Skip to content

Policy regarding AI assited contributions #740

Description

@StaticRocket

We haven't gotten anything blatantly tagged with AI assisted tools until #642 and #643. I think we need some policy in place to control documentation quality and insure we don't accidentally trigger any copyright issues.

Quality: I'd hate to have a customer informed about a quirk that doesn't exist, an API who's interface doesn't match what we ship, or a step in some walk-through that's generally ill advised. The review process currently is more strict that it used to be, but the content still depends on multiple reviewers familiar with the component in question. To be quite frank, we don't have the resources to handle a full guide's validation in a timely manner at the moment.

Copyright: Seems a lot of projects have given up on waiting to see how that plays out. We could do the same, but we should look to see if there are any automatic copyright detection tools we could use to prevent issues, regardless of AI assistance. I know there are tools frequently used in Academia for this purpose. The specialized database that also requires frequent updates usually means these are paid and will be a pain to automate. They are also frequent to false positives, which is not something we need more of here.

Poisoning: There have been multiple attempts to roll all of TI's documentation into a single chatbot like thing to answer product related queries. I don't know what the current status is, nor do I care to know. I figure this will be used as one of the many sources regardless. Should we be concerned about the feedback loop we'd be creating by allowing assisted commits here?


Proposed solutions:

  1. Adopt Linux kernel / Mesa / LLVM contribution guidelines and use Assisted-by trailers.
  2. Adopt a permissive policy that preserves the content like NLnet Labs.
  3. Reject any assisted commits outright. We haven't taken any (disclosed) tainted commits yet.

I'm of the opinion nothing generated by a bot deserves the attention of a human operator. Barely anyone reads these documents on a good day. If it gets any worse we might as well not publish anything. I know this is somewhat extreme compared to my coworkers opinions on the subject. I'm hoping to get information about TI's general guidelines regarding tool assisted documentation and what should be enforced here.

Metadata

Metadata

Assignees

No one assigned

    Labels

    questionFurther information is requested

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions