Session Date/Time: 16 Mar 2026 03:30
The following transcript is a complete and verbatim record of the AI Preferences (aipref) Working Group session at IETF 119.
Mark Nottingham: Okay. It is 11:30 local time, so let's go ahead and get started. As I said, this is AI Preferences. I know that in the last session, a number of groups had some issues with remote participation, so apologies if that happens again here. We'll do our best. If you are remote, please bear with us and try and get our attention through chat if there are issues, but be aware this meeting, I don't think we're going to be having terribly momentous discussions. This is mostly a status update for the rest of the IETF community and kind of contextualizing it for folks, so if there are issues, hopefully you won't miss out on anything critical. And of course, we always make our decisions on the mailing list.
We'll start. This is the Note Well. If you're new to the IETF, you may be becoming familiar with this. This is the policies and procedures under which we operate, especially regarding things like intellectual property and also things like your behavior, so harassment and the Code of Conduct. This is important. Please do take time to understand this. You can find it by following the handy QR code there or by searching using your favorite internet search engine or chatbot for IETF Note Well, although please don't allow a bot to interpret it for you.
I want to highlight one aspect of this. And if we could close the doors in the back, that'd be appreciated so we don't have the noise filtering in. I wanted to highlight that the Code of Conduct, we try and keep this a professional and welcoming environment. We do take that seriously, and especially in this group, we know that it's sometimes contentious, sometimes feelings run high. We ask people to remember that, especially when you're talking about other people and other people's motivations or their state of mind. We've had a few instances where people have sailed a little bit close to the Code of Conduct, and we want to keep it a professional and welcoming environment. Anything to add there?
Suresh Krishnan: No, I'm good. I think you summarized it well. Thanks.
Mark Nottingham: Right. And if you have any thoughts about that or concerns or want to know more, feel free to talk to Suresh, myself, the Area Directors, or anyone else.
Mike has pseudo-volunteered to take minutes. You're the best. Let it be put on the record by the minute-taker that the minute-taker is the best.
Speaker 1: Oh, cool.
Speaker 2: Yeah, please. Thank you. Thanks.
Mark Nottingham: Okay. So our agenda for today is, as I said, mostly a status update.
Okay, Mark Dan is saying they're not seeing slides on the Lite client. Okay, that's not good. I can share the links that I am projecting into Zulip if that helps folks. So right now we are on the famous Note Well slide, which is that. Goal setting, interesting. Okay.
So as a reminder, the working group has a homepage. I will paste this into chat as well. So if you're not familiar with this, this is where you can find links to all the materials for the working group. And of special note might be to folks our charter, and I'll paste this link as well.
So you can read through the whole thing at your leisure if you're not familiar with it, but the important part for this working group is mostly in the first sentence: "The AI Preferences working group will standardize building blocks that allow for the expression of preferences about how content is collected and processed for artificial intelligence model development, deployment, and use." And especially the phrase "expression of preferences" is important to understand there. We're not creating a technical enforcement mechanism, and in fact, later down you'll see that's out of scope. We're just allowing people to express their preferences about how things happen. And indeed, that's much in the line of how robots.txt works, which it turns out is the basis of a fair amount of our work.
Mark, are you sharing through the share slides stuff or through the...?
Mark Nottingham: I'm sharing the screen.
Suresh Krishnan: Okay.
Mark Nottingham: And so we had a meeting, an online meeting, about a month ago now, where we spent some time going through the issues list. So for those who haven't been following the work closely, we were chartered and we started work early last year, in January of last year. And we had some, a couple of successful interim meetings. We had some documents that we had some early momentum behind, and we went to working group last call with those. But because of various things happening and folks joining the work and some objections that were raised, we weren't able to complete that working group last call.
And in the subsequent discussions, we've had some more interim meetings and kind of pulled those apart and figured out what we can do. We've been paring down the scope of the work until we can come to agreement on the core aspects of the work. And so what you're seeing here is after most of that, I think in the last revisions of the drafts, we've taken a lot of the more contentious terms out so we can focus on refining the semantics of the vocabulary that we're working on so we can have something as that core. And then we'll see if people have proposals that they can bring to the table to add functionality that they want back, and that's still to be determined.
And so what we're doing right now as a working group is working through the issues list and trying to get people to think in terms of making proposals and keeping the discussions going, because what we've observed is we schedule these interim meetings and we make some amount of progress and then there's very little progress in between them. And so we're trying to encourage working group participants to, you know, work offline, write internet drafts or just write standalone proposals, contribute to the issues list, but to keep the work going between these interim meetings.
And so the interim meeting we held in March, the purpose was to go through these issues list and try and get the disposition of each issue, assign issues where we could, or spark some people's discussions about what they could bring to our upcoming interim meeting, which is in April, for a deeper discussion there. So that was the, if you didn't attend that meeting, I'd encourage you to go in and look through the minutes of that meeting, which are on the screen now, and I just pasted into the chat, because that is a pretty good snapshot of where we're at right now in combination with the current drafts.
And so our current issues list is much as it was in that meeting. I'll paste the link into the chat here. A couple of things have come out of that meeting. One is we found we kept on circling back on the usage proposals that were coming up. A number of people have wanted to express the need or desire to control how AI uses their data after training time, so at inference time, we've called it at one point. There have been a couple of different terms for this. And so we took those terms out of the drafts and we replaced them a few times, and there have been proposals that have not been adopted.
We've collected kind of a running commentary of what's happened there on this webpage, which again, there's a link in the chat as to why that happened, because what we're finding is people are joining the group and bringing the same proposals up or wondering why we're not considering that at this time. This is trying to capture that. And especially the most recent go at that, we spent a lot of time in the Zurich interim last year now, coming up with a proposal for what we called AI output, which was describing preferences for how data is treated in the output of an AI model. And that seemed promising in Zurich; unfortunately, under later scrutiny, that kind of fell apart. And it got to the point where as chairs, we felt that it wasn't suitable to keep in the draft because people were focusing on small aspects of criticisms against it, and it was clear that what we really needed to do was to step back and have a bigger discussion about what the appropriate approach to solving this area of issues is.
Likewise, there was a discussion around search, and we put together a summary page for search as well, describing the evolution of the terms around that. This was started as a carve-out that working group members felt was important that if I have content on the internet and I want to say, no, I don't want you to use that for AI, but with a caveat around that of what exactly does that mean, of course, but there was a concern that if you did that, you might be taken out of search rankings. And many websites, especially, derive a lot of value from being locatable by search. And so it was felt that that was a, or it was put forth that that was an imbalance of power there and it was not a good outcome if search engines could say, well, you can only be in the search engine if you allow our content, you allow us to use your content for AI training, for example.
And so the search term originally was a carve-out to make sure that it was clear that I don't want you to use my content for AI, but I still want to be in search. As the discussions have evolved, that has been a bit shaky because we've come to understand that almost all search now is, has AI under the covers. And so the real question that's kind of come up is: do we really want to talk about this in terms of controls for AI, or is this more about controls for how search results are presented and how data is used when it's used for, for example, presentation in search or for example, a chatbot or some other interface? And part of the problem here is that the field is evolving so quickly, it's difficult for us to describe universal semantics for these things. So this page is trying to capture the state of the discussion there.
I'll take one or two other things to cover now, but I'll take a break. Martin Thompson, are you online? Did you have anything to bring up about the drafts in their current state?
Well, he's in a remote room, so he's probably... yes, there's lots of different things to consider there. Oh, and our... okay. All the bits ran away. Martin is typing his response. I see a closet. I see a hand. And now nothing.
Martin Thompson: Oh, it's up there? Sorry about that. No, no, no, this is all fine. All right, so I've put some messages in the chat, but essentially, not a lot has changed since we last met on this front. And I think part of that is because of the call from the chairs to have folks develop new proposals around things like the overarching category that we talked about or around inference and so forth. And so it would be good to have more discussion about those things, but we haven't seen anything sort of coalesce yet. I know that people are working on some things.
But until we have that, we're sort of waiting, I think, from the editors' perspective. Paul and I are discussing a few things that we think we can improve in the drafts around the edges, particularly things like the text around conformance and whatnot. But none of those really have much bearing on the real problems that we have, which is exactly the sort of things you already discussed there around search and whatnot. So until we have more, then I think we'll be waiting. Hopefully, we'll have some more to discuss ahead of the meeting that we have next month coming up, the interim.
Mark Nottingham: Okay. Thank you, Martin. I agree. I think there are some things the editors can do, but they are largely around the edges that can help us avoid a few distracting discussions that keep on coming up if we, you know, if people keep on fixating on certain words and we can get past that. But the bigger, meatier issues are ones that we need to address as a group.
Suresh Krishnan: Yeah, I think the use case proposal is a really, really good way to tackle this. So if we can start documenting what are the use cases and see if they're addressable, I think that could probably go in a more productive direction, right, Mark?
Mark Nottingham: Yes, and funny you mention that. Although, Meetecho, hello? Okay, so we'll pause the meeting until... okay, okay. And for people who are bored, you can think of the Monty Python-like people who are responsible for stuff skit in your mind because I cannot share the link.
So we have a little problem. The server is currently unable to handle the request. Okay. Nope, nope, nope, nope. Okay, let's just talk then.
Suresh Krishnan: That's fine. I'm not in the meeting Wi-Fi. That's probably the difference.
Mark Nottingham: Interesting. We are getting live translation still, which is kind of interesting to me. But anyway, so speaking of use cases, yeah, one of the things that really came up repeatedly, both on the last call that we had and in the past, was that if we had some focus on use cases, that might help move the discussion forward. I think as chairs, we're conscious that that can be a tool that can help us. It can also be a tool that can be very distracting and make the group faceplant, so we want to handle it delicately.
I've just created a new wiki page that we can start submitting those use cases to. I split it up for the time being into three categories: training, use, and presentation. I would ask that folks concentrate on use cases for the expression of preferences because that's what the working group is chartered to do. I understand that there are other use cases that might intersect with the interests in this group, but we need to focus on what we're chartered to do. And hopefully, we can use that to kind of at least have less misunderstanding or talking at cross-purposes when we have discussions as a working group. That's to me the real value of having use cases that we can collect. So we'll send a message to the mailing list about that to get that kicked off and moving along.
And again, our goal, I would present this to the screen, but I can't do that right now. We have a working group interim for two and a half, three days scheduled in, it's three days, isn't it? Three full days scheduled in April in Toronto. We really want this to be a productive meeting, and certainly we can try to hone the stuff that remains in the drafts around training and other terms so that we can make that closer to being ready to be shipped. We would also like, however, to have other proposals on the table for the other things that people want, if they want other things, that have a chance of gaining consensus. And that means you need to incorporate all the discussions we've had to date and maybe think outside the box a little bit. I would say from what I've observed, we've had a number of times people bring up this idea of purpose-based preferences or display-based preferences, things along the lines of how the data is used rather than the mere fact that AI is involved. Personally, I would encourage folks to consider that carefully because I think that seems like the most promising path we have forward if we want anything beyond a basic training signal. Suresh, anything to add there?
Suresh Krishnan: No, I think that makes a lot of sense. And one thing we kind of talked about in the interim and we want to emphasize here too: so like try to bring up the issues like as early as you can so we can have some substantive discussion before we go into the interim and focus the high bandwidth time on trying to resolve those things because issues come up during the meeting, we may not have enough time to soak on it, and we don't want to go on the cycle again where we wait for the next interim meeting or a face-to-face meeting to make progress. So please, thank you very much in advance for bringing up issues before. Thanks.
Mark Nottingham: All right. And to be very clear, that's not saying that if you don't bring something in time for this April meeting that it's, you know, can't happen. It's just that if we want that meeting to be productive, we can't have, you know, it's not helpful to have something pop up on the mailing list the night before. People don't have time to evaluate it and to talk to people about it. So if you are going to make a proposal, please make it substantially before the meeting if at all possible. And I do think that at that meeting, one of the things we'll do is step back and say, okay, where are we at in this work? We've been doing it for more than a year. What do we think is achievable? You know, how do we get this to some sort of success as a group?
Suresh Krishnan: Yep. And I think like one of the things that has happened is like we always get like new participants coming in with new perspectives and that's awesome. And I think this use cases would be like really good building blocks when people talk about like what are the things they want to do and then we can do a cross-check on whether that can be achieved with the tools that we have. I think that could be probably the highest impact thing we can do before the April interim.
Mark Nottingham: Right. So I think that's our summary of where we're at. I think we can just open it to any comments, questions, discussion if folks have particular ideas they want to bring up, feedback on where we're at, happy to hear it. Otherwise, I think we're done. And especially anybody remote, who apparently can still hear us.
Go ahead, Benjamin.
Benjamin Curtis: Hi, Ben Curtis. So I have a question related to the charter and the specifics of AI Prefs as it relates to initiatives that go beyond just AI preferences but contain AI preferences. So I've been working on an IEEE standard around user preferences as a whole, such as for placing cookie banners, terms of service, etc. And a large part of that is AI preference: can my data be used for any AI period? As part of that standard agreement, it's something we've been talking to UK government about, WordPress, etc. So it's starting to get traction. We have considered coming to IETF as a way to build a protocol around that. Is that given that a large portion of it is AI preference but much of it is not as well because it's more generalized, is that something that the AI Pref group might be interested in, or is that completely out of scope?
Mark Nottingham: So it's an interesting question. I think there are aspects of what you're describing that could fit into our vocabulary. What we're running up against is because we're chartered in a way where the vocabulary is separate from the attachment mechanism. You know, in our terms, I could imagine what you're talking about, you know, there are aspects of the vocabulary we can reuse and then you might talk about how do you attach that vocabulary to particular pieces of content. We focus, we tend to focus on content hosting, but that's not exclusive in our charter. We also have folks in the working group who are interested in, you know, just attaching it to content that then can be sent to other people or whatever.
But part of the challenge there is that because there's that separation between the vocabulary and the attachment, it the unintended effects of expressing preferences can mount up a lot of concerns, especially around open web, as people have been calling it, where if you allow people to say I don't want this data to be used with AI, then that precludes things like assistive technologies or you know, and then there's a user agency aspect of that where if I have data that I am using, I you know, want to summarize it for example. Should the, you know, even if the person did license that data to me or has the, still the ownership of that data, what's the balance of equities there in terms of whether I should be able to use AI tools on it? This is why there's this focus on presentation, for example, or display because that's a little bit of a narrower use case. And that is still a very active discussion. I think that, you know, in one perspective, people are very concerned about the effects on the open web and about user rights. In the other perspective, it's you know, we're not a legal framework, we're just a preferences framework. So we're not telling people you can't do that, but it could be that some implementations will still you know, apply those preferences in perhaps inappropriate ways. So I guess, you know, what I'm saying here is that this is a bit of a minefield. It's it's a difficult area for the working group. So if you do want to come and talk, that's great, but be aware that it's not a it's not an easy area for us.
Benjamin Curtis: Yeah, sorry, just quick follow-up. That makes total sense. This is coming out of customer common, so a lot of it is legal based as well as the preferences on top of that. I ask this because given the potential timing of that, I I could drive to Toronto. So would this be something that would be of interest for just a review or putting on the mailing list to see what someone would we do have a draft actually already?
Mark Nottingham: At this point, I think we are trying to to scope down our focus to you know, minimal viable product and then see if we can build past that. So I wouldn't want to lead you into thinking that I know. Totally fine. That's what I... but but please, I think your input would be valuable.
Suresh Krishnan: Yeah, like throw it on the mailing list. I think it's always good to understand like what you're proposing. Yeah, I can put it in my... yeah. So I think like throw it up and we can see what it is. But like Mark said, like we're trying to finish off like a MVP at this point, but yeah.
Mark Nottingham: Anyone else? Or anyone remote? The network is back, yeah. You come on... ah. Okay. So it was relaying Miria. She was back on the network, but she had to re-enter the credentials. So in case if anybody else cannot get on. Right. I had to refresh. Right. We'll give the remote folks just another minute just to make sure that... oh, they've been on the whole time. Okay. We have not. Okay. I'm on, thankfully. So this is what is called diversity in connectivity.
So Martin Thompson is typing. Let's see what he has to say. Although his user indicate no act not active in the last year. Wow. There's some mismatch state. Huh. Martin, did you want to say something? The room is back on the network if you want to say something. Ah, here he is.
Martin Thompson: I'd certainly be interested in what Benjamin is doing, though I can't promise that I'd be able to give that positive a positive reception. Attempts to provide user-level controls is fraught.
Mark Nottingham: Yes, fraught is a good word here. Yeah. But but interesting. All right. So if we don't have any other comment, I think we can close early and give people their time back. Going... last chance.
Suresh Krishnan: Yeah, thank you. And like for if you cannot make it to Toronto, the meeting is like a real full interim meeting, so please join on Meetecho just like this and you can like have all the fun without driving there. Thank you.
Mark Nottingham: Lovely. Right. Thank you all. Leaving short. Yeah, thank you. We can have an early lunch. Yeah. And thanks Michael and Laura for the notes. Yes, thank you very much.