[ai-control] Re: -04 drafts for evaluation

[Leonard Rosenthol <[e-mail]>]

What we are building, as we have said before, are “preferences”.  By themselves, they are not binding (technically or legally) in any way.  Our vocabulary will - by itself - not restrict users from doing anything.

But as been voiced by others many times is that we MUST work with policy makers (and implementers) to guide them about the threats and harms that can be posed if the Vocabulary and/or Attachments are mandated in inappropriate ways.

Leonard

From: Farzaneh Badiei <[e-mail]>
Date: Monday, November 3, 2025 at 6:22 AM
To: Paul Keller <[e-mail]>
Cc: [e-mail] <[e-mail]>
Subject: [ai-control] Re: -04 drafts for evaluation


EXTERNAL: Use caution when clicking on links or opening attachments.


Thanks Paul and Leonard.

Sorry to be pedantic about this but if we have asset based signaling, instead of location based how are we going to ensure that this doesn’t affect end user accessibility features? I know we discussed this and there is a sentence in the ID that addresses it but I don’t think a one sentencer is enough. How are we going to ensure that emails can be summarized for those who want it. And summarization is an accessibility issue, both for people who have low bandwidth connectivity and people who have other accessibility issues.
I am just worried that more and more  we are moving towards protecting business models instead of serving all end users.

On Mon, Nov 3, 2025 at 6:00 AM Paul Keller <[e-mail]<mailto:[e-mail]>> wrote:
Thanks for your response Farz!

> On 2 Nov 2025, at 18:50, Farzaneh Badiei <[e-mail]<mailto:[e-mail]>> wrote:
>
> If Foundation Model Production replaces AI training and Generative AI training why not just use those terms? And in my opinion (and I have been trying to educate myself on this AI terms for the past two years so bear with me) Model Production is the stage after training and production  means producing something. So in a way I think Foundation Model Production is dealing with after training but in our case it includes both production and training. Which makes it broader.

I do not think that the “production” in the name of the category refers to the production of something with foundation models. The draft clearly states that it refers to “the act of using an asset to train or fine-tune a foundation model.” That’s why I see “Foundation Model Production” as a replacement for the earlier “AI training” and “Generative AI training” categories, but not as something that can substitute the overarching category — which also includes the categories dealing with the use of assets by deployed models.

> I have another question, which we have raised a number of times so I apologize if you answered already: why do we need an "all" or a "tdm" or an "automated processing" category when the bot can be blocked anyway at robots.txt level should they want to allow or block it? I think it's important to have a clear cut answer. Krishna has thankfully raised the issue here again and he explains it better than me: https://github.com/ietf-wg-aipref/drafts/issues/178

I agree with Leonard’s answer on that point. The whole reason we have two different drafts (attachment and vocabulary) is that the vocabulary is meant to ensure interoperability across different attachment mechanisms and robots.txt is just one of them (albeit an extremely important one).

/Paul


>
> I am working on my other comments and will send it to the list.
>
>
> On Sun, Nov 2, 2025 at 12:23 PM Paul Keller <[e-mail]<mailto:[e-mail]>> wrote:
> Dear all,
>
> I also have a question of understanding.
>
> @Farz – in your last message you wrote:
>
>> On 1 Nov 2025, at 21:13, Farzaneh Badiei <[e-mail]<mailto:[e-mail]>> wrote:
>>
>> I have to think really hard about  replacing training generative AI with the Foundational Model Production category. Now that we have that change, then automated processing definition has become even more problematic. Too broad, it includes everything.
>
> This echoes something Jo wrote earlier in this thread:
>
>> On 30 Oct 2025, at 23:20, Jo Levy <jlevy=[e-mail]<mailto:[e-mail]>> wrote:
>>
>> Moreover, the change to the term "Foundation Model" instead of AI Training, makes it even more apparent that a top level category is not needed.
>
>
> Could either of you help me understand why you think this is the case? I’m trying to follow the reasoning, but I’m not sure I do. In my reading, the new “Foundation Model Production” category simply replaces the previous “AI training” and “Generative AI training” categories. I don’t see how that change affects the rationale for having (or not having) a top-level category. But maybe I am missing something here?
>
> Best, Paul
>
> --
> ai-control mailing list -- [e-mail]<mailto:[e-mail]>
> To unsubscribe send an email to [e-mail]<mailto:[e-mail]>