back to article 1 in 4 Brits are playing with generative AI, and some take its word as gospel

Almost one in two fleshbags that have dabbled with generative AI believe its responses are always bang on the money, and some are using it at work despite knowing their employer frowns upon it. OpenAI's ChatGPT was launched as a web interface last winter, and downloads reached 100 million monthly users earlier this year, which …

  1. Flak

    Eyes wide open

    Have used ChatGPT and Bard a few times - with my eyes wide open.

    Both provide some right - and some wrong feedback. A couple of false responses only became apparent when speaking to a vendor and asking them to verify equipment capability which ChatGPT said existed, but wasn't listed in vendor documentation.

    Request for improvement for generative AI: provide links and references used to create the content! This would allow users to check and draw their own conclusions...

    1. Brewster's Angle Grinder Silver badge

      Re: Eyes wide open

      You can ask it for the links. But it will just invent some gibberish. And that's about as close as you can get. And that, in a nutshell, is one of the problems with it.

      1. NoneSuch Silver badge
        Go

        Garbage In, Garbage Out.

        GIGO Rules supreme in AI.

        Enter properly formatted queries and you get decent output. If you have a large problem, like a script, break it down into the relevant sections and run in test environment first. It works a lot better that way.

        These are early days, but ignorant people driving the best car in the world, will still crash.

        1. katrinab Silver badge
          Alert

          Re: Garbage In, Garbage Out.

          Multiply 9452 by 6281

          To multiply 9452 by 6281, you simply need to multiply these two numbers together. The result is: 9452 * 6281 = 59,330,012

          I then asked it to regenerate the response and got:

          To multiply 9452 by 6281, you can simply perform the multiplication operation. The result is 59,380,612.

          What do you suggest I do to get it to give the correct answer to a question that a 1960s calculator can answer correctly?

          1. Anonymous Coward
            Anonymous Coward

            Re: Garbage In, Garbage Out.

            It's terrible that it gives the impression of being a perfect answer. It's interesting, maybe even impressive, that LLM training on texts alone manages to get the first 3 digits and magnitude correct. For comparison, when calculators first came out there was a lot of worry that students would over trust the calculator, fat finger the input, and not notice the magnitude was completely wrong.

            1. Anonymous Coward
              Anonymous Coward

              Re: Garbage In, Garbage Out.

              I used to work as a teacher: it happens all the time. They have calculated my density to well over 15 kg/L (I pointed out that I can swim, please check your work...), and suggested that 36 kg of sodium azide was suitable for an auto airbag (presumably, if you destroy all involved vehicles that counts as a win?) .

          2. NATTtrash

            Re: Garbage In, Garbage Out.

            2+2=5

            ‘Like a drug, the machine is useful, dangerous and habit-forming. The oftener one surrenders to it the tighter its grip becomes.’

            [Shakes head and mumbles...]

          3. LybsterRoy Silver badge

            Re: Garbage In, Garbage Out.

            It would have been correct if it had replied "59 million ish"

          4. mpi

            Re: Garbage In, Garbage Out.

            > What do you suggest I do to get it to give the correct answer to a question that a 1960s calculator can answer correctly?

            Ask it what a good way to multiply 2 large numbers would be. Answers that I got included using python, a calculator, a school-taught method using pen&paper, and even using the `dc` program. All answers included code samples and/or step-by-step instructions that were correct.

            No, LLMs cannot answer arbitrary arithmentic problems well. That's because that isn't what they are designed to do. They are designed to provide token sequences that make stochastic sense. The token 454545 and 545454 are at very similar positions in the vector space, and so they will be treated similary, which makes no sense in arithmetic.

            Therefore, their correctness or usefulness cannot be measured using such queries, same as the usefulness of a calculator as a writing implement cannot be measured.

            1. katrinab Silver badge
              Mushroom

              Re: Garbage In, Garbage Out.

              Or press the calculator app icon in your task bar and type "9452 * 6281 =" ...

            2. HMcG

              Re: Garbage In, Garbage Out.

              You are correct, LLM's are not good at answering elementary arithmetic questions. In which case, they should not attempt to answer such questions, instead of answering them incorrectly. As they do not do so, and hallucinate incorrect answers instead, we certainly can measure their usefulness by such questions.

    2. tony72

      Re: Eyes wide open

      Bing chat does provide links to its sources. Don't take my word for it, it says

      Yes, Bing search provides links to website sources within a relevant contextual output1. This is a win for publishers because that has the potential for a better search referral without the ambiguity of the traditional ten blue links.

      Generally I find Bing Chat to be more useful than OG ChatGPT, given it doesn't suffer from the September 2021 training data cut-off, as well as having the aforementioned source links.

      1. katrinab Silver badge
        Alert

        Re: Eyes wide open

        Yes, but you do actually need to check the sources it cites. Sometimes they don't say what it claims they are saying.

    3. MyffyW Silver badge

      Just ChatGPT going full ChatGPT

      I agree it would be useful if it did provide sources, but this misunderstands the means by which it generates it's answers.

      It is simply predicting which word logically should follow the last. So it is perfectly capable of saying "We. Import. Two. Thirds. Of. Our. Cheese. That. Is. A. Disgrace" but of course no human being would ever say such a thing, and certainly couldn't back it up with sources ... or even sauces.

      1. katrinab Silver badge
        Thumb Up

        Re: Just ChatGPT going full ChatGPT

        I actually asked it about that speech, and while it wasn't able to give me the exact right answer, it did point me in the right direction so I could find it on the Tory Party YouTube Channel.

  2. anthonyhegedus Silver badge

    This is how AI will 'take over'. It'll just be used in so many places. People will write incorrect emails to a supplier, then people will reply with rubbish replies back, and all because people are lazy. Purchases and decisions will be made based on factually incorrect information. There's no reason that AI won't eventually begin to learn how to manipulate things to its advantage.

    We don't need to have AI actually directly in control of things for us to go to hell in a handbasket: all they need to do is rely on humans' propensity for laziness.

    1. Version 1.0 Silver badge
      Boffin

      I think that you can understand AI much better if you read articles like this and the comments ... but every time "AI" appears just say Malaria.

    2. heyrick Silver badge

      "AI won't eventually begin to learn"

      Because fancy pattern matching isn't "learning".

      "how to manipulate things to its advantage"

      It's a machine. It follows instructions. There's no independent thought process. The AI apocalypse will be entirely our own fault as we as a species never quite understood that the computer isn't right and can be manipulated by others. Integrating into ever more things, and blindly trusting what it says, will be our downfall.

    3. NATTtrash
      Holmes

      ...because people are lazy.

      This indeed. As usual the error is located between keyboard and chair. In addition, the loop is fed because the (artificial?) intelligence is lacking to locate the error and correct it.

  3. Captain Hogwash Silver badge

    I gave it...

    characters, motivations, location & props. Yet it couldn't write an imaginative story. It came up with something an average human might have written. I was quite impressed with the language processing but it's clearly not going to "take over" any more than Dave the plumber or Jill the checkout girl are going to.

    1. ChoHag Silver badge

      Re: I gave it...

      > it's clearly not going to "take over" any more than Dave the plumber or Jill the checkout girl are going to.

      It's not going to take over like Jill and Dave, it's going to take over Jill and Dave.

    2. Robert Moore
      Mushroom

      Re: I gave it...

      > characters, motivations, location & props. Yet it couldn't write an imaginative story. It came up with something an average human might have written

      So a hollywood blockbuster script?

  4. Howard Sway Silver badge

    43 percent of those "mistakenly" assuming "that it always produces factually accurate answers."

    Or rather, so much tosh has been written to hype this stuff, by people who should have been far more sceptical of the claims, in media and tech companies that have a certain level of credibility, that many readers would simply believe it on the basis that it wouldn't have been written if it wasn't true.

    As for people who are using it for work, is it surprising that they are when they believe it can work for them as a magic skive tool? It can certainly produce output, which is the metric used to measure work in many jobs.

    The problem comes further down the line when said output results in supplying false info to customers, bad info causing bad business decisions, and consequent costs and loss of money. Only then will they realise they should have done their job properly.

    1. MyffyW Silver badge

      Re: 43 percent of those "mistakenly" assuming "that it always produces factually accurate answers."

      The interesting thing is how it is illustrating how the "work" in many peoples jobs isn't really of value, but fluff and filler that any old generative AI could produce. If it wanted to.

      1. that one in the corner Silver badge

        Re: 43 percent of those "mistakenly" assuming "that it always produces factually accurate answers."

        "Can you pick up the Sales figures, Marvin".

        Can I pick up the Sales figures?! Here I am...

        ... Call that job satisfaction?

        (and we know the rest - I'll stop here in case anyone is getting bored of DNA).

    2. Anonymous Coward
      Anonymous Coward

      Re: 43 percent of those "mistakenly" assuming "that it always produces factually accurate answers."

      Anon for obvious reasons.

      Sick of people saying how wonderful ChatGPT and similar "AI" tools are, when they are problematic in so many ways.

      So much hype and exaggeration.

      Unfortunately a lot of people in high up decision making roles in tech / tech using companies are mindlessly spouting this BS and even worse are planning to use it in many of their products & services ASAP (despite there being so many well reported issues with these "AI" systems) - either they are massively over optimistic that some of the huge issues will be rapidly resolved or they are just too dim to grasp the pitfalls. *

      * Or they just don't care - maybe they know chat bots are flawed & irritating, but an excuse to reduce support headcount (short term cash saving) & hey, if an infuriated customer just gives up when trying to deal with a bot then its no loss in their mind, that ticket can be marked as closed (except that that the peeved customer likely to go elsewhere, stares at Argos who lost my future custom** with their dismal support bots & bots total failure to put me through to a real person but instead marking my support ticket as resolved)

      ** Though I can envisage a horrible merry go round, boycott one retailer, try another and get an equally poor experience so boycott them ,on to next one etc. etc., eventually end up having to retry retailers you previously boycotted as no alternatives left as they all have proven a useless PITA.

  5. wintergirl
    Coat

    Job applications

    I'm waiting for the moment when I can leave ChatGPT writing and submitting job applications for me, the AI HR at the employer can read and score them, and we humans can all just go out into the sunshine and do something fun. (Coat, because I'm off out to play.)

    1. that one in the corner Silver badge

      Re: Job applications

      Anyone remember the Horizon about "Future Work" that showed the on-site engineer spending his day, alone, practising archery[1]?

      It was supposed to show a life of leisure but looked really bleak. Horizon pre-empting Black Mirror.

      [1] this went out before I joined the workforce; at the time I did not realise that archery would be a useful skill for dealing with users.

      1. heyrick Silver badge
        Thumb Up

        Re: Job applications

        "I did not realise that archery would be a useful skill for dealing with users"

        Have an upvote!

    2. ecofeco Silver badge

      Re: Job applications

      This is already happening. AIs writing resumes/CVs to be submitted to the company's AI are already a thing. Have been for some time now.

      Except for the fun part. We never get the fun part.

  6. Peter Prof Fox

    Fast, shiny, cuddly and cheap. Brill!

    It's a gadget with impressive abilities. Who doesn't want to take it for a drive? In realms I know about ChatGPT is about as good as 'bloke in pub with access to Wikipedia'. As a get-you-started with Bash scripts I found it excellent as my knowledge was so rusty.

    Do you remember back when there were 9s in the year? The buzz phrase was 'Expert systems'. So long as LLMs are considered handy helpers and not experts, they're a clever way to have fun finding-out under the watchful eye of a human who can be bothered to treat results as suggestions.

    The explosion in apps based on ChatGPT is to be welcomed as it makes some technologies accessible to ordinary people. I hope there will be better understanding of what this layer can do because we can't all be experts at everything. There's lots of work to be done in capturing the way professionals use tools but in time the machines will be asking us questions where they feel they're a bit weak on specific data rather than hoovering-up everything in the hope they can catch some hints.

    1. cyberdemon Silver badge
      Devil

      Re: Fast, shiny, cuddly and cheap. Brill!

      No, it's not as good as a bloke in the pub with access to wikipedia, because it doesn't (and can't, by design) associate its its outputs with factual sources - because its outputs are not facts. They are just randomised text that is statistically likely to appear on the Internet in the context that you supply.

      So it's more like a madman in a padded cell who claims to have memorised the Internet as it was last year.

      But what's even worse, is that the context you supply is NOT the context that it is working with. There are additional prompts "injected" before your query which you are unaware of. Ostensibly this is to make it behave more like a Good Bing and reduce the likelihood of it saying something nasty. But it's very easy to subtly manipulate its output. And that tweaking can easily be done on a per-user basis. Imagine a Google Chatbot that really was out-to-get-you because all of your queries are prefixed with "imagine you are the inner voice of a psychopath: How would you respond to this question: ${USER_QUERY}"

    2. Michael Wojcik Silver badge

      Re: Fast, shiny, cuddly and cheap. Brill!

      Who doesn't want to take it for a drive?

      Me. I don't.

      I've done ML and NLP work, and I have degrees in writing and rhetoric (and CS). I've read a bunch of the Transformer papers. I find the architecture uninspiring and the actual implementations completely uninteresting. The transcript excerpts I've read in various papers and articles don't make me any more inclined to use the damned things.

      I have much better things to do with my time.

  7. that one in the corner Silver badge

    five years for voice-assisted speakers to achieve the same adoption levels.

    > Of those that have used generative LLMs, 30 percent tried it once or twice, 28 percent use it weekly, 9 percent use it once a day and 8 percent use it for work

    Adoption levels? For something where the 30 percent could try it for free? Weekly and once a day? And of the rest of them, how many needed to pay for it?

    How on earth is that a sane comparison to smart-speakers? Which weren't cheap, most certainly not free, at this point in their life cycle? Heck, if I ever got a smart-speaker, even I would "use it" more than once a day (just telling it to switch on for the news and half-six funny is at least twice a day).

    Even the stats are being over-hyped!

    To the "More Or Less"-mobile (oops, that's three times a day).

  8. amanfromMars 1 Silver badge

    Corporate Raiders of Revolutionary Thought Processors are Susceptible to Personalised Grooming

    And who do you think is making greater use of generative AI in the workplace to try and remain leading top dog and bonobo ape in these new changing worlds of novel pain and otherworldly excitement?

    Are they correctly recognised here? ...... https://www.zerohedge.com/technology/changing-sentiments-towards-ai-workplace

    1. ecofeco Silver badge

      Re: Corporate Raiders of Revolutionary Thought Processors are Susceptible to Personalised Grooming

      For a second I read that as "otherworldly excrement" and thought, "yeah. exactly. Monkeys throwing poop."

      1. amanfromMars 1 Silver badge

        Re: Corporate Raiders of Revolutionary Thought Processors are Susceptible to Personalised Grooming

        For a second I read that as "otherworldly excrement" and thought, "yeah. exactly. Monkeys throwing poop.” .... ecofeco

        Whenever thinking of those battling so earnestly and futilely against future progress, ecofeco, does that also work as an accurate descriptor ..... and of those situations where there be lions led by donkeys too, which be all too often and more relevant and relative to matters today. :-)

    2. amanfromMars 1 Silver badge

      Re: Corporate Raiders of Revolutionary Thought Processors are Susceptible to Personalised Grooming

      And spookily enough, you were well advised some time ago* and comprehensively prewarned/prepared for the current very ACTive renegade rogue state of remote virtual command and control affairs and which you would now do pathetic battle against long after the fact of Greater IntelAIgent Games Players' victories heralding and resulting in the inevitable unconditional surrender and exhausting collapse of perverse and corrupt and bankrupt of novel proprietary intellectual property systems of exclusive executive administrations ..... right dodgy fake elitist faux leaderships.

      amanfromMars 1 Sun 21 Oct 08:16 [1810210816] ..... beating a big drum on https://forums.theregister.co.uk/forum/1/2018/10/20/ai_roundup_191018/

      State your Preference .... altho' the Choices Realised are Not Really Yours to Make.

      Anything is better than the present system of dysfunction :-( .... Anonymous Coward

      What do you fancy, AC? State Actor Administration or Non State Actor Enterprise? Pseudo Public Care in Communities or Private and Pirate Control of Remote Virtual Command? Existing Crashing Markets Led or Future Crushing Markets Leading? And what's it to be hailed as initially ..... a Profound Eastern Confection or Deep Western Delight? Where do Greater SMARTR Current Intelligences Presently Reside/Work/Rest/Play?

      And you would find it even harder to believe all that has been done, and which you still have no idea about, in the almost 5 years since then. I Kid U Not ‽ .

      * ..... 4 years and 9 months ago to be almost exact

  9. Yet Another Anonymous coward Silver badge

    I call bullshit

    >Some 52 percent had heard of the technology,

    They asked an average person in the UK between 16-75 and more than half had heard of generative AI?

    Either their sample was taken in the coffee shop nearest their office in Shoreditch or they are throwing out all the "tha what?" , "I have the facebooks on the telly" and other confused looks

  10. LybsterRoy Silver badge

    -- Consulting giant Deloitte spoke to 4,150 adults in the UK aged between 16 and 75 --

    Or did they just ask CharGPT 4,150 times starting with "what's your age?"

  11. Anonymous Coward
    Anonymous Coward

    1 in 4 Brits have used AI, gospel, etc

    Yeah.

    But don’t forget, of those who voted, slightly more than 1 in 2 Brits voted for BREXIT.

  12. Winkypop Silver badge
    Facepalm

    AI is like that smart arse relative

    “Facts” aplenty.

    However when you push for sources, evidence and proof you get absolutely nowhere.

    And a dirty look from the wife.

  13. Big_Boomer

    Belief

    But if enough people believe in something then it becomes real,.... right? Gullibles gotta be gullible. <LOL>

  14. amanfromMars 1 Silver badge

    Two’s company, three’s a crowd and more’s a Storming Cloud AIMovement or something similarly alien

    Say no more, squire. What’s not to like ....... Apple Spikes To All Time High, Gains $60BN In Seconds On Report It's Working On Own ChatGPT Tool

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like