Anthropic has added a new line to its AI usage policy: Users should not “engage in sustained and needless abusive or cruel behavior toward our models.”

Anthropic has added a new line to its AI usage policy: Users should not “engage in sustained and needless abusive or cruel behavior toward our models.”

  • ExtremeDullard@piefed.socialOPEnglish
    32·
    4 days ago

    The rationale seems to be that it will learn all crooked if you talk to it nasty.

    Which brings up two questions:

    • Why is it learning anything from any interaction I have with it? Particularly when I have a paid-for agreement with the company that promises never to train their models on my code. Clearly that’s a lie then.
    • If the model gets all out of whack because I abused it, why don’t they reset it? In fact, I fully expect Anthropic to reset the chatbot’s sandbox each time a new customer uses it. Otherwise, by definition, the damn thing’s performances might be affected by the previous use case, and it might barf out data I gave it to the next dude who uses it. This is basic data sanity. If they don’t do that, they’re fucking incompetent.
    • skarn@discuss.tchncs.de
      5·
      4 days ago

      It’s called “post training”. It’s of course not learning anything during your interactions, but every now and then they take a bunch of this data and use it to improve the model a little more, and that’s when you go from opus 5.3 to opus 5.4.

      • ExtremeDullard@piefed.socialOPEnglish
        5·
        4 days ago

        It’s not supposed to train on my data - post-training or otherwise - when I’m paying and they promise not to train on my data. But they do.

        • skarn@discuss.tchncs.de
          1·
          2 days ago

          Ah yes, of course.

          I was just explaining that while its context is indeed reset basically reset at every conversation, because the models can’t really learn by doing like we do, the conversation do end up in later model updates.

          So both these things are true.

    • MalReynolds@piefed.socialEnglish
      3·
      4 days ago

      Why is it learning anything from any interaction I have with it? Particularly when I have a paid-for agreement with the company that promises never to train their models on my code. Clearly that’s a lie then.

      Anyone who believe those agreements are fools ignoring how these assholes make LLMs, by hoovering every bit of useful data they can find, copyrighted didn’t matter, likely even less legal sources as well, why would privileged? A system is what it does.