Blake Silvercrest

—

Q&A Section

22:28 – They define what Generative AI is (sort of).

23:40 – Q: What does opting-out cover, and what does it not cover? Short answer: It opts you out of model-training, or as Mike Minton describes it, Generative AI. Mike and Mary also clarify the distinction between Generative AI and the type of AI used for things like safety tools, since it's an important distinction.

25:17 – Q: If I opt-out, will AI-powered features, like captions or safety tools, still work on my channel? Short answer: Yes

27:33 – Q: How do I opt-out right now? Mary shares her screen and shows where the opt-out option is in the Settings page.

(RA Note: It’s under the “Security and Privacy” section)

29:17 – Mike jumps in to add that almost every content service makes users opted-in by default with no ability to opt-out (re: Generative AI). Twitch’s Terms of Service already permits the training of (Generative AI) models, but they (Twitch) went the extra step to add the control to let people opt out.

30:02 – Q: Why is it not opt-in? Mike: Well, there’s an honest answer, and I think most of you can probably appreciate this, like—If it was opt-in, nobody would opt-in. That’s honestly the answer, so by- by– It’s gonna be on by default, and I said almost every content service in the world is on by default, the- the– I think the thing that we’re doing here that is unique and different is- is respecting your wishes to opt-out of model training. And I- I know this is not a fan-favorite, I know it’s very upsetting to the community, but this is- this is, you know, where we are.

Mary: Yeah, we don’t expect you to be happy or excited about this, and… I… I don’t expect anyone to react to this favorably. It’s across the board. There are people who are like, “We’ll go somewhere else, we’ll stream somewhere else–” This is very industry-standard. Streaming somewhere else won’t absolve you of training, and things like this, and GenAI. Being able to opt-out is so important for you to share with others, and recognize that you can opt-out. So… please let other people know that you have this option.

31:30 – Mike offers some more clarification about specific types of use-cases (of different types of AI) like general-purpose LLMS (large language models) general doing things.

32:25 – Mike: We want to make a distinction that there’s no Twitch-specific use cases that we’re working on, as it relates to content-creation or anything like that.

32:36 – Q: Why can’t people opt-out of ALL AI-uses at Twitch? They covered this a little bit in the first few questions, but Mike says core things that improve people’s safety or experiences on Twitch (ex: discovery / improving safety / captions / etc.) rely on underlying systems that are understanding content to be able to operate. Letting individual channels opt out of that would break their overall promise to the community of safety (and also in the context of advertisers).

33:46 – Mike once again emphasizes that they will respect and allow opt-out for things that are working on behalf of a creator, especially their content.

34:07 – They start taking questions from chat, explaining that they might answer ones they already answered at the beginning of stream for folks who weren’t able to tune-in earlier.

34:25 – Q: Why not have it opt-in? Have you already been training models? Mike: Twitch has not been training models; I can’t speak to what Amazon is training or not training, as it relates to any type of specific usage of content.

Mike continues talking about how there are (depending on how you count it) 10 - 25 companies around the world that are creating what he calls ‘Frontier Large Language Models.'

35:00 – Mike: I don’t know this conclusively, but I think it’s a pretty reasonable assumption that almost any publicly-available content is being used for model training in one form or another—whether- whether permitted or not, so I think we also just need to acknowledge that, you know, there’s- there’s a lot here that is even beyond, you know, our direct control, but in terms of—as we said here— […] I can’t speak to what models have been trained or not trained, I can tell you that Twitch has not been specifically training models for our purposes.

35:45 – Q: The language in the new AI setting implies we are to choose between discoverability or having our content fed to AI, can you speak on this? Mike says this setting doesn’t impact discoverability, this setting is allowing your stream’s content to be used to train models that may help those models operate better in the future. He again clarifies the difference between Generative AI vs. other types of AI that help improve features like recommendations or discoverability.

37:48 – Q: Who are you selling our data to? Mary: No one, and I think that’s an important distinction. We don’t sell data.

Mike: No, we don’t sell data. […] That’s the most clear answer—we do not resell the data to other companies that are training models, so that’s not what’s going on here. We are not profiting or making any money by selling your data.

Mary: Amazon also doesn’t sell your data. Sometimes I think people are like, “You don’t, but you’re Amazon, and Amazon does.” Amazon doesn’t sell your data either, they have a strict policy on that. So, when I say “We’re not selling it,” I mean us. The royal “us.”

38:42 – Q: Given the response to the AI toggle, will you consider changing how AI-related features are rolled out in the future? Perhaps giving users a timely notice to toggle-off AI-related features before they’re implemented instead of auto-enrolling all users? Mike: Yeah, I think that’s especially applicable on the creator side, where we have things in the future that may use AI to help improve those features, you know, we’re— It’s kind of a– We’re trying to be as clear as possible about what’s happening, but at the end of the day, and I firmly believe this, it really comes down to the end-product experience. Like, if it’s a good product experience, and it’s solving a problem you care about, how it’s done is less important than what the quality is, in terms of that, and we saw that feedback quite a bit with stream summaries.

Mike talks about the good and bad feedback they got about that feature, and how if stream summaries worked well, they might’ve had different feelings about it.

40:11 – Q: Did Twitch ask the Ambassadors and Creators their thoughts prior to rolling this out? Mary says she can’t speak to all creators / ambassadors, she did not ask ambassadors their thoughts.

Mary: We were pretty tight-lipped on what we were able to ask or provide.

She says she did read through all the user voice entries and she’s aware of the current vibe / reaction from the community re: AI. She says she’s sorry she couldn’t invite ambassadors for their reactions early. They really wanted to try and release this after getting all their communications approved, but it was a really complicated rollout. She apologizes again for not being able to include the ambassadors early-on.

41:10 – Questions about comments on clips (not related to the AI discussion, so I’m not including these).

42:37 – Q: Have my VODs already been used for training? Mike: [Long pause] In- in- in terms of— I don’t actually know the answer to that question, ‘cuz I don’t know what Amazon—or well, Amazon in this case—has done in terms of model training, and what they’ve used and not used.

Mary: Opting-out removes your previously-collected content from future training, and it makes sure your content isn’t used in future Generative AI training, so I would—and I think it works immediately, so… I would turn it on ASAP.

(RA Note: She specifically says "future" training, so it's unclear whether it's already been used for past training).

Mike: Y- yeah, but let’s—I mean, again, let’s try to… to level set here. When you think of the people that provide Frontier models to the internet (Frontier LLMs) and are doing this at very large-scale, Amazon is not one of those companies, turns out. The biggest ones most people are familiar with are going to be, you know, Chat GPT or OpenAI, to be specific, and Anthropic and (Clod?) etc., so… [Stammers] I think the honest answer here is like, if models were trained on your VOD, it may or may not have even been Amazon, it was likely any one of the, as I said earlier, dozens of companies that are creating models that are using any publicly-available content they can find.

44:09 – Q: Did the anti-GenAI user voice have ANY impact on the rollout of the GenAI update? Mostly want to know if user voice did anything, thank you. Mary says yes, she reads the user feedback and other people do too, and it was used to impact what happens.

Mary: This has been in the works for longer than a week, this has taken weeks, if not months, to put together. Your voices and user voice ABSOLUTELY matter, and as (Mike) Minton was saying, I don’t expect anyone to jump for joy or be happy about this, but what I am saying is like the opt-out is a reflection of reacting to this community’s voice that you are not wanting to train GenAI models, and we want to give you that option. This toggle if a good example of hearing you and doing something for that opinion, so yes. Yes, we did.

45:14 – Q: There was a 13k user voice asking for it to be opt-in originally, did you use that? Mary says yes, but the user voice was constructed before that, the user voice wasn’t asking for it to be opt-in originally. She says most of the feedback they see asks for no AI whatsoever, or more options about AI. She says they read that and utilize it for their feedback and the way that they adopted the rollout. She knows this isn’t a satisfying result but they did read and use feedback.

45:52 – Q: Will channels who remain opted-in be labeled for audiences as such? How do we know if a user has it on or not? Mike says that depends on the feature, but there won’t be any label. Things like stream summaries will be obvious because users can see that that feature is turned on.

46:44 – Q: Why did Amazon and Twitch not use existing content models? Why is Amazon choosing to create a proprietary model? Mike: For Twitch purposes, to be clear, we are almost exclusively using existing content models. When we train models specifically, they are not Generative AI models, they are models that are more into what we would call a “Machine learning model,” which outputs the score, a number, or something like that. We do train those types of models for specialized safety-use cases, etc., but when we’re doing content understanding, things like what’s being said in the stream and creating a transcript from it, we’re using an off-the-shelf model, it’s not a proprietary model.

47:40 – Q: Honest question, who’s this for? I don’t see this benefiting content creators, I see it used to replace you. Mary says she thinks they can say something about that, she says she doesn’t think any of the use cases are to replace you (ie: streamers), replace your voice, replace your character, or any of that.

Mike says they’ve already answered this, they’re not trying to replace streamers. They have no plans or efforts for it, and they think it’s a terrible idea. He says their energy is around helping their creators be more effective, work harder (in terms of keeping their communities safe and creating content), etc. He gives examples of auto-clips for smaller streamers and captions on clips, and improving the lives of the community on Twitch.

49:57 – Q: Why frame this as being about accessibility and community? That feels really misleading. Mary: It is not my intention to mislead you in any way, I think when we have these conversations, it’s really immediate and sometimes reactive, and this makes sense to me, because you don’t like AI in any form of it, and so we see a lot of chat that’s just like “No AI in any form, period. I don’t want it,” and I hear you. We’re not trying to distract you or, you know, make light of this, but I do want you to recognize that AI is used when we have a community of this size, and at this scale, to help us be more efficient, and that can be used for safety features, and for community-building features that you like, so it’s- it is important to have that nuance. So, I’m not trying to trick you or anything like that, but there is a difference between training AI Gen models and AI on Twitch, and so when I talked to the ambassadors about it this morning, they were like, “We’re not stupid, we know the difference,” and I don’t mean to make it sound like we think you’re stupid, or that you don’t understand, but I do think there are people out there who don’t fully grasp that difference, or they basically combine them as, “I just don’t like AI.” I don’t think we can function without some forms of AI, and that’s why you can’t opt-out out of some of them. And so I apologize for that, but it is a distinction we have to bring up.

Mike talks about how some survey said that more than half of adults use some form of AI on a daily basis, “Now you wouldn’t know that from reading Twitch chat, of course,” but he says it’s a general population-type study. He also says that the anger and anxiety about AI is also high, and those two things are in conflict. He says their bet is that in the long run, all creator concerns and participants within the (Twitch) ecosystem will be addressed in the same way that past technology transformations have been addressed over time. He once again talks about some different examples and ways it can improve user and creators’ experience on Twitch.

Mary says it’s a real tough time for AI right now and a lot of things are being scraped, and she reiterates that folks should utilize the option that they have to learn about training for GenAI and use your best judgement whether or not to opt-out.

Mike advises checking settings on other sites too if you want to avoid having your content get trained on.

54:47 – Q: Is the opt-out only available on desktop or mobile too? Mary says it was a dual-release on IOS and desktop, she recommends updating things if you’re not seeing the option to opt-out on mobile.

Mike reminds folks it’s in Settings under Security / Privacy.

56:32 – Q: Would any art and assets that we use on our streams be used for training? Is it mainly our voices or the camera? If we are Vtubers, what is being used? Mike says he’s not responsible for Amazon’s training efforts or terms of use / GenAI training, but he’d assume anything that’s part of your channel content wouldn’t be used if you opt-out.

Mike talks about how it can be difficult doing these kinds of Q&As for a variety of reasons.

Mike: What we’re trying to do is triangulate answers here that are complex answers to questions that have, in some cases, concerns from a ‘What we say can get us in trouble’ perspective, and not just with the community, but like literally in trouble in different ways around the world. We have—obviously I am not speaking on behalf of Amazon on any of their efforts, so there’s this triangulation that’s going on to try and find an answer that is as, you know, helpful as possible to explain what’s happening here and why without breaking into jail in one of many different dimensions that we—or that I specifically—could be breaking into jail here, so I appreciate and know that this is a very sensitive topic, and I am honestly doing the best I can to try and find answers that answer your question and respect your feelings and frustrations about this specific use of technology. I understand that I live in a bubble that is not your reality, as it relates to this technology and your use of it, and so I, you know, just want to emphasize I respect your opinion, I respect your feelings, and we’re doing the best we can to try and thread a line here that meets the needs of everyone.”

58:44 – Q: Will this make us more money, if users stop using Twitch because of this? Mary says it’s a delicate topic, and the community had the reaction she expected us to have, which is that we don’t like this. Because this is industry standard, when people say they’re not going to use Twitch anymore, going other places won’t absolve you of this.

Mary: This opt-out is something that we really encourage you to use and share with your communities so that they know, I know that I will tell my community, ‘Hey, just so you guys know—I have opted-out for GenAI,’ because I have community members that would not join my stream if I opted-in. So I’m gonna tell them and be vocal, and I encourage you guys to do what’s best for you and your channel settings, but it’s not as easy as saying ‘everyone’s just gonna leave’ because this is something across the board. It’s something that’s happening on livestreaming communities across the space.

1:00:20 – Q: Captions on clips is great, but we don’t have the option to edit the captions, and sometimes they’re incorrect. Are you gonna let us add that or fix that? Mike says they’re working on that. Sizing, positioning, and colors are also being worked on.

1:01:00 – Mary mentions there are some questions about laws and complying with EU laws, she mentions it’s gone over a lot of lawyer scrutiny, so it should be compliant with all laws.

1:01:29 – Q: You mentioned most people use AI nowadays. You understand how much of it is being forced onto us against our will, right? Almost every website I go to these days, some AI bot is shoved into my face within minutes. Where does it being forced against our will factor in? Mike acknowledges that, but the specific survey he was referring to were using chat bots like ChatGPT to solve their own problems or their own usages specifically, it was not counting the people who didn’t have a choice (like Google).

Mike says they’re trying to answer as many questions as they can.

1:02:42 – Q: [To Mike] If you were a streamer, would you opt-in? Mike says for him personally, just because he’s more comfortable with what’s going on and what the technology is doing and not doing, he would not bother to opt-out. He says everyone needs to make their own decision about it though, and that’s why people have the ability to opt-out.

1:03:07 – Q: What happens if a creator who opted-out is collabing with a creator who opted-in for a stream together session? Mike says what they should do there is respect the opt-out. He says they’ll consider that use case.

(RA Note: In other words, there isn’t something in place for that situation yet and/or they haven’t considered it).

1:03:58 – Q: They want to have a way to see what channels have opted out. (Mary paraphrases this question since they already answered it earlier) I don’t want my chat participation used, so I’d like to avoid those streams. Mary says right now, there’s nothing that lets you know someone is opted-in or out. Mike says you could go channel-to-channel and ask people, but they don’t intend or have a plan to make an indicator of that.

1:04:43 – Q: Will Twitch’s official streams be opt-in or opt-out? Mary: I think we’re gonna opt-out.

1:04:56 – Q: Wasn’t this clear in the user voice [feedback]? Mary: Yeah, I mean, I think you guys made yourselves very clear in the user voice. I don’t think that can be misinterpreted. I think this community’s stance has been pretty clear on your views of AI. There’s no misunderstanding, I hear you loud and clear.

Mike: Yeah, not tone-deaf here in terms of knowing where the community stands, the feedback’s been quite clear.

1:06:03 – Q: You don’t know what it’s being used for training? Mary: I think the better answer to that question is Mike does know, we do know that it’s being used to train models, but it’s a BROAD stance of what you can train a model for, so giving you a list of 500 use cases I don’t think is actually very helpful, the point is it helps train models, and the use cases are endless, probably. So if that makes you uncomfortable, or you don’t like that, then I encourage you to use the opt-out function, and to tell your friends.

1:06:48 – Q: Why didn’t Twitch email people? Mary says that’s a fair question. After a pause, she says they have conversations internally about how to release this information, and they use Twitch Support most regularly, and they moved the release date to be on Patch Notes so they could talk to the community about it directly and have this conversation. She says social media is a very effective tool and that not everyone emails creators. Social media, DMs, and live streams tend to get the information out further than an email, and she acknowledges that this isn’t the best answer, but it’s also true that creators don’t always read their emails.

1:08:00 – Q: Concerns about voice cloning [Paraphrased] Mike says that’s specifically not one of the things that’s being done, it’s maybe used for speech understanding, but it’s not used for voice cloning specifically, and that’s one thing he can confirm isn’t being done. He says, “the people responsible for it did confirm that detail.”

1:08:51 – Q: Are chats opted-in? Mary says it’s creator-based, so if a creator has opted-out, then their chats are not included. “You, the creator, make the call for your chat, not chat.”

1:09:29 – Q: Hey, [Twitch] is just trying to hide this Mike says this is the opposite of hiding it, acknowledging it publicly isn’t hiding it. He says they’re here and doing the best they can to address user questions, and they’re doing their best to make it clear that the opt-out option exits. They published a blog post and tweeted about it, he says he thinks it’s good feedback in terms of email or notifications, they’ll look at other ways to make sure this message gets out as broadly as possible.

1:10:06 – Q: What stops Twitch from using AI in the future, from adding more features that can control creators? Mary: I would just reiterate that user voice is a powerful tool that we do read, and it does impact things that we do or don’t do. Having a user voice with that many upvotes and that many comments is something that I can utilize. Your voice is important and I think—I encourage you to continue to be vocal with your opinions on AI, which do help us make better decisions for you guys over time.

1:10:47 – Q: Will opting-out be used to lower eligibility for front page? Mary says no, it’s not related and creators won’t be hurt for turning it off. Not discoverability, not eligibility for ads or marketing or any opportunities, it won’t impact creators. Mike adds that it has zero connection to site functionality, this option is only about training models, “Models, at least in our current reality, Twitch is not using.”

Mike clarifies what he means by “current reality” is that nobody can predict the future. Things change, but he thinks a lot of these concerns will be addressed. Mike also says, “I do know from just the very basic functionality, there are certain things that AI does better than any other technology we’ve had before, as it relates to content understanding at scale. That understanding then lets us do a lot to improve the viewer and streamer experience on Twitch.”

1:11:27 – Q: Can there be a Twitch notification for the browser and mobile apps? Mary said there could be, and they should probably continue to get the word out about this.

1:12:34 – Q: Would it be better to have people opted-out by default? Mary says they talked about it earlier, but the opt-out function is available and it’s something they really encourage people to use.

1:12:50 – Q: [General question] What are you using it for? Mary: To train GenAI models. Opting-out stops that immediately, like, the second you click it.

Mary encourages folks to continue being vocal, especially in spaces where they can continue to have this conversation.

1:13:42 – Q: Will GenAI be trained off our VODs, or are they trained off of livestreams? Mike clarifies that his earlier answer to this question wasn’t as helpful as it could’ve been, so he clarifies:

Mike: I think the important thing is that if you remain opted-in, any of your channel content, whether VOD, clip, live, chat, anything in and around your channel, may be used for training. The point I made earlier is that I don’t know specifically how it will be used in a given thing, we’ve given one example, which is- we think is a reasonable assumption, which is for audio. You know, one of the biggest use-cases for AI is translating audio into text, and so the use of your audio may be used to improve the ability to do that translation, and then that would improve future features potentially, like captions or even– well, live captions, or clip captions.

1:14:55 – End of Q&A, Closing Remarks Mary says there will be a lot more conversation about this in the coming days, and she’ll do her best to consolidate a lot of this feedback from chat and user voice in different spaces so they can use it in their internal conversations at Twitch. She says she really appreciates people’s passion and thoughtfulness in this conversation, and she says, “I hear you fully.”

She says they will have “Patch Notes” streams like this every month, and they will be communicative and tell the community what’s going on. She appreciates folks talking with them and participating in the conversation. She thanks Mike Minton for joining this “tough stream” and for being, “As honest as you absolutely could.”