Using AI Assistants

[plehegar]

AI assistants during teleconferences can be used for:

1. translation of spoken languages
2. voice transcripts (accurate or polished)
3. note transcripts
4. ask follow-up questions
5. identifying takeaways, summaries, tasks
6. creating new content from meeting notes
7. suggesting next steps

So far, we only have a policy that covers transcripts.

Should we look at policies/guidelines for all of the other cases?

cc @swickr @dontcallmedom @tidoust @iadawn

[amyvdh]

Please be sure to add privacy considerations (all parties agree to be recorded and their words to be sent to a 3rd party). In addition, I think we should state every instance where AI is used needs to be added in any resulting info and that there's always a commitment that an actual person will verify the information.

I wish there was encouragement to consider, like printing out paper or taking flights, some consideration of the ecological toll of using AI.

[koalie]

I wish there was encouragement to consider, like printing out paper or taking flights, some consideration of the ecological toll of using AI.

Thanks @amyvdh

STRONG +1 to this, @plehegar

This draft guidance seems premature to me. We started a team-internal conversation a few months ago about whether we wanted to adopt a stance to allow or restrict our usage of AI. This draft circumvents this conversation. Please let us not put the cart before the ox.

That being said, in case where we opted to welcome some usage of AI, it would be ideal to me (wearing my Comm hat on) if we were able to craft the right warnings and be specific what we consider OK and NOT OK.

I want to avoid at all costs that W3C be seen as giving anyone carte blanche vis-à-vis AI. Nothing is normal about AI, unfortunately, from provenance of collected data being unfairly or illegally done, to unreliability of tasks, to the appallingly disproportionate amount of energy and water AI uses to complete tasks more or less badly.

[cwilso]

This draft guidance essentially says "don't use this", does it not?

I would absolutely agree that W3C should not give carte blanche vis-à-vis AI tools. At the same time, having to rely on volunteer human scribes is a poorer experience than using transcription tools, and with some management of the process, I believe it's possible to avoid at least some of the concerning aspects of automated transcription. (We have, for example, worked out agreement in order to use automated transcription and summaries for WHATWG triage meetings. We do human editing passes, and do not share the raw automated output.)

[iadawn]

Part of the reason this came about was because of an event in one of my meetings. This relates specifically to Zoom's AI Companion. It is worth noting that the AI Companion is not necessary for transcripts or for using transcripts as an approach to minuting. Nor is it required for translations (I believe).

The key use case we have for switching on AI Companion is to support participants with cognitive disabilities. This is an important use case that it would be remiss of us to forget.

The challenge we had was that a participant from a public sector organization identified internal policies they had which imposed limits on their participation in certain circumstances related to privacy and data protection when such AI tools were involved. This covered the legislative region in which the data was stored (if at all) and whether data was used as training data for AIs.

There a couple of options available that can mitigate the concerns raised, although there is a good chance they will cause other problems:

• Managing data and storage locations - the storage location for recordings, recorded transcripts and other things can be specified. This option is available on any paid account.
• Retain and access meeting transcripts - this option (I think) stores the transcript of the meeting and makes it available for the AI companion . It is not clear if this means that it is used for training purposes.

There are a number of other options related to to AI companion and data protection:

• Auto-delete meeting summaries
• Only share meeting summaries by email
• Show a custom disclaimer when turning on AI Companion in a meeting

This is all by the way of background information for the discussion. However, I will reiterate that AI Companion has a significant use as an assistive technology. There is often a tension between security and accessibility. Can we please not fall into the usual pattern of security trumping all and people with disabilities being excluded as a result.

[reidmore-online]

FYI unfortunately in the case of Zoom, AI Companion is required to get transcripts, it looks like they got rid of the functionality to save transcripts separately from AI Companion. Zoom does say they do not use transcripts or the meeting contents to train AI Companion here: https://support.zoom.com/hc/en/article?id=zm_kb&sysparm_article=KB0057623

We have used it in meetings in a CG setting and it has been successful for us so far, I think we should be having a conversation about encouraging this if Zoom AI Companion meets our security/privacy concerns. It should be up to groups how to use it, when to use it, but it's worth discussion.

Just as an example, my CG's workflow for using it currently is that I get the transcript, I do a first pass and clean it up/format it, I then copy it into pad.w3.org, share that link with the CG, leave it open for about a week, then close it and save the minutes in HTML to our repo. I delete the transcript off zoom (I'm not using W3C zoom for this, but my org also has a retention policy of only 30 days).

[hidde]

Should we look at policies/guidelines for all of the other cases?

Yes. Many groups struggle with deciding on AI companions.

While I have sympathy and understand that:

• AI companions can be useful (or, if we're honest, perceived to be useful) to people with cognitive disabilities and non-native English speakers.
• volunteers for scribing are hard to find in some groups.

I would like us to consider disallowing (or at least discouraging) AI companions, for these reasons:

• possibility of misrepresenting, misgendering, misattribution is inherent to the technology and likely to disproportionately affect minorities, non-native speakers in our community.
• uncertainty about value because of subtle incorrectnesses: in our work details matter a lot. I know Zoom says you can get questions answered and identify takeaways, what's the value if the answers are potentially wrong? If it took away the wrong things? This could sit in the way of effectiveness of our community, shifting our focus to whatever was statistically more likely.
• decontextualisation: these tools miss a lot of context and context is an essential part of our multi stakeholder consensus process.
• climate and ecological impact weighed against value: as Amy mentioned, there is an ecological toll, and, while not denying usefulness, I think we need to question how it weighs up.
• shifting of responsibility and accountability: when the bot does it, who is accountable?
• budget: chances are they start charging more for this functionality as it is expensive functionality, may higher prices change the cost/value equation?

(We can circumvent the first three with thorough human checks, but that could easily cost more time. And yes, humans make mistakes too, but with Zakim the whole group can check and improve in real-time).

Instead we could:

• seriously look at ways to make scribing more attractive. Wild scenarios could include paying scribes, making it explicit part of group joining.
• publish audio or video recordings of meetings (some groups do this already) so that people can understand discussions in context.
• replace, in some groups and contexts, scribing with “light minuting” and manually answering the questions @plehegar mentions in his post: maybe spend the last 5 minutes of each meeting writing down Important Things like takeways, tasks, next steps. Zakim already supports tasks.
• reach out to Zoom to allow transcripts outside the companion.

On a personal note… I am seriously considering not participating in groups when it is turned on. I think we should not facilitate and normalise this at scale.

[xfq]

FWIW, our WG uses AI companion, but we don't use it as a replacement for manual note-taking. We still take notes manually, and we also use AI companion for meeting summaries, and we publish both versions.

For the AI ​​companion's summaries, I manually review them carefully each time to check for errors, such as misattribution, or parts that shouldn't have been recorded but were, etc., and then I publish both (because the two versions differ in their level of detail).

[reidmore-online]

One thing I wonder for our use case, and considering the audience we serve, is whether it is worth investing in a lightweight tool of our own. There are open source packages for audio transcription which we could set up in a closed environment to mitigate concerns about privacy or retraining. We also need a tool where there is a human closely in the loop, something most of the tools on offer today can't provide, everything is closed off from live editing. I think it is possible for us to have something that bridges concerns and benefits, because while I think human scribes are essential, it's a role with a lot of challenges that can't be easily overcome by many. A scribe role where someone is taking minutes at the level we need now is really hard, but if the scribe role becomes more of an editor/checker, it could be much more accessible to more people, or it becomes a group activity, as happens in some groups that scribe via Google Docs or Ether Pad.

[iadawn]

@reidmore-online I think this is something that would benefit from a little bit more investigation both of the needs and the options available.

Note, that I would maintain that from an accessibility perspective scribing is not the primary need - although there is some work required to explore this a little bit more.

[plehegar]

Thanks all for providing input to this! I feel we'll need more input before attempting to push forward on this.

So, as part of that, I want to make sure to link to other relevant thread/issues:

• https://github.com/w3c/AB-memberonly/issues/180
• pointer to the ongoing work on AB position paper on AI? cc @hidde

Anywhere else for past discussion/conversation?

Anyone got a link to existing guide/policy on AI assistant usage (within or outside W3C)?

[plehegar]

cc @TzviyaSiegman , who is looking at the AI experiment happening at GSF (which I believe falls under item "creating new content from meeting notes" but also goes beyond teleconference assistants.

[TzviyaSiegman]

@plehegar more info to come about GSF's tooling in January, but it falls under the category of creating new content with AI from prepared materials. Content refined by humans.

[RByers]

For the specific question of scribing, I want to point out that research suggests that women are ~50% more likely to volunteer for such jobs, and that volunteering to scribe tends to make one less able to participate deeply in the discussion. My personal experience in W3C meetings is strongly consistent with this research, to the point that I regularly feel compelled to volunteer to scribe because of how uncomfortably obvious the gender bias is.

I'd like to propose that, knowing this research, it should be against W3C policy to call for volunteer scribes and instead become an essential job for W3C to fund one way or another. Realistically I think that means either:

1. Asking W3C staff to scribe as part of their job (some already do this consistently), or
2. Increasing dues to hire professional human notetakers for all WG meetings, or
3. Building some infrastructure (and training?) to allocate scribing duties to all meeting attendees equally, or
4. Accepting automated note-taking followed by offline expert human review/editing/augmentation of notes prior to publication.

My personal vote is that (4) is the least bad option (as someone who already relies on AI note-taking and summarization in most of my work meetings), but IMHO any of these would be far better than the status quo.