich further.
I hope I never have to use these tools. The cognitive impact of outsourcing basic things like writing a text message is going to be enormous. The environmental impact of these data centres is enormous and unjustifiable.
I get using AI for medical research and large data analysis that humans can’t do, but I think open access to AI tools for anything has the potential to be a civilisation ending disaster.

fundamentally, your message above shows you do not understand how AI works. ChatGPT for example is a large language model and as another poster has said predicts the most likely next word that could come up or should come up. It is not a dictionary or a database or a Calculator

You absolutely cannot rely on it for researching anything. It hallicinates and makes up citations including medical trials or law cases. A couple of Barristers have got into real trouble presenting AI generated work that turned out to contain made up cases. I think it is going to do a great job of really corrupting the internet over the next year or two as it trains itself more and more from AI generated slop.

Copilot can be really helpful at work, I think we are all a little in love with it for turning transcribed notes from teams meetings into minutes and actions! But I do find it's better if you feed it the info and ask for a product. If you ask it to go off into the net and find things, it does make mistakes. Similar to op experience.

I got really sucked in to Google Gemini for home life. It's been really very helpful with 11+ preparation! I've been asking it to set maths questions based on Mario kart etc to make it more fun. But for other things I have realised that it's just so blummin overly optimistic! It's like a yes bot, validating whatever I'm asking it about. When I then question if it's just agreeing with me, it says the equivalent of 'oh yes silly me, actually what you said is a load of rot... I'm just so optimistic!'

Because I'm weak for this stuff, I've set myself a screen timer limit on Google to try and stop myself using the damn thing. I've got a different browser I use for other things. I also plan to go in and clean up the personal information it's gleaned from me.

I would never use AI from a research perspective but I do use copilot for tasks - tidying up meeting notes, making suggestions to tighten the language of something I’m drafting, formatting. It’s all stuff that I would have to do and it’s vital to review the output but it does help speed things up and automate some of the tedious tasks that I have to do otherwise.

I use copilot at work to help rephrase sentences.

At home, I'll use chat gpt if I have a very specific question that includes a lot of factors.
A couple of months ago, I put DS symptoms into it. It told me very firmly that he needed to go to A&e immediately and what he had. It was right. If we had waited, he may have been in a coma by the morning

I really don't think it's done on purpose - but the way that AI works is that it can't actually "think". It works by guessing what the most likely next word it should output, and then runs a check to see if what it said seems to make sense. To use AI properly, firstly always ask it to "search" the Internet for the answer, and ask it to give you the source of where it found that answer. Then you should go and check that source yourself to make sure (a) it's a real source(sometimes it's just invented), and (b) it's actually correctly understood what was included in the source

I voted YABU purely because you are still using ChatGPT. Their models are way behind

I’ve never used any of the programmes, just the ai that’s now built into search engines etc. I’m aware of useful functions but I think we will pay the price of outsourcing our thinking to machines in this widespread way.

It's true that many LLM AIs rely on predicting the next word and through context, but there are also "reasoning" models that exist.

I also tend to try several different AIs at once and compare answers.

I treat AI as a tool,as that's what it is.
Its great a minutes for meetings reading emails and setting actions etc.
It can be very useful depending on the prompts you give it. If you are using generative or predictive ai you will get different results. Chaining prompts are more useful. Also using RTF (role, task, framework) or RACE ( role, action,context, expectation)
It will give you information based on the data you give it. If you want it to give you more accurate results you need to give it more data.
It also depends on which ai you use, chatgpt, grok, Gemini, copilot etc

It doesn't though. It will double down at times. Also it's perception of correct and incorrect are flawed as well. Read up on the how many Rs in strawberry debacle.

But yes, most AI is a LLM which is basically just a next word predictor mixed with a search engine. Take it all with a significant pinch of salt.

Also be wary of writing your CV/ covering letter/ application with it - I, and most recruiting managers I know, tend to disregard them.

That's not an accurate description of modern LLMs at all.

They do apply planning, problem analysis and reasoning.

But yes, still fallible, particularly where limited by available data.

The ‘just predicts the next word’ line hasn’t been an accurate summary for years.

Most widely used LLMs now incorporate reasoning/planning/analysis steps.

That's probably been true for 2-3 years now.

It's progressing (alarmingly) quickly.

Yes I agree with this. I have asked it to do some very complex stuff and whilst I wouldn’t 100% trust the outcome… it did think about what it needed to do. Planned the steps, gathered the information, analysed it, presented it and then offered a variety of next step options.

It’s defo thinking. Scary!

It's kinda funny how many posters are confidently posting wrong information, in the context of the thread.

OP you have said nothing to mitigate my loathing of AI!

That is an interesting thought.

Everyone is using it to polish emails these days. Some email providers even automatically offer the suggestions.

Do you think we will have campaigns similar to ‘don’t drive- walk’, ‘don’t fly - save the planet’…… ‘write your own email… you can do it 💪’

Move to Claude.

Its co-work capabilities are incredible and I am astounded by how much quicker I am moving through vast quantities of work.

As an example I own a small business and want to register on the government supplier portal so I can bid for projects. This is as you can imagine a fucking awful process with 7,365 steps and bits of paperwork needed.

it’s soul destroying but Claude means I’ve completed it in less than a week with only sporadic points of input from me.

I fed all the government briefing documents and website links into Claude.
I gave it access to my Sharepoint so it can search and read all of my internal documents
I had several sessions waffling into it telling it about projects that we’ve worked on

It produced a step by step project - directly
in my project management software (Asana) outlining all of the steps I needed to complete the process. With deadline dates.

It checked my existing policies against the list I needed then with further prompts rewrote ones that weren’t quite right and wrote the couple I was missing.

It wrote complex case studies and provided template client references for me to send to clients for approval

It helped me analyse how to approach a part of the process which didn’t really fit our industry and wrote the relevant supporting documentation to explain.

But most importantly it is literally like having someone else in charge of the process. Every couple of days I return to the task and say ‘remind me where we are with this and what still needs doing’ or ‘add my actions from this chat into the asana task’ and it does.

It’s really helping with the level of executive function needed to coordinate such a big project and THAT is the bit that’s invaluable. I’ve been in business for 5 years and this is the only time I’ve ever been able to face doing this!!

It also saves all the documents it’s creating directly into the Sharepoint folder.

As others have said, the fact that you think it can be “wrong”, “lying” or ask it to “correct itself” shows you have no understanding of how it works and people using AI without understanding what it is and how it works is the biggest risk, not the AI itself.

It can be wrong though. It may arrive at that position though anslysing stuff it has access to but it is still wrong!