OpenAI have claimed that a teen violated terms that prohibit discussing suicide or self-harm with the chatbot after the AI allegedly encouraged him to take his own life.
  • lemmy_outta_here@lemmy.worldEnglish
    162·
    4 天前

    And I bet teens are going to go on violating the TOS. Maybe we better restrict AI to people who could potentially read and understand the TOS if your product is so dangerous that lifesaving instructions are contained within.

  • Peruvian_Skies@sh.itjust.worksEnglish
    8·
    3 天前

    So if a teen shoots up a school even though the gun manufacturor says not to shoot up schools, suddenly we don’t have a gun control problem but a teens don’t read the TOS problem.

    Yeah, that sounds right for the USA.

  • Die4Ever@retrolemmy.comEnglish
    1·
    3 天前

    This isn’t the first time ChatGPT has had this issue? ChatGPT supports “tools” which means the LLM can call to a traditionally programmed function to do things like search the Internet or do math, why don’t they just add a tool for it to be able to report/flag a chat? Everyone knows ChatGPT already isn’t private. The prompt could say to report/flag chats about suicide or illegal activity, I think the LLM would pretty reliably report them before they get very far.

    I guess this case started back in April though, so maybe they’ve already done this.

    I asked ChatGPT what tools it has, but it’s possible there are some it isn’t allowed to talk about.

    tools

    Yes — I support a number of built-in tools. Here is the complete list of tools available in this environment:

    ✅ All Available Tools

    1. web Search the internet or open webpages. Used for fresh, factual, time-sensitive information.

    2. python Run private Python code (not visible to the user). Used for internal reasoning, analysis, parsing, etc.

    3. python_user_visible Run Python code with visible output — charts, tables, files (PDF, DOCX, XLSX, etc.).

    4. canmore Create and update canvas documents (markdown, code, multi-file-style editing).

    5. bio Memory system: store or forget user information when requested.

    6. api_tool A file-system–like abstraction for discovering and invoking additional tools. (Meta-tool.)

    7. image_gen Generate or edit images based on descriptions.

    8. container Run commands inside a sandboxed shell-like environment.

    If you want, I can also describe what each tool is best for, or show examples of how they’re used.

    https://chatgpt.com/share/692f1cc7-3a80-8010-afa2-6576474fae38

  • w3dd1e@lemmy.zipEnglish
    2·
    3 天前

    The transcripts from that ChatGPT log are bad. OpenAI is fucked if they can’t settle