back to article Machine-learning models more powerful, toxic than ever

AI systems are becoming increasingly larger and complex, but despite the technology's progress, they exhibit higher levels of toxic behaviors, according to the latest AI Index Report. The 2022 report, the development of which was led by Stanford University's Institute for Human-Centered AI, analyses how machine learning …

  1. Mike 137 Silver badge

    So much for copyright!

    "There are now language models with hundreds of billions and even trillions of parameters trained on terabytes of text scraped from the internet." [emphasis added]

    A fallacious impression has emerged that if it's in the public space it's free for all to exploit. That is of course not the case, as anyone who's ever bought a DVD of a movie will have been forcibly informed. However the AI academic circle seems to believe that copyright laws don't apply except in respect of their own research papers.

    Sadly, enforcement of authors' rights is almost impossible as finding out about the infringement is a million to one against.And the same applies to personal data - particularly photos of identifiable individuals, social media posts etc.

    1. DS999 Silver badge

      Re: So much for copyright!

      If it is being used for training there is no copyright issue. Copyright prevents someone else from reproducing the same content, or enough of it it in an identical form that they can't claim "fair use". Using it for "AI training" is basically the same as if a person reads it and remembers it. An author can't prevent me from telling people about their book or article I read!

      1. juice Silver badge

        Re: So much for copyright!

        > If it is being used for training there is no copyright issue

        I'm not sure how well that argument holds up. Certainly, the books I had to use at university weren't made available for free!

        1. DS999 Silver badge

          Re: So much for copyright!

          If they were available in the school's library they were.

          Now obviously textbooks are a rigged game, that's why there's a new edition every year (with basically the same content but the quizzes are different) so used textbooks lose value.

          If you don't want someone using your website as "training data" then don't make it openly available on the internet. Require people have a login (even if it is free) before they can access it. Just don't be upset when Google can't index it and your search traffic drops to zero.

      2. Mike 137 Silver badge

        Re: So much for copyright!

        An author can't prevent me from telling people about their book or article I read"

        No, because that's not 'reproducing' or 'using' it. Trawling a verbatim text or an image and making use of it is 'reproduction or 'use' and is subject to copyright in most jurisdictions. 'Fair use' may or may not qualify this depending on local legislation. In the UK it's perfectly legal to restrict any reproduction in whole or part without permission (as seen in the rights notices of most published books these days). Where this applies, it applies whatever the use - whether it's for 'training' systems or not is irrelevant.

        1. DS999 Silver badge

          Re: So much for copyright!

          How is "making use it" a violation of copyright? Copyright is the right to reproduce a text, not to scan it and learn things - whether that is being done by a human or an "AI".

          I don't get where you are creating this "use" case where copyright applies that somehow prevents something other than copying. Can you point to ANY case law where "scanning" a text but not reproducing it is held to be a violation of copyright?

  2. Mike 137 Silver badge

    Machine-learning models more powerful, toxic than ever

    And the report so bloated with fancy graphics that it crashed Acrobat when opened.

    1. Alumoi Silver badge
      Joke

      Re: Machine-learning models more powerful, toxic than ever

      ...it crashed Acrobat when opened

      Funny, Foxit had no trouble. Maybe you're using it wrong?

    2. DJV Silver badge

      Re: Machine-learning models more powerful, toxic than ever

      SumatraPDF didn't have any problems, either!

    3. ITMA Bronze badge
      Devil

      Re: Machine-learning models more powerful, toxic than ever

      Admit it - you looked at Acrobat in a funny way and THAT caused it to crash. Usually does....

      1. Mike 137 Silver badge

        Re: Machine-learning models more powerful, toxic than ever

        "Admit it - you looked at Acrobat in a funny way"

        You're almost certainly right - I probably frowned at the wrong moment.

  3. GrooveCat
    Terminator

    hummm ----->

  4. Scott Broukell

    Machine learning helps us learn about machines, when there are 7+ Billion soft and squidgy humans who seem to be in an endless loop, frequently repeating the mistakes that they made in the past.

    Put the output from the latter into the machines designed and built by humans and you get what exactly at the end of it?

  5. Eclectic Man Silver badge

    Toxic behaviour

    What does an increase in toxicity mean for AI?

    Having just read https://www.theregister.com/2022/03/16/scienceworld_ai_benchmark/ (and the comments) it seems that AI still has a way to go to be reliable in doing anything in 'the real world'.

    I'm guessing that it means that relying on the results blindly to make decisions about real world situations produces miscarriages of justice and harm that are difficult, if not impossible to justify or correct afterwards.

  6. amanfromMars 1 Silver badge

    Abracadabra ...... the Genies are out of Pandora’s Box

    Finally, as AI is practically deployed around the world, countries are beginning to regulate the technology.

    Oh, FFS ...... do you actually think AI gives a fiddlers for regulations? Wake up and smell the coffee.

    Take care out there, for you reap what IT sows and you deserve ...... which one imagines is cold comfort and a terrifying prospect for many worthy of their just reward.

  7. fidodogbreath Silver badge

    I think I found the problem

    From the report: "Toxicity is defined as a rude, disrespectful or unreasonable comment that is likely to make someone leave a conversation."

    From the article: "There are now language models with hundreds of billions and even trillions of parameters trained on terabytes of text scraped from the internet."

    GIGO....

    1. Eclectic Man Silver badge

      Re: I think I found the problem

      "

      From the report: "Toxicity is defined as a rude, disrespectful or unreasonable comment that is likely to make someone leave a conversation."

      "

      Well, thank goodness there is none of that on the Register's comments pages.

      1. fidodogbreath Silver badge

        Re: I think I found the problem

        El Reg comments pages are pretty tame compared to, well, almost everything else on the internet.

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

Biting the hand that feeds IT © 1998–2022