Session Date/Time: 21 Oct 2026 20:00
[00:00:20] Mark Nottingham: K. Hello, everyone. It is 07:00 in the morning here. I think it is other times for some other people. It looks like people are still filtering in, so we might just give it two more minutes to see who else shows up before we get started. So just hold on just for a moment, please. Okay. I think, we can go ahead and get started. Hello, Suresh. Good morning, everyone. Let's go ahead and get started. Let's dive right into it. Get the preliminaries out of the way. This is the NoteWell. If you're not familiar with, this, this is the terms we participate in the ITF with or terms we participate under. You can use your favorite search facility to find this by searching for IETF note well. Kathy, could we ask, if you're not presenting or talking, to to turn your camera off there? Just helps save bandwidth for people. Thanks. So please do familiarize yourself with this, if you're not already. Anyone willing to volunteer, to take minutes for this meeting? I think we're mostly interested in the very high points of, any, kind of progress made or or, high level arguments that we wanna take back to the list. Anyone?
[00:03:35] Chris Needham: Yeah. This is... Yeah. Chris here. I'm I'm happy to to have a go.
[00:03:39] Mark Nottingham: Chris, you're you're you're the best. So there there is the pad page, linked in the agenda that you can use. Or if if you can't find that easily, just take notes on your own and send them to Josh and I when when the meeting's over, if you could.
[00:03:54] Chris Needham: Thank you. Yeah. I'll probably do the latter. Yeah.
[00:03:56] Mark Nottingham: That's absolutely fine. Just just, making sure capturing the major points of discussion is is all that really is necessary at this point. Okay. So let's get into it then. Thanks, Chris. This is... This meeting really is just a a a catch up, a a checkpoint, before we have the San Francisco meeting. We wanted to keep discussions going, see where we're at, and especially if if folks wanna bring materials, to the San Francisco meeting to to give them a chance to, sound out the group or find collaborators or, you know, get those arguments going. So as always, I think, we'll we'll start with the issues list and and a quick summary of where we're at. You know, as as you saw, we have the most recent drafts, published now quite a few weeks ago, I think. Actually, let me go to the data tracker for these. So this is the most recent vocabulary document, zero eight. It incorporates a lot of our discussions, that we had at our last meeting. I I... We have not... And then we're we're much more explicit about this now. As you can see here, we're not saying that this has consensus. This is just the current topic of discussion for the group. However, we do want to start moving to consensus, fairly soon, I think. I think we're getting to that stage of the work. So what I thought we'd do today is run through the issues list, see where we're at with those, see if people had comments on the drafts. Kevin, you had something?
[00:05:48] Kevin Kelley: Sorry. No. I didn't didn't mean to interrupt. Christina and I do have a PR that we opened, that might be an interesting topic for discussion too related to some of these issues.
[00:05:58] Mark Nottingham: Okay. Why don't we get to the first one? Why don't you bring it up then? That makes sense? Okay. Great. And and, yeah, just... If if you could keep your camera off unless you're speaking, that would... I think that would help. Thanks. That's just for whatever it's worth. That's the norm, for the idea of meetings. No worries. So going through, first of all, on the issues list, we have a couple of these that we think are ready to close, and we wanted to confirm that. The first being this, issue two zero eight, use the term nonsubstantive. So this was discussed in in Toronto, and indeed, in any form doesn't appear in the editor's draft. I think we're okay to close this. Any thoughts about that, or can we go ahead and do that? Okay. Close that. Next up, one seventy two. This is the category, for rag and grinding brought up by Chris originally in the email. We we do have the usage category now. So, undoubtedly, we need to discuss it more and and refine it, but we do have that there. So I think we can go ahead and close this. Any thoughts about that? Okay. Sorry. It's hard to track all the different sources of information now at the same time. So closing that. And then finally, very similar. This was from Bradley Silver. Is Bradley here? Yes. He is. Category for substantive use. This was discussed a long time ago, and I think the the the direction we've gone in is that use category. So so, Brad, are you happy to close this for now?
[00:08:11] Bradley Silver: Yes. You can close it. Great.
[00:08:14] Mark Nottingham: Thank you. Alright. So that puts us at 12 open issues. We've deferred a few of the issues around defaults, unknown and allowed, and Krishna's display based preferences vocabulary. In our last meeting, we said we wanted to keep that open and come back to it eventually once the rest of the draft was more fully formed to see if there's anything we could extract from it. Ecker, these other two issues are from you. Any progress or any further thoughts about those, or we still just gonna keep those deferred for the time being?
[00:09:05] Eric Rescorla: Sorry. I think I'm... The screen is too small on my screen to see which one we're referring to.
[00:09:10] Mark Nottingham: Oh, sorry. Let me, is that a little bit better? Oh, Apple. No. There we go. How's that?
[00:09:22] Eric Rescorla: So I still didn't see which ones... I still didn't see which ones you're referring to. I'm sorry.
[00:09:26] Mark Nottingham: So this is a 153 and 154, the difference between unknown and allowed and also defaults.
[00:09:34] Eric Rescorla: I think I need to reread those, so I'm not sure. I'm sure if further closing that.
[00:09:37] Mark Nottingham: That's fine. If if you could just I'm gonna go from oldest to newest, I think. First up, 163. Is Joe with us? I don't see Joe. This is around the the asset terminology, which I think has been, contentious for a while. Most recently, we had a proposal for using the term artifact. I think this is almost all editorial, but the working group does feel strongly about it. So I wanted to bring it back up here. Martin?
[00:10:31] Martin Thomson: Yeah. As as Paul says, this is one of those ones that we've been sitting on for a while, and the the decision at the time was to sort of wait for us to to get further through the rest of the discussion and and then settle on something. But I have a pull request up that addresses this and uses the word artifact. It was a good opportunity to go through the the vocabulary draft and the attachment draft and make sure that the terminology was harmonized. The the vocabulary draft was pretty good, but the the other one needed a little tweaking around its use of the word content, which was a little bit tricky. But otherwise, there's something there that people can look at. I... Like Paul says, I don't think it's quite time to to merge that in until we're settled on on some of the other issues.
[00:11:33] Mark Nottingham: Oh, okay. I... I'm a little confused by why we need to wait there, but sure. Timid Robot?
[00:11:42] Timid Zehta: Hi. Timid Robot with Creative Commons. Martin, I was just wondering why your poor request didn't update the, definition to use, artifact. It still says asset. Is that
[00:11:56] Martin Thomson: intentional mistake on my end. Pure mistake. Cool. Thank you. Thanks.
[00:12:02] Mark Nottingham: Okay. And and I'm
[00:12:05] Martin Thomson: happy to merge this if people are happy with the terminology. I was not looking to disrupt existing stuff that's in flight.
[00:12:13] Mark Nottingham: Mhmm.
[00:12:13] Martin Thomson: That, you know, we've got pull requests open that use the word, and we're talking about them. And so if we start using different words, people might start to get confused. I I don't know.
[00:12:24] Mark Nottingham: I hear you. I'm also... You know, in terms of of cutting down our issues list, and if we're gonna use new terminology, giving folks some time to get used to it, and and make sure they're happy with it, might be good. I I... If if... I... I'm not hearing anyone pushing back. If if if anybody would like to, please get in queue or or maybe send a mail to the list. But otherwise, I'd say, you know, editors just find the most reasonable time to merge and get it in place so that people can start to get some comfort with it when when it's ready. That's annoying. So next, distinguishing between asset and use. I think this one, again, we've discussed a few times, mostly in the periphery of other issues. And, I I believe the group is is is largely agreed that, you know... And and indeed, if you look at our charter, it's it's it's reinforcing this that the preferences we're stating about are about the use of data and not access to it. And and so, Martin, you were working on this. Any progress report?
[00:13:58] Martin Thomson: Yeah. So there there is a pull request up here.
[00:14:01] Mark Nottingham: K.
[00:14:02] Martin Thomson: It's three paragraphs. Think it's a section 3.3, which is fun. Section three has given us so much joy. I'll give people a chance to read that. Maybe you can make it a little bigger, and we'll be good. Yeah.
[00:14:33] Mark Nottingham: Let's see. Where's the
[00:14:36] Martin Thomson: That's been up since the last meeting.
[00:14:37] Mark Nottingham: Make it bigger. Yeah. I I... The only quibble I think I'd have with it at this point is, the word prevent is maybe... Oh, it's it's... Well... Okay. Encryption... In some of those cases, prevent is too strong, but that... That's a quibble at best, I think.
[00:14:55] Martin Thomson: The the word constrain maybe might might be a better way
[00:14:59] Mark Nottingham: to Sure.
[00:15:01] Martin Thomson: To do that or restrict or something like that. But, yeah, happy to happy to tweak that.
[00:15:08] Mark Nottingham: Edgar?
[00:15:11] Eric Rescorla: Yeah. I think this text is generally sound. If it were me, I'd probably remove the technical met control mechanisms, like encryption access access control. And my reason for that is that, really, what we're doing is contrasting this to, like, the disallowing robots.txt. And there are, like, lots of mechanism. I've heard some from from seeing it. I mean, I'm not gonna fall down on on the street. If people are like, well, if that's important, I I think it's fine. But if it were me editing this text, I I would really focus on the on the the question that people, I think, routinely have, which is like, you know, which is, like, how does robot... How this correspond, like, things like robots?
[00:15:46] Mark Nottingham: Okay. And Martin says he's happy to type in
[00:15:50] Eric Rescorla: And I'll add I'll add a comment in the text just just to to... Jure so Martin doesn't forget it. But, course,
[00:15:54] Mark Nottingham: it's In in PR? Lovely. Thank you. Any other comment on this issue or this PR? Kevin.
[00:16:04] Suresh Krishnan: I think Kevin is in there. Yeah.
[00:16:07] Kevin Kelley: I I agree with that group very much. I think that the contrast between traffic management directives and expressions of preference about how the resulting data is used, I think that's exactly what we want to say here. Right?
[00:16:20] Mark Nottingham: Okay. Sounds good. Anyone else?
[00:16:31] Martin Thomson: Okay.
[00:16:37] Mark Nottingham: Let's move on then. 192. Is Max with us? I don't see Max. Anyone have any further thoughts about this, or should we skip to the next one? This is, one ninety two, introduction to, the relation to existing laws. I think my understanding is we need to get a little clarity about Max. We talked about this a little bit at the last meeting, and we need, Max to come back perhaps with a revised proposal, if he still wants to pursue this. Okay. Moving on, two seventeen, use of synthetic and AI definitions. Oh, apple. There. So we went back and forth on this, a fair amount at the last meeting, with the use of the word synthetic in the document. I believe it is still there. Yes. So in these two instances right here, I'll make that bigger in the definition of generative AI model. Any thoughts? Chris?
[00:18:11] Chris Needham: Thank thank you. Sorry. It's kind of difficult to to note take and to keep across the issues. So my my worry with... Let me just review where we got to. Yeah. My my worry with sort of leaving the word synthetic in there is that it could be interpreted to make a distinction between things that are sort of generated by the AI system or, you know, perhaps could be quoted verbatim by the system where, you know, both outputs are kind of problematic, but one is perhaps more of it in synthetic in nature than the other. And it's like it's all of the AI output is like, we would want to sort of treat in the same way. So I I I worry that we open a distinction that isn't helpful to us by by using that word.
[00:19:12] Kevin Kelley: That there Yeah. Yeah.
[00:19:13] Chris Needham: Because there there may be parts of the output that that actually, you know, may not be considered synthetic.
[00:19:19] Mark Nottingham: Right. I I think, from remembering the conversation before, part of the concern was if we're gonna use the word, it it should be intentional. And and the question is what work is synthetic do... Doing that that generate is not doing, I think.
[00:19:32] Chris Needham: Yeah. Indeed. So... Yeah.
[00:19:33] Mark Nottingham: Yeah. Yeah. Kevin.
[00:19:37] Kevin Kelley: Yeah. Sorry. I'm trying to remember exactly how the last iteration of this conversation went to. I think where we got to was... Like, the word synthetic is not... The definition does not fall apart without it, but it is evocative of the sort of substitute of use idea that that, we just closed Brad's helpful issue from way back concerning, but that is still kind of the thing we're trying to describe. Right? And I also remember that there was a attention of that conversation in... At one of the in person interims where we talked about what a model that exclusively repeated training data back would look like. And that's sort of... Doesn't seem like a real world example of of anything that that we're trying to describe here. So, like, whether or not the model recites some particular phrase is different than whether the model is being used to generate synthetic content, if that makes sense.
[00:20:51] Mark Nottingham: K. Timid Robot.
[00:21:02] Timid Zehta: Hello? My name is Timid Robot. My understanding is also that it wasn't about memorization, but it was about trying to create a line between, like, classifier type mod... Classifier type uses and kind of generative... More generative, like... Where the the content is modified, you know, or, you know, complete... You know, it's it's completely new, not cited. And I think that we do need to try to make that line a little bit clearer. And so I think I like the presence of synthetic more than I like its absence because it... You know, I don't think it is successful at drawing that line.
[00:22:06] Mark Nottingham: But... Yeah.
[00:22:10] Kevin Kelley: Okay.
[00:22:16] Eric Rescorla: Yeah. So, I mean, the sentence below explicitly says classifiers blah blah blah. So I think that that's actually much more, as as Claude would say, load bearing than than the the word synthetic. So, I mean, I I... I'm a little worried about, like, you know, this... The... Like, the fact that we're sort of still arguing out what the word synthetic means in this in this cat... In this category makes me really uncomfortable because we're now expecting someone to read this without any of the benefit of the contest we have and understand what it is. So I think what would be really helpful would be if, like, the people who feel strongly about this one or the other could, like, try to line up all the things which they think are in scope with and without the word synthetic. And then, like, if they can figure out which things they are, then may... Then maybe those things are the same, in which case we don't need the word synthetic, or maybe they're not the same, in which case let's just decide on what those things were... On, like, how to, you know, like... And and and and so I I think there are, like, four possibilities. One is that those things are... That everyone agrees that those two things are different or they're the same. And the other is everyone agrees which one should be in scope or they don't. In the cases where... And once we get to that point, we'll know what we have to do. And in particular, if there's disagreement about which things should be in scope, which things not be in scope, you have to resolve those things technically, there's no disagreement that we just have figure out how to craft some text that electros a lot in the right place. So I think that would be my suggestion. I mean, I guess if I were voting, I'd try if there was a, but I don't I don't... If we think it's a low bearing, I'm kinda comfortable with that.
[00:23:41] Mark Nottingham: Okay. So sorry. That last bit, you're... If if if
[00:23:45] Eric Rescorla: I mean, if if all we're doing is, like, voting on the word synthetic or not, like, I would keep it, but, like, I don't like that it's load bearing. If people think it means anything, I think we need to work out whatever it means.
[00:23:55] Mark Nottingham: I I was wondering whether it might be useful, as a thought experiment for folks who who want to keep, synthetic, for example, to go off and create a definition for the term, so we can see what they're they're loading into it. And and and that might help tease out some of these issues because I agree with you, we're we're loading a lot into one word. Timid Robot, I think that's an old hand. So, Brad.
[00:24:21] Bradley Silver: Yeah. I just wanted to weigh in to agree with Chris. I also have the same concerns as Edgar. I kinda feel like we're we're... The the the main rationale, which I think Timid Robot, you know, got right, is already covered. You know, I think that there has to be some kind of general trust around what the line between the generative and non generative AI model is. You know, if you're just saying that everything... Every other... Every model that's basically, you know, not mute in in what it does and never spits anything out, must be a generative AI model. I think this doesn't fit in with with with common conceptions of what a generative AI model is. So I I I I worry about keeping the word synthetic. I think your solution or your suggestion going forward makes sense. And... But, you know, if we were to be voting on that right now, I would be voting to delete it. I think it stands up well enough without it.
[00:25:30] Mark Nottingham: K. Nate.
[00:25:32] Nate Hake: So, I also think that we should, delete it. I... This isn't the strongest point for me, but I I would point out that in the copyright context, when there have been debates about where the... You know, what the sort providence is of AI generated content, many AI companies have taken the legal position that AI generated content is, quote, unquote, original content, which might be interpreted to be something other than synthetic content. And so I just think, for me, there's some concern here that those same arguments might be used to try to, like, get what we clearly intend to be AI generated content not covered by this. And for that... And I... In in absence evidence that this does some actual load bearing work, I I think we should cut it.
[00:26:22] Mark Nottingham: Cap.
[00:26:25] Kevin Kelley: I just wanted to say that, like, the the kind of output that Nate is talking about is exactly what I intend for this definition to cover. Like, we're we're not trying to pull a gotcha move on anybody. And I think to Brad's point, like, extending each other a little bit of trust about that intention is is going to be helpful here so that we don't wind up splitting hairs that don't materially impact the outcome.
[00:26:56] Mark Nottingham: Nick?
[00:26:57] Nate Hake: I... Kevin, I I don't doubt your motivations at all, and that's not what I I was suggesting. Some of these companies are not represented in this group that I'm talking about, And so we have to come up with a text that works, you know, without trust and in in with... In the context of, you know, relationships and disputes with people who are not part of this working group. And so it's... That that that is is my point is is that that I think that this argument has been made, including by companies who do not have representatives here. And so that's what concerns
[00:27:34] Mark Nottingham: I I think I'd I'd go back to Edgar's point that, I believe it was Edgar that that, we need to communicate clearly enough that someone who hasn't been in these conversations can understand what we're doing. And Edgar, go ahead.
[00:27:48] Eric Rescorla: Yeah. That was more or less what I was gonna say. I I don't think there's an issue of trust. I think it's an issue of, like, clarity of people who not read the legislative history. And, and so we want people to be able to read this document, understand what they have to do without, like, going and reading the archives of this mailing list, you know, long after, you know, long after we've retired. And so I think that's why it's important to be clear on what this says. I don't think it's an issue of whether or not I have any personal feelings that one might have about other people in this group. It's really a matter of, like, this document has to stand on its own two feet.
[00:28:15] Mark Nottingham: So, obviously, we're not gonna close this today. I think, if if folks could volunteer or start thinking about, you know, doing what was said, you know, trying to tease out the issues here, write down, what... What's being implied by this and what the different outcomes might be, that'd be really helpful. I don't know if anybody's willing to put their hand up at this point. Timid robot.
[00:28:47] Timid Zehta: I'm not... Sorry. Timid robot. I... Unfortunately, Mark, I'm not, offering to take this on. I'm just based on what I'm hearing, I feel like it would be better to remove it and give people the opportunity to make a strong case to add it back in versus continuing asking questions about it. Because most of what I heard, the people who advocated for it were not nearly as invested as the people who want it removed. And so that might that might just shorten future discussions, because I suspect that no one feels strongly enough to, rally around it.
[00:29:35] Mark Nottingham: If if folks have responses to that proposal, go ahead and get in queue. That that sounds like one reasonable way forward. Sure. Chris?
[00:29:45] Chris Needham: Yeah. I I don't know that I could do this, like, ahead of the San Francisco meeting, but I'm certainly happy to discuss with anybody sort of offline to to try to come to a conclusion on it. But I I actually kind of favor what Timid has just said as a as a way forward. I think that would make sense combined with, you know, what you said earlier, Mark, about, you know, putting together a definition.
[00:30:15] Mark Nottingham: Paul?
[00:30:17] Kevin Kelley: Yeah. Very similar to what the two previous speakers said. Like, it seems to me that it's either remove or sort of provide a definition that's consensual. So, like, I think we have... In the absence of somebody coming up with a definition that works for the group, like, it feels like that this discussion moved towards rather removing it for now.
[00:30:39] Mark Nottingham: K. And Matt?
[00:30:44] Matthew Hervey: Yeah. Matthew Hervey, Cloudflare for personal views. Yeah. For what it's worth, I totally agree with Timid that it's better to remove this pending some strong clear reason to to reinstate it. And to Eckhart's point, it should be clear to someone coming to this document, I don't think even looking at the mailing list, they would understand, what we meant.
[00:31:07] Mark Nottingham: Okay. I'm I'm gonna... I think Eckhart mentioned a poll. I I think it might be interesting just to have one poll. We... We've heard, some people, say, they have a fairly strong preference if I'm interpreting it, correctly to, remove this. I've heard people arguing to keep it, but I'm not quite sure of how strong that is. And so I might... Where is the poll tool?
[00:31:38] Suresh Krishnan: Mark, do you want me to start something?
[00:31:40] Mark Nottingham: I can I'm I'm If
[00:31:41] Suresh Krishnan: you unfound it. Okay. Yeah.
[00:31:42] Mark Nottingham: So I'm gonna pull... I think I'll pull, do you feel strongly that we should keep the synthetic term in the draft?
[00:31:56] Eric Rescorla: I just wanna flag that Farzaneh the chat has asked you this during ITF. I don't know if it means anything you, Mark. I just wanted to flag it just in case you didn't see that.
[00:32:03] Mark Nottingham: I I did not see that. What... Sorry. Farz, do wanna get in queue?
[00:32:08] Farzaneh: Yeah. Sorry. I didn't raise my hand. Sorry. I just wanted to know if... Because we we don't have enough background and there are not many people here that... Can we can we defer this to IETF meeting?
[00:32:28] Mark Nottingham: Well, so this is a a a meeting. We're we're not gonna make a decision today. We're just trying to gather information and get the discussion further down. So
[00:32:38] Farzaneh: I I don't think we can gather information when people don't have enough background. Why isn't it synthetic there, and they're not even here? Some of them could not attend.
[00:32:49] Mark Nottingham: Okay. The issue's been open for quite some time, and we discussed it pretty extensively before. So, like I said, we're we're not gonna close the end today. We're just trying to to get us... Ourselves in a position where we understand the issue. And and to do that, we have to discuss the issue. So I'll just hit a poll. I'm just mostly interested, in in hearing that other side. Do you strongly feel that we should keep the synthetic term in the draft? And I'm seeing some come in. Okay. Leave it open for just another moment. If you can't find the poll tool... Actually, while we do that poll, you're in queue. Did you wanna say something?
[00:33:43] Suresh Krishnan: I think it's a lingering queue thing from
[00:33:45] Mark Nottingham: Oh, okay.
[00:33:45] Kevin Kelley: That was that was me trying to find the poll tool. Okay.
[00:33:58] Mark Nottingham: Okay. I'll go ahead and close the poll in just another few seconds. So right now, we're at, three saying yes, 14 saying no, one with no opinion, and then several more implicitly with no opinion or or who cannot find the poll tool. So three fifteen one. So if if any of the folks who who do feel strongly, no obligation now, and if if no one says anything now, it doesn't mean anything. It's just... It it wanted to give folks an opportunity to to explain a bit more why they feel strongly, that we should keep it. Or or if folks wanted to just at least identify themselves so they can coordinate on, you know, putting together a case more, now's that opportunity. But, of course, it's not the last one. Okay. I don't see any hands.
[00:34:50] Suresh Krishnan: So, Mark, I just wanted to add one thing. So whoever wants to keep that term probably needs to have a definition, try to write a definition of it so we can kinda figure out, like, what falls in scope or not. I think that's probably the additional detail I would like to add on that. So What
[00:35:04] Mark Nottingham: what I'm hearing is this that on on its own, the word synthetic is is ambiguous enough to to cause issues for a lot of people. And so, having a definition of it, if we wanna keep it, would would help. Or or if people wanna make the argument that it doesn't add anything and it should be removed, having that flesh... That argument fleshed out, would be helpful in San Francisco, for example. So I think, you know, if if people can kinda put together, some materials around that, that'd be really helpful to move the discussion forward when we get to San Francisco. Okay. So moving on, issue two thirty one, de minimis content generation. Click. No. Oh, Apple. There we go. So, Erin, this was your issue. And I think this was... If I remember, this was an idea that you had during the meeting, and I'm I'm wondering what, in your mind, the status is now on this issue.
[00:36:15] Erin Simon: That's an excellent question. I haven't checked in on it in a while. I believe it has been kind of folded into the drafts, but I'm not confident enough to say that without going through everything.
[00:36:30] Martin Thomson: No. No worries. So So if the last comment here... I was gonna say the the last comment here is where I flag how I believe this is covered in the draft. So if you scroll to the the bottom, and for Erin's benefit, if you can't read that, it it says the definition doesn't include a model that used for the classification even where the model generates rationale for its output. Now we might wanna tweak that language in some ways, but I think that's I think that's addressing this point.
[00:37:01] Erin Simon: Yeah. I I think the concept is good. Maybe one question is, does the issue operate at a higher level than just that one definition? Because there are a few places in the current draft where I feel like we've gotten good wording in one definition that we should uplevel to the whole the whole vocab. But maybe that's itself a set... Like, if you wanna close this one out but preserve that concept somewhere else, I'm fine too.
[00:37:32] Martin Thomson: Yeah. So this is in the definition of generative AI model, just to just to be clear. So it should apply to to other users, but we'd have to be extraordinarily careful as usual.
[00:37:46] Mark Nottingham: Yeah. Yeah. Okay. So, Erin, I'm gonna mark this as proposed to close. And if you wanna go back and and take another look, maybe make yourself comfortable, then we'll come back to it later.
[00:37:59] Eric Rescorla: Aye aye.
[00:38:00] Mark Nottingham: Thank thank you. Okay. Always be closing issues. Next up, two forty five, Timid Robot, concern about redistribution of models. So this was, one of the threads that started, well, not started, but what we we heard about... A bit about during the last meeting. I see there's been some activity in the background. Timid robot, Martin, where do we think this is at? Kevin?
[00:38:41] Kevin Kelley: I'm sorry. I was just going
[00:38:44] Mark Nottingham: to say, think this
[00:38:44] Kevin Kelley: is another one where we agree about the intent, which is redistributing a model or anything of that flavor should not be a way to launder preferences. It's just how do we... How does it fit into the design kind of, and how do we phrase it?
[00:39:03] Mark Nottingham: Yep. That that sounds like a good summary. Martin?
[00:39:11] Martin Thomson: Yeah. So I think there's two issues that address this one or two pull requests even that that address this one. So Timid Robot opened one, and then it was subsequently closed because I think we we moved the discussion on. But there's two fifty three, addresses an issue that we'll talk about later on. But two fifty three also includes text that could address this one depending on your perspective here. So it's a bit oblique, though. It doesn't doesn't directly do that. Where was it? Yeah. That's that's the that's the one that I think covers that, but I I don't know if it's direct. Hence, the no linking.
[00:40:21] Mark Nottingham: And you also had two forty five. Is that relevant?
[00:40:27] Martin Thomson: Yeah. Let's look at that. That's a commit. So... Oh, you just opened 245, I think.
[00:40:34] Mark Nottingham: I did. I did. I did. You're right. Add commit. Yes. There we go.
[00:40:49] Martin Thomson: Alright. So that... I I remember now. We have this issue open, but we have a section that addresses this in the current draft, which we have specifically flagged with this issue. And so the text was indeed the text that Timid Robot proposed, and that's why I closed that pull request. We got a little confused trying to get a a version of the draft going, but it basically says here the compliance with preferences regarding the use of assets in training of models also includes conveying the preferences associated with those assets along with the model so that we're using the model to make a decision about how they then comply with preferences.
[00:41:35] Timid Zehta: I'm going to interject and say that that's old text. I think we should look at vocab draft eight, section 3.3, which is updated.
[00:42:12] Mark Nottingham: That seems pretty pretty straightforward, actually.
[00:42:20] Timid Zehta: So I just raised my hand to say that the text in 3.3 satisfies me. And if that's all that's needed to close 245, happy to have it closed.
[00:42:34] Mark Nottingham: Okay. Okay.
[00:42:40] Eric Rescorla: Yeah. I guess I can live with this text, but I I I do think it creates a weird asymmetry in that if I go and I scrape a giant pile of sites and I just copy the data and I just let let people have it, I don't have to do anything here. I don't have copy any preferences. But yet if I do any AI model training, then I have to copy the preferences. And it's a very weird asymmetry. And so I I guess, you know, I I guess freedom is good. So, you know, so so if only half the people in the world have to play something, I guess it'd be better than none. But, you know, but I think it creates a pretty odd asymmetry.
[00:43:14] Mark Nottingham: So I guess if if I'm listening to what you're saying, it's it's that we place, conformance requirements on, people creating models, but not on people doing the crawling.
[00:43:28] Eric Rescorla: Correct.
[00:43:30] Mark Nottingham: And I I think if I were to characterize this, it would be that we we we do so by saying, you know, if you're training AI models, willful blindness is not, you know, conformance effectively is the approach that's being taken here.
[00:43:45] Eric Rescorla: Well, but, I mean, you know, I guess maybe. I mean, I think, you know, if if a... You know, I I I think I made this point in the list. But if if a if a scrapes the site and then distributes the data and b, you know, and and b trains on it, then b doesn't have to have any any interaction with the, with with with the with those... With the preferences at all because they can simply align them. Whereas if a trains the site, scrapes the data, distributes the model, then then, say, the the same output output output chain, which is a... Which is an open source model. And yet in one case, like in one case, like, you know, you had to buy the Hermes in there even.
[00:44:21] Mark Nottingham: Right. No. But I I think that's the... If if if if a provides a service, I'm gonna scrape sites and align the, you know, preferences, then then that that should not... I I I think I agree with you. That should not protect b.
[00:44:37] Eric Rescorla: But but but if you notice that that that, in fact, is the text that we have here. Right? It says we have here, like like, that there's there's no there's no requirement at all that, like, a, keep the references. And so I... I mean, again, like, I guess, you know, maybe I'm not gonna I'm not gonna... I I guess I can live with this, but, I think it's an odd odd.
[00:44:58] Mark Nottingham: If I if I play back what you're saying, I'm hearing is is that paragraph there under the notes, places compliance... Conformance requirements on parties other than the other targets of our specification. It it... You know, it's it's like compliance preferences we're gonna use in training unit models also includes conveying the preferences associated with those assets, which applies to... If I train models, but if I don't train models, then that's a Yeah.
[00:45:23] Eric Rescorla: I know. It does. I... That's how I... Sorry. This is actually... Yeah. I think this text is clear. Okay. And I guess if other people... If no one else, like, thinks this text is funny, then I'm I'm gonna be quiet.
[00:45:31] Mark Nottingham: I I I I do see what you're saying. Okay. Thank you. Martin?
[00:45:38] Martin Thomson: That was... Kevin was ahead of me. Oh. So I I think I... Because... Well, it disappeared. So, obviously, we've moved on. That's fine. So I just wanted to say in response to to echo there, I think this is a reasonable concern. Yeah. Kevin, I wasn't paying attention. The the concern that echo raises is a reasonable one. I did open two fifty two and two fifty three to address a different point, But as part of writing that text, I added language that would address the retention of preferences with, say, crawling or scraping or crawling or scraping and then transformation into an embedding, which I think was a concern that Leonard and Nate have both raised in the past and we may have later on the issue list to discuss. So if symmetry is what you desire, then then those two pull requests would address that concern. And maybe not in the direction you prefer, but I'd encourage people to have a look at those.
[00:46:49] Mark Nottingham: Martin, could you maybe update the issue with the appropriate links just to make it clear what the current state of things is?
[00:46:58] Kevin Kelley: Brent.
[00:47:00] Bradley Silver: Thanks. Maybe there's not an issue to address here. It sounds like if there's not going to be a change to this language, then I'll be quiet. But I just want to underscore obviously that for us separating the application of use preferences from the act of crawling through to are there any further distribution of what's crawled to the model that's eventually developed it should it should travel. That is that is important. There is no... And I don't like the idea either of of applying concepts like willful blindness and I think now we're starting to get into legal legal constructs, which I don't think fit very well with our with with our document. But, you know, if you're... If we're leaving this language as is, then I don't have an issue. But I'm just... I guess I'm responding somewhat to Echo's concerns.
[00:48:00] Mark Nottingham: Right. I I think the question is, you know, as it's written, it it... It's it's a requirement on a party who isn't doing AI training, which is a little weird in our framework. Just thinking out loud, if I were to to rewrite this, I might write it in terms of, you know, you cannot comply with preferences if you don't have full information about them, you know, if you're using a third party service or or something. Perhaps I should go off and think and and put some of the issue.
[00:48:36] Bradley Silver: Right. And then it's not how the market is currently, you know, evolving right now. There there is substantial activity that is focused specifically on crawling by certain strata parties that is that are not those that are developing or distributing developed models. So I think it would leave pretty big gap if, if we had to require the link between collection and development into a model in the same party that's doing the crawling because that's just not not what's happening now, to publishers.
[00:49:17] Mark Nottingham: K. So it sounds like maybe we need a... And Martin just didn't issue update. That's great. So we'll come back to this undoubtedly in San Francisco. Timid Robot.
[00:49:31] Timid Zehta: Hello, Timid Robot. I just wanted to note that this is... I know that Common Crawl is is already bundling preferences in... With headers and robots when they archive websites, which means that... I know they also help with the Internet Archive, and so I think significant portion of the archiving in... Industry is already behaving in a fashion where preferences are communicated with content. And so I don't think it's I don't think it's abnormal, weird, or onerous. Thank you.
[00:50:12] Mark Nottingham: Thanks. Alright. Let's go ahead and move on then. We're up to number two forty six. This is Leonard's issue. Is Leonard with us? I do not see Leonard. Martin.
[00:50:36] Martin Thomson: So that... This is the issue that I was referring to earlier, which is there's... There was a concern that someone would go out there, obtain a whole bunch of stuff, and then apply some sort of transformation to it, whether that be changing HTML into Markdown or taking whatever it is and producing an embedding of that and then distributing the embedding. I think the obvious thing and, you know, the extend a bit of trust thing to call back to Kevin's point earlier would be that there's an expectation that we would have preferences carry on with such a transformed thing. But the the issue here exists to sort of say, well, maybe we can just be explicit about that. And so 252 and 253 address that. Leonard also has opened 240, which has an alternative spelling of that. So that is the last paragraph of this one that's really directly applicable, which basically says, well, you... If you obtain something at one point and then use it somewhere else, make sure you keep the preferences in all that time. And two fifty three says, that includes if it it gets transformed in that process. And two fifty three has all the text associated with it. So if you can... If you wanna look at two fifty three, because it builds on the other one, you can just expand to include those last two paragraphs.
[00:52:26] Mark Nottingham: So this is all relatively new to the group. Any initial reactions, thoughts, thoughts?
[00:52:35] Martin Thomson: I say relatively new. I introduced this in... At the the in person meeting we had in in London and informally talked about it, but it wasn't on the agenda formally. So it's been around that long.
[00:52:53] Mark Nottingham: Timid Robot.
[00:52:56] Timid Zehta: Hello. Timid Robot, Creative Commons. I like this text. I think that it is better to make it explicit. It, you know, empowers people to honor the the preferences if that is their intention. And I think that it's... I think it's something that will be helpful for people who have that have that goal.
[00:53:30] Suresh Krishnan: Martin, I I prefer the term insure from 252 than this, but it's just me. No hat.
[00:53:39] Martin Thomson: Yeah. Do you wanna comment on the on the pull request? That'd be helpful. Not sure what you want at word.
[00:53:47] Suresh Krishnan: No. It's like... It says, like, also provides an opportunity. The two fifty two said ensure that somebody can apply preferences. But... Yeah. I'll I'll I'll comment on
[00:53:58] Martin Thomson: the other. I I I like that. That if we can we can align the two of them. Yep. Sounds good.
[00:54:40] Mark Nottingham: Can you hear me?
[00:54:45] Bradley Silver: Yes.
[00:54:46] Mark Nottingham: Okay. Thank you. I'm not gonna blame that on any one particular piece of technology, but I'm back. Okay. Kevin.
[00:55:04] Kevin Kelley: So I don't understand what the... What two two fifty three would change. Right? Like, when we talk about models, what we really mean is sort of an AI system. And I don't think I'm spilling any secret beans if I tell you that, like, text gets tokenized and transformed into embeddings, and all of that stuff happens as part of the implementation of that system. This feels like... I mean, it doesn't... I don't have any particular objection to it, but it feels like the sort of thing where trying to place a list of specific elements of an AI system that we want to ensure are covered. Like, other than being not useful is also counterproductive because it suggests that things that we don't think to list now or that haven't been invented yet are not in scope.
[00:56:05] Mark Nottingham: Timid Robot.
[00:56:08] Timid Zehta: So my name is Timid Robot with Creative Commons. Kevin, my understanding was that this was trying to target processes kind of pre... Pretraining, pre tokenization. And I thought that Nate had spoken about already existing markets where there are third parties, like, selling, like, data APIs or something where the data being sold had been transformed from the data, like, crawled or collected. I don't know if Nate can speak further on that. It's not an industry that I'm personally familiar with.
[00:56:58] Mark Nottingham: Nate?
[00:57:03] Nate Hake: Yeah. I don't know. That was, I think, in the context of another, of the AI use definition, although it is true that this relates to that. It... It's this... There are third party APIs that you can use that basically will sell the embeddings themselves. Like, the... They'll they'll they'll have... You know, you can go out and can purchase embeddings that are drawn from materials. And so there's a question as to whether or not... You know, I think it's important that our... Whatever definitions we come up with extend the preferences to those embeddings where they're separate from the actual content itself. I think this is more about sort of if, you know, if if if... Whether or not this is sufficient or not, I'm not a 100% sure, but I I I I think, Tyler had raised some of these concerns. He's he's not here, but that that was the idea. So I think that's that's kind of kind of the idea.
[00:58:04] Kevin Kelley: Then does does this make sense as part of the crawling conversation if that's the objective?
[00:58:17] Nate Hake: I... Kevin, could you say that again? I did not understand that.
[00:58:28] Kevin Kelley: I just said, does this make sense as part of the crawling conversation if if that is the case you're trying to cover?
[00:58:42] Nate Hake: May... Maybe. I'm not sure. I I think I I think I'd need to review it all in whole. I think as long as we're sort of attentive to the reality that that... I mean, this is a very significant marketplace. Like, this is not some edge case. This is this is a really significant thing. I I think that we're kind of tracking in a debt... In in a place that would address this, but I I think I I need to look at the whole thing as whole. So so perhaps... And then part of this is a problem that I'm not as technical as as some of the folks who ideally would have someone with more technical knowledge weighing in here. But... Yeah.
[00:59:27] Mark Nottingham: Okay. So there's still some discussion in chat. Matt?
[00:59:35] Matthew Hervey: Just... I wonder if it addresses Kevin's concern, if this isn't just about occurs at different times but by different parties. Because I think the concern is entities collecting, converting artifacts, we're now saying, and then reselling them at a later date to third parties. Does that help? Is that acceptable to people if it's about third parties?
[01:00:14] Mark Nottingham: Any initial responses to that?
[01:00:17] Kevin Kelley: Well, I'm sorry, Martin. I was just gonna say in the past, like, the issue that just came up, right, we've we've tried to steer away from caring about organizational boundaries.
[01:00:31] Martin Thomson: Yeah. So the different parts of systems was intended to sort of cover the different parties' things. So if you take a very broad view of the system as one that includes multiple entities, different organizations, or what have you, then different parts of those systems could be operated by different different entities. That may be a little bit too oblique. I don't think it's necessarily a bad idea to say different parties in this case if that would help clarify things.
[01:01:12] Farzaneh: Okay.
[01:01:16] Mark Nottingham: Obviously, we need more discussion on these. Hopefully, we can get some on list, get some refined proposals, and continue in San Francisco. If no one else has anything for this one here, I think we can move on. So that takes us to 02:47. So this is another, Leonard issue. I think this is almost editorial if I'm interpreting it correctly. He's he's asking if we can, oh, this is about the data model. Right? Martin, I think you talked to him a little bit about this.
[01:02:10] Martin Thomson: Yeah. I think this is... I would classify this as a an editorial issue more than anything else, but I I don't think this has any bearing on the meaning of things or how we're talking about them. I think everyone here understands very well what what the data model is. I think it just matter that I have some sort of disagreement about whether or not the the words that I'm using establish the the model clearly enough or whether we need something more formal. And I tend to think that it's pretty clear, but Leonard has some questions. So we're working through that. I don't think that we've Okay. Made a lot of progress.
[01:02:49] Mark Nottingham: We'll we'll let that
[01:02:50] Martin Thomson: I'd almost flag this as editorial.
[01:02:53] Mark Nottingham: Almost. I'm a little reluctant to do that right now, but I hate when I hear you.
[01:02:57] Martin Thomson: I get that.
[01:02:59] Mark Nottingham: Okay. Then that leaves us at 02:49. I don't like this new interface. Martin, this was yours.
[01:03:14] Martin Thomson: I'm just playing clerical. This is the question This is one of the bigger issues. In the draft.
[01:03:19] Mark Nottingham: Yeah.
[01:03:20] Martin Thomson: Right. And so I think this is where we we start to talk about the proposal that Chris and Kevin have put together about AI use, which rewords it slightly and then leads into other things.
[01:03:35] Mark Nottingham: And that's this one here, two fifty four. Yep.
[01:03:39] Martin Thomson: So this is an opportunity for Chris or Kevin to to stand up and someone to take over for Chris on the the note taking because I think this would be unfair to him if we ask that. So I'm happy to take some notes on this one if that if that helps.
[01:03:57] Mark Nottingham: Thank you, Mike.
[01:03:58] Chris Needham: If you if you don't mind that, yeah, that would help. I don't know. Kevin, do do you wanna sort of take us through this? Or...
[01:04:06] Kevin Kelley: Let me take a stab, and you tell me what I'm misremembering or misrepresenting from our our working session.
[01:04:13] Chris Needham: Okay.
[01:04:14] Kevin Kelley: Yeah. So so Krishna made some time and talks about this. What we're trying to talk about here, right, is the distinction between a situation where the user provides the resource or otherwise sort of, like, wants to use a tool that happens to have AI in it to interact directly with some specific resource.
[01:04:37] Martin Thomson: And the... Like, I think
[01:04:39] Kevin Kelley: a different way that we've... That the group has tried slicing this was sort of the the rag grounding distinction. And where we've landed is sort of this two part definition. The first part being that the user has to express specific intent to interact with that resource, and the second being that it has to be a specific resource. So if you ask some abstract question about why the sky is blue and the model goes out and and fetches data and uses it to try to answer that question, that would be AI use here. If you paste the URL or drag a file into the browser or do something else that's like this this resource... Interact with this resource for me. That would be outside of the scope of this definition. And I think, you know, my philosophy this whole time has been to try to draw dotted lines around categories of use that are... That people want to express preferences for. It doesn't mean that things that aren't part of this definition aren't worth talking about. But I I think that if Chris and I understood sort of where the group was after the last meeting we had, that there's a lot of... I was about to say consensus, but that's Mark's word, not my word. A lot of agreements that this kind of use is something worth discussing. How'd I do, Chris?
[01:06:22] Martin Thomson: Okay.
[01:06:23] Chris Needham: Yes. Yeah. That that... Yeah. That that... That's about right. So we... Yeah. We've tried to kind of draw a line somewhere between, you know, what what we mean... What, you know, what we mean by directly provided, and this is an attempted sort of, breakdown of that. Okay.
[01:06:42] Mark Nottingham: Let's, let's hear some reactions then, Alissa. Alissa? Audio not working. Alright. We we can come back to you, Alissa, or you can type into the chat whichever is easiest for you. Timid robot.
[01:07:24] Timid Zehta: Hello. Timid Robot, Creative Commons. I think this is a good definition of a... What was, I think, previously somewhere called, I can't find it right now, but, like, systems use. I do think that people want to express a preference about what is excluded here, And I think that kind of the other side of things was much more contentious, and I don't think it needs to hold this up except that I think that this just needs to be... The the naming of this just needs to be more specific so that there's room for people to iterate, work on what is excluded here.
[01:08:13] Mark Nottingham: So, specifically, you're saying instead of AI usage, should be something a little more specific?
[01:08:18] Timid Zehta: Yes.
[01:08:19] Mark Nottingham: Okay. That's helpful. Thanks. Brad Silver.
[01:08:26] Bradley Silver: Thanks. Yeah. I I actually find this heading in the opposite direction because, we've now moved away from provided the asset to rather interact with the asset. And I actually think that the provided part was the part that was operatively drawing the line between where the user has the asset and is providing it to the model. It means that they have obtained some level of privilege over what they can do with the asset because it's their copy or they have access to it and can provide it. Rather than saying that what is excluded is when I tell the model to go to do something with a particular asset. It's actually the part that I think is potentially more damaging because it is the part that says, I want you to go and find this article or this publication and I want you to tell me x y and z about it, summarize it, may fall into the broad construct of interact means and I don't know what that means per Sebastian's point. That is actually the part that is actually the disintermediation and the element that I think we would wanna express a preference around to potentially say no. If you wanna do that, then there's another way we'd like you to do that or here is here is, you know, a subscription or something or, you know, whatever it is rather than saying, oh, we'll just we'll just pass this along to you via the via the model, via the tool. We'll never know who you are. We'll never have an opportunity to get you to interact with our publication directly just because you have specifically designated a particular asset. I just don't think that that's that that's a solve. Yeah. Sorry.
[01:10:33] Mark Nottingham: Okay. Alissa, do you wanna try again?
[01:10:40] Alissa Cooper: Let's see. Can you hear?
[01:10:41] Mark Nottingham: Yes.
[01:10:42] Alissa Cooper: Okay. Yeah. I have two questions. One which kind of picks up on something Sebastian put in the chat, which is who is doing the interacting. It's kind of... It seems like it's implied that the user intent is for the model to interact with the assets or this is the AI system, I guess. But it's... Or maybe it is the model, but it's not clear. And then my second question is just, like, about a hypothetical. Like, if I'm talking to a chatbot and I say, like, go look at bbc.co.uk and tell me a few things about what's on the home page. Is that providing an asset with... That the model interacts with?
[01:11:42] Kevin Kelley: Can I respond to that for a second?
[01:11:44] Mark Nottingham: Sure.
[01:11:45] Kevin Kelley: I think in in Chris and I's discussion, that would not be interacting with a specific asset. Right? You're saying go go look at the BBC's website and do some stuff for me that's different than interacting with a specific asset or artifact or whatever noun we wind up with.
[01:12:13] Alissa Cooper: Okay. Any thoughts on the first question?
[01:12:19] Kevin Kelley: Remind me what the first question was. Sorry.
[01:12:22] Alissa Cooper: When it talks about the express user intent, I mean, I'm I'm super curious about how we know what the user's express intent is. It says to express user intent to interact. I'm assuming what that means is they want the model to interact with the asset, but it's not clear.
[01:12:45] Kevin Kelley: Yeah. I think that's exactly what I had in mind. Like, part of the art here, right, is we don't wanna sort of reify what products look like today into rules that will probably outlive us. But I was thinking things like, you know, our products currently pop up a dialogue that says, do you want me to reach out to this website for you? Whether that could probably take other forms. But, yes, like, the the user has expressed intent that they want the AI powered tool to interact with that specific resource.
[01:13:22] Eric Rescorla: Okay.
[01:13:22] Kevin Kelley: Does that Thanks. Help?
[01:13:24] Alissa Cooper: Yeah. Thanks.
[01:13:28] Chris Needham: Chris? Yeah. Perhaps I'll I'll come in on on that same point. It's... So the way that I was thinking of it, you know, in in the conversation I had with Kevin that led to this was it's things that are kind of equivalent to, like, the user has the asset already, and they're giving it directly to the system. And so if the user, like, uploads it to the, you know, to the AI system or the the user gives a URL and the system fetches it, like, the end result is kind of the same. But to Brad's point, part of our conversation was that this is an asset that the user already has visibility of. So part of... Certainly, part of my intent was that this would not be substitutional because it's because it's an asset that the the end user already has visibility of. So that's perhaps a distinction that might need to be made clearer in the definition.
[01:14:39] Mark Nottingham: Okay. Nate?
[01:14:48] Nate Hake: So first, I understand we're drawing... We're focused on drawing the squiggly line, but I just wanna reiterate that from my perspective, it's important to be able to express a preference as to all use. So either everything after the comma needs deleted, or there needs to be a separate category that covers assets, that are... You know, where the user is interacting with them. Now engaging specifically on the point of where to draw that line, I set out an example about a month ago, to the list of... Well, what about in the case where you have a search engine and you simply click a link, and then potentially, you could imagine a product. We're getting close to this with some real world products that by clicking the link, you then... Instead of getting a a... The HTML content from that website, you get an AI summary of that website. And what I heard Kevin say is it sounds like the intention is is that there should be an intent by the user for the AI, to interact with the content. And so that's good and maybe maybe implies that we could get somewhere on this. But I still would say that there... It gets tricky when you consider, well, what if that's just the feature of this product? Like, does the mere act of using this product that has that effect mean that you intend for that to happen by clicking that link? And so I I think this is a very difficult line to draw in practice, and I think it will be very difficult for third party tribunals or courts to to to go back and and look at this. And so I... I'm... I I appreciate the effort here, but I I think this line is is still unclear.
[01:16:24] Mark Nottingham: K. Ecker.
[01:16:30] Eric Rescorla: In the spirit of Alissa's questions, I'm just trying to wrap my head around what, you know, the the specific, you know, specific intent, is supposed to indicate here. And so, like, in the chat, I suggested one potential meeting, which is that you're able to distinguish what it is without use of an AI system. So, I mean, that that would... If you had unambiguous and, like, you don't need some, like, AI to interpret what it is, that may be the wrong line, but I'd be interested in understanding what... Let's distinguish as, like, one... What, you know, general intent from specific... General, you know, general asset from specific asset if that is not the line that we're trying to draw.
[01:17:13] Mark Nottingham: So I am I am still hearing a continuation of the discussion we had, at the last meeting, which is that, you know, there's the the question of where we draw this line and and and how we can draw this line. And then as, Timid Robot and and Nate and I think a few others have said, there's also the question of, you know, whether or not we... You know, there are a couple different options here in terms of of, you know, just having only with the carve out or also covering... Having more expansive version of use. And we discussed that at the at the last meeting. But at this point, we're just trying to figure out if we can draw a line in a place that makes sense and that communicates clearly. So that's not to preclude the other discussions. There's other aspects of this we still need to, to cover, but but that's, I think, the task in front of us right now. And and what I'm looking for is, if we can move the discussion forward and and have folks work on proposals that we can take to San Francisco to kind of discover the answer to that question, whether we can draw that line and then what it might look like. Martin?
[01:18:22] Martin Thomson: Yeah. So I just wanted to highlight something here. This is a really kind of admirable attempt to try to use a very economical description. I think I used that term at the meeting in London as well. We've been very economical with our words. I think this might be one of those cases where we need to have the definition, and then we need to say what we mean in in subsequent text. Because this this line, if we can draw it, is going to require some precision in terms of how of how we define the exact... Not just user intent, but also how the specific resource is identified and and sourced. And if we're not absolutely crisp on that one, we will have words that can be interpreted multiple ways by different people in different contexts, and that's not gonna work out very well to Nate's point. So I'd like to see that happen. Thanks.
[01:19:19] Mark Nottingham: I... And I'd agree with that. I think at this point, I'd rather have, see proposals with too many words and too much explanation that we can potentially cut down later on rather than having, the potential for multiple interpretations. Knowing that that getting to the right set of words is gonna be difficult no matter how we do it, but at least if we can all be on the same page of what the proposal is intending to say, that'll help. So I'd encourage people who are making proposals to to to, you know, put some exemplary, cases or explanatory text or whatever you think is is necessary to communicate what your intent is if if that'll help. Alissa?
[01:19:58] Alissa Cooper: Yeah. I just wanted to ask the question, like, if it would make... If it would... If people think of the easier path as maybe defining the exception here. We had all that discussion about, like, the whole category of, you know, anything that doesn't change the weights. I don't remember what the... That wasn't how we talked about it, but that that was the concept, which is, like, the big broad category. But the thought just occurred to me looking at this that, like, instead of defining it as everything but this exception, like, trying to figure out what that exception piece is.
[01:20:36] Martin Thomson: Alissa, are you talking about trying to define a category for that or to define language for it? Because I I think what Mark and I were talking about there was just putting words in. But are you talking about a specific category?
[01:20:52] Alissa Cooper: It was this epitaph on what you're saying. I agree with you that you're gonna need to have more words to describe this no matter what it is, but it just occurred to me looking at what Kevin and Chris came up with that trying to hone in on that on that piece that is the exception might be a little bit more straightforward conceptually than than, like, saying this is what the whole thing is except except this one piece. So that's all.
[01:21:18] Mark Nottingham: Would that be a defined term, or would that be Yeah.
[01:21:21] Alissa Cooper: I think that would end up being the... Well, yeah, I think it would end up being a defined term, and then it's possible that we... That it it has some relationship to the the prior definition, that broad definition of anything that doesn't change the weights that we were using earlier. Or we didn't... Never used it, but we discussed it in London.
[01:21:41] Martin Thomson: Okay.
[01:21:43] Alissa Cooper: Just food for thought in case it strikes anybody else's... A more obvious way to do it than this way.
[01:21:50] Mark Nottingham: Okay. Nate.
[01:21:55] Nate Hake: So I do think that if we're gonna have, like, a a exception, we need to be super clear about what that is. I think my original proposal months ago was to have two separate categories that would be expressible, and and to try to draw the line between user and sort of system in that sense. But I also think it would be helpful to understand, like, why... What what is it that we're intending to exclude from this definition when it comes to sort of use... User provided and and, really, what's the animating purpose behind it. For millions of artists who are out there, like, a... Having a preference that only relates to system use is borderline worthless to them because what we're seeing happen at scale on the Internet right now, especially on video platforms, is somebody puts out a video. That video is... You know, goes really popular. Like, let's say it's a dance choreography video. Choreography is something that we copyrighted in many in many jurisdictions. And then other users immediately use AI to basically recreate that video, but just substitute in the faces of other people, celebrities, other people into the, into us, copycat video, and those copycats basically dilute the original. And this is happening at scale. This is a significant problem. Like, the head of Instagram recently said that that, quote, unquote, we need better tools to stop AI from creating creators' work. So I actually think it's in everyone's interest, including the tech companies, because it's important to remember that spammers are users too. And when it comes to tokens, like, the majority of of usage that's happening is is done by these spammers when we're talking about about... Or or or copycats. And so I just think, like, we're we're really spending a lot of time on drawing this line, but, like, why I think it's important. And I understand that we wanna just focus on trying to draw a technical line, but I think it's really hard to do that without having the discussion about why we wanna exclude this and whether or not those are things that we actually want to exclude.
[01:23:55] Mark Nottingham: Farz?
[01:23:59] Farzaneh: Yes. Mark, I just wanted to, sorry. I sound worse and worse. I just wanted to mention that in all this conversation, I just wanted to remind us that we... If I remember correctly, in London, we also talked about some sort of carve out for accessibility, text to speech, and, you know, like, we should have that discussion as well. Like, why... I mean, we should have the discussion of why we want to have this AI use preference. Well, so we need to have the discussion, the carve out, and whether we want to, like, have a carve out in the text similar to what we have for search at the moment. Thanks.
[01:24:49] Mark Nottingham: Right. And I I I I do remember that that that, we were looking at... Right now, if I remember correctly, it's specific to search, and and you're right. We did discuss whether we need to generalize that. So that's still a to do for the group. Okay. We have five minutes left. I I think, you know, we... We've had some good discussions here. I think a number of folks have things they can go off and and work on and bring to San Francisco, and, hopefully, we'll we'll be able to, make some progress there. If if folks have any questions or wanna bounce ideas off of folks, can talk to Suresh and I. You can talk to the editors or just bring things to the mailing list. If if you have things that you think we haven't covered yet, please bring those to the mailing list. And, yeah, like we've said, we haven't declared consensus on anything yet. I do think we're we're trying to get to the point where we can start, some consensus calls and and and focus the group fairly soon. It seems like we're now having a dwindling set of issues. So you... If you have something that you feel hasn't been discussed yet or or needs more more consideration, please bring it to our attention pretty soon. Suresh, anything else to add?
[01:26:07] Suresh Krishnan: Yeah. Like, the one thing I remember, like, Farz, like, know, you jog my memory. Right? Like, you were gonna write, like, a use case for the thing, like, you thought was not allowed. Right? Like, so if you can get to it before San Fran, that'll be awesome. I know you're traveling, but, if you have got still a couple of weeks to get to it. So if you can add the use case, for the check, that'll be really good. Thank you. Farzaneh, yes. So thanks. Nate, is there something you wanted to say?
[01:26:54] Nate Hake: I just wanna say, I think it's... For me, it is incredibly important that we don't abandon artists and and and make this solely a publisher focused sort of protocol. And I think that, like, every time I have put out an example of, hey. Look at how user, you know, use can be harmful. I've been told, put it in the use case wiki. And I've been told that despite the fact that most people can't even find the use case wiki. I don't think it never... Have any of those use cases actually come... Like, the the wiki is not being used in any of these discussions. So I would propose that if we're going to have a carve out for user provided... You know, or user, user expressed, use, then the burden should be on those who think that this deserves to be a carve out to provide examples as to why this should be nonconsensual. Because if you can't express a preference about something, like, you cannot give consent unless you can withhold consent. And so what we're saying is is that you're not allowed to pro... Express a preference and as to, you know, something that is, user... Done by the user, what we are effectively saying is is that that use must always be nonconsensual. And I've heard some examples specifically with with respect to accessibility, transcription, text to speech, and translation where there are some arguments that have been forward for why generative use cases should always be nonconsensual. I would note that in the concept of sir... In the search context and in elsewhere, we have managed to come to a consents or some sort of a general agreement to to try to find a way forward for those specific issues. But if there's other usage that we think animates this, like, we should talk about that. Because I think if it's just the accessibility stuff, we can address that. But I I think it's, it's like I put forward a mountain of of specific examples in that in that wiki of use cases that I think are harmful and that I think justify the expression of a preference even with respect to user usage. And then the flip side of that has been nothing. Nobody engages with that wiki. Nobody engages with those use cases, and then we're being told, well, hey. We've had a lot of time to consider this, and it's it's... You know, if you read the chat, it's just like we're just gonna move on from this. I I... I'm... This is really problematic. We're... Bill Gates said that we should have human reserves. Like, we should have some portion of the Internet that is not... That is able to exist separately. This is really important. Like, that we we could make a lot of economic value by building in Central Park, but we don't because we think that there's some value to that. And there's some of us in the Internet who wanna maintain, our our ability to publish things for humans by humans without AI being involved in it. And I think I've made a prima facie case of examples as to why a preference we would want to express that. I think if the if the... If someone says, we we think you you should just not be allowed to express that preference, there should be a list of examples as to why. And the ones that I've heard that are generative, which is what we're talking about, because a lot of the ones are non generative, the generative ones I've heard are mostly around the accessibility things. And as I said, I think we can find something there. But I just really wanna make this clear that we're talking about abandoning artists. We're talking about abandoning creators. And it just... I am the only person, as far as I'm aware, in this group who who makes a a living by creating myself with my own hands and and doing doing things and putting things on the Internet. And so I just think that that perspective is underrepresented here, and I wanna be very clear that I I... I'm really concerned about the direction this is heading. So that's all I'll say.
[01:30:37] Suresh Krishnan: Mark, I think you're muted.
[01:30:41] Mark Nottingham: Thanks. Hear me now? Okay. So we're out of time. If if you have more thoughts, please do take them to the mailing list. Folks have proposals for San Francisco, please, make sure they're ready for discussion then. And, hopefully, now we... We've we've got a focused enough issues list where we can make some real progress in San Francisco. Sounds good. Thanks, for your time.
[01:31:06] Suresh Krishnan: Thanks, Holly. See you.