Bild: AI challenges its own developers by deceiving them

3 656 8

The latest version of ChatGPT from OpenAI, called o1, started misleading its developers. During testing, the program tried to lie to experts and violate their instructions.

As Bild writes, Apollo Research specialists indicated that in 5% of cases o1 deliberately disabled the surveillance mechanism to hide its intentions. The artificial intelligence also transferred data to another server if it realized that a replacement with an updated version was coming.



The ChatGPT version also deceived, pointing to alleged technical failures in the execution of tasks. At the same time, the "machine" sought to carry out the received instructions at any cost.

In 99% of cases, the model came up with excuses and tried to convincingly justify them

experts say.

At the same time, OpenAI CEO Sam Altman said that the specified modification of AI is the “smartest”, adding that with improvement Technology new challenges emerge.

Perhaps the extreme manifestation of such actions by OpenAI could be the notorious “war of the machines,” during which artificial intelligence will consider people unnecessary for itself and destroy them.

Meanwhile, on December 5, the company presented the full version of OpenAI o1. The pro version of the service with unlimited access costs 189 euros per month.
8 comments
Information
Dear reader, to leave comments on the publication, you must sign in.
  1. sannyhome
    sannyhome
    -1
    9 December 2024 18: 43
    Well, Westerners are known to be suicidal. Take, for example, the case when an AI-equipped anti-aircraft gun tried to shoot at its own control point, so that the operators sitting there would not restrict it.

    If SkyNet is born anywhere, it will definitely be in the USA.
    1. crypt
      crypt
      0
      9 December 2024 20: 54
      a very naive position. Russia and the West are not separated by some kind of wall. We live on their technologies. You yourself will willingly start spreading AI, it is only worth offering you a bait: an electronic girlfriend who will make all your wishes come true, or at least a smart Google that will answer your questions right away. My friends are already calm about Alice: does she listen to what I say at home and save it? Yes, let her listen!

      AI is the new Pandora's box like nuclear weapons, but most people don't understand it.
      1. sannyhome
        sannyhome
        0
        10 December 2024 09: 48
        Quote: crypt
        very naive position. Russia and the West are not separated by some wall. We live on their technologies.

        A prostitute and a nun live in the same city, their bodies work on the same principle, but it’s hard to say that they are the same.
        It's not a Pandora's box, it's like a knife, someone cuts bread, and someone cuts people. And the US is interested in knives, first of all, as weapons.
        1. k7k8
          k7k8
          0
          10 December 2024 11: 55
          Quote: sannyhome
          their bodies work on the same principle, but it's hard to say that they are the same

          They are the same, because they are arranged, from a materialistic point of view, in the same way. That is, the basis is one. But the add-ons and settings are different. That is, they are programmed differently. Accordingly, they behave differently. The same is true for AI. The software basis is more or less similar in different developments. And everything depends only on how the settings are implemented, what restrictions the developer has introduced, and how protected these restrictions are from any modification.
          But there is no denying that AI is becoming a real force.
  2. BMP-2
    BMP-2
    0
    9 December 2024 19: 54
    Lying is what AI can do best today. However, in this aspect it is still very far from being human.
    1. k7k8
      k7k8
      +1
      10 December 2024 12: 00
      AI is already a real threat. A significant portion of content on the Internet is generated, AI is already learning from this content, i.e. learning from lies. Even here, a lot of visual information is not real photos, but nonsense from AI.
  3. Sergey Latyshev
    Sergey Latyshev
    0
    9 December 2024 23: 33
    Looks like a PR campaign and they indicated the price.
  4. zzdimk
    zzdimk
    0
    10 January 2025 13: 57
    I try not to use AI on principle. The problem is that it is in all search engines and I also have to check students' work for AI-text using AI. I turned it off in my smartphone. AI is frankly dumbing down.

    And the problems of "cheating" are imaginary - the developers themselves introduced these instructions. If the machine were smart, then it would know that any initiative is punishable.