I think "take responsibility for the [AI-written] code" is a hard hill
to climb. Even with the "No code gets accepted unless it is reviewed
by a human" statement, you still don't know the origin of the
AI-generated code. (i.e. what it was trained on, and if AI reproduced
some code it was trained on.)
As an example: Way back at the beginning of this thread, I mentioned
that I've seen AI "vibe coding" regurgitate code I wrote elsewhere. (I
recognized my own code from an article I'd written.)
:: I've also seen examples where AI “vibe coding” regurgitated
:: some of my code. That was a personal experiment for a “niche”
:: topic where I asked copilot to write a version of nroff; the
:: AI cited an article I wrote in its summary about “here's how
:: this works” and I recognized my own code in the generated
:: output. This is dangerous because in the online article it
:: pulled from, I wrote about how this isn't a safe way to write
:: this (doesn't catch edge cases, etc) but it was an easy way to
:: show it to non-programmers. But copilot didn't add that caveat
:: to the code it generated for me.
Let's explore this further: Some time ago, I wrote an op-ed that
discussed AI coding issues in a "quadrant" model. I can't find the
article right now, but imagine "your project's license model" on the x
axis, and "the code base the AI pulls from" on the y axis. It's a
quadrant, so each axis only has two values; "open source" or
"proprietary." I identified 4 issues:
1. Open source project, AI reproduces open source code
Not all open source licenses are compatible with each other. Consider
an MIT-licensed project, but AI reproduces code from [someone else's]
GNU GPL source tree. You can include MIT-licensed code in a GPL'd
project, but you cannot include GPL'd code in an MIT-licensed project.
AI has introduced a license violation, but you don't know about it.
It's a matter of time before someone finds it, and now you have to
audit the code, rip out the GNU GPL stuff, and rewrite the offending
code.
2. Open source project, AI reproduces proprietary code
In my example, I used GitHub Copilot. Is Copilot trained on
proprietary code trees? Best outcome is a "cease and desist," and you
rewrite the proprietary stuff. Worst case is another "SCO v Linux"
lawsuit.
3. Proprietary project, AI reproduces open source code
4. Proprietary project, AI reproduces proprietary code
Both of these are bad for enterprise customers that use AI for coding
assistants. #3 might be found during an internal code review. #4 might
be found during a "due diligence" review (such as a corporate merger).
(The case where an AI generates purely original code wasn't part of my
discussion.)
I'm very interested in protecting FreeDOS against issue #1 ("AI
reproduces open source code") and issue #2 ("AI reproduces proprietary
code"). Let's say we didn't have a "no AI" policy for the core parts
of FreeDOS. I wouldn't want to find out 2 years from now that
someone's AI-generated code has a problem, and then have to rip it out
and rewrite it. We don't have enough active developers to do that.
And as I said at the beginning of this thread, there's another legal
issue: wholly AI-generated works cannot be copyrighted. (I'm unclear
about "AI-assisted" works, but I'd rather not find out the hard way.)
I've always been very careful about contributions to FreeDOS that
could be "tainted." (See "OpenDOS" and "Computer History Museum
MS-DOS" from many years ago, for example.)
But I only care about the core parts of FreeDOS. If a developer uses
AI to work on a "third party" DOS tool that we happen to include,
that's up to the developer -- as long as it's not included in a "plain
DOS" install. Games, apps, editors, .. etc .. are outside the "no AI"
policy.
On Thu, Jun 18, 2026 at 11:13 AM Michael Brutman via Freedos-devel
<[email protected]> wrote:
>
> Trying to determine the correct uses or the amount of correct usage for
> an emerging technology is probably a mistake. AI tools (not "clankers"
> please, that's a trendy but worthless term) are changing rapidly.
>
> Consider doing the following:
>
> - The person who submits the code must acknowledge if an AI tool
> was used, and to what extent. (Did it generate code? Test cases?
> Language translation?)
>
> - That person takes responsibility for the code, as they are the one
> that submitted it and invoked the AI tooling.
>
> - *No code gets accepted unless it is reviewed by a human.* This
> is the big one as it hopefully filters out code that is incomplete or
> possibly dangerous. Many people are using AI tools, but they don't know
> how to evaluate the output. Human review by a domain expert is needed.
>
> - The code may be removed in the future if the AI contribution to it
> becomes problematic. (I'm thinking about copyright issues here.)
>
> Ultimately, accepting code or contributions is an editorial process.
> The editor has the right to accept or reject for many different reasons,
> and using AI tech is just another factor in that decision process.
_______________________________________________
Freedos-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/freedos-devel
