And for Meta, it is going to take more time, right before your chatbot reaches Taiwanese people. So it allows us to develop a more comprehensive testing plan based on API’s hopefully, with AIEC here and with Anita — I’m sorry about the mistranslation before — but we would love ...

One from OpenAI, about the research access program in which AIEC will also have a window in addition to your window there so that when we detect significant regression or things like that, we will let you know before it’s deployed to the general public or just as it’s being ...

Excellent. So now we have essentially two commitments.

Okay, so we’ll ask your licensor, OpenAI. If you have any comment on this particular issue.

Okay, excellent. Thank you. And then we have Microsoft.

We’re not talking about a model. We’re talking about an API. So API, just like the existing API that takes a prompt and normal prompt and feedback, something so that the evaluation is done not on the model weights itself, but rather as API user is just that it’s more ...

So I think we’re talking fundamentally about two different things. This is about pre-deployment or just about to be deployed, this kind of testing. Do you have the same concerns, as you outlined?

Okay, we get your idea about post deployment evaluations that would need to respect the privacy of these users and trade secret and so on. But for this one, I think what we’re asking is essentially, pre-deployment testing, meaning that instead of just testing Gemini, as it is something more ...

We also earlier I heard you saying that’s the Mandarin version or the Taiwan version is not going to be deployed to production soon, which means that we will have plenty of time to do the testing and evaluation and thank you for committing to taking this back to your ...

Meaning we just go ahead and download the 140 billion model of Llama 3 ourselves and make the test would you be interested in receiving the testing reports?

And what we will require, of course, is some sort of researchers access to your API or in case of open models that are yet to be released. Some early access to your models. I hope that this slide speaks for itself. So again, maybe let’s begin with Meta which ...

And so, as a pilot, we’ve already commenced evaluation for the TAIDE model, which is now an open model with that you can all download with hugging face. And so we here invite all of you to also join as TAIDE did, to the pilots program so that we can ...

So, to ensure that the content generated by Gen AI because there’s a characteristic fairness, reliability, safety, etc. We are working with the the NIST as I mentioned, the AISI within the NIST in the US, the AISI in the UK, and other counterparts for example, the one in in ...

And we understand there’s also ways to just take the classification models in a kind of remote testing way and so on. So while we will leave the technical details to working over the discussions, I just want to use the event consensus slide to remind that it is not ...

Okay, great. Yeah. And we look forward to working with you on that. I think there is a strong consensus during the deliberation that there needs to be like useful monitoring the levels of like, if you say that your identify and manage capability is at x percent. People expect ...

so that whether the system you mentioned that maybe Ross and so we learn is May entirely developed in house or it is somehow in partnership with some other third party system or things like that.

So legislators who pay strong attention to these consensus are expected to take some of them into their platforms and we also have a upcoming anti-Fraud Act that will incorporate some of these measures. So while this one is just to gauge people’s general understanding, and make sure we’re all ...

Because we have 14 days to provide written addendum. If you have timelines and such please do add to the transcript. Okay, great. So I think we have covered all three points in some detail. Does any of you have your own, like additional points to make or additional responses ...

I think that this address point two. So you’re essentially saying that if it’s deceptive in the sense of is Gen AI or its deep baking someone in depth causes harm, then that is already governed by your community rules. On the other hand, you strongly encourage people to label ...

And is this coming to Taiwan like is this experience how we could have in Taiwan?

So that there is a direct correspondence, right because Facebook or Instagram you can also post pictures and videos and also advertisements. So for this point I would like to hear Meta’s response.

Now YouTube products that you’re also rolling it out such as system?

YouTube allows YouTubers to or creators to label Anita was it only like videos or just general purpose like pictures or sound recordings and so on?

Okay thank you. So you will provide a response to these three points in written form. Okay. Understand that. Thank you. And now let’s circle back to Meta. On q1 and q2. We talked a little bit about how those deceptive contents who are persistently posted maybe detected in the ...

Understand. So we’ll continue to build a relationship with your team. Thank you. Now we have Microsoft back.

Thank you. So you mentioned like significant self-harm and so on. So if our community already have a list of such resources as well as the triggers, let’s call it the fight them for such resources, would it also be in a local affairs team? So that we can get ...

Sure. Basically it’s about you talk about a specific case. But Google just mentioned for Gemini, there’s general purpose double check tool that would allow for like, orienting fact checks or an misinformation attributions. You mentioned that ChatGPT when it’s allowed access to real time user information marks the source ...

So on your last point, when you say that, for things related to local elections, who point them to a canonical resource, what is the process of for example, adding such canonical resources? Is it a government representative relationship with OpenAI? Is it civil society relationship? How is the keywords ...

Okay, so while Allen is figuring out some technical issues with Microsoft Teams, we will go to Open AI. So, please would you like to talk about in these three points?

Now, Meta pointed out is not entirely foolproof or is not unbreakable. At least it provides an industry standard to go forward with. And finally, you also mentioned that your ads and so on are already comprehensively being rolled out to adhere to these signature and watermark related provenance schemes. ...

Because there are 14 days to co-edit this transcript. So if you can get the answer within the next 14 days, we can retro actively add it to the transcript. And it sounds as if you answered this question right here. But for the other points, I think yes, you ...

Thank you, that is very comprehensive, and indeed addresses all three points. I have one follow up question. So when you say that for users who persistently post deceptive content, you need to give creators some time before you rule out the rule that suspends them. When they do so, ...

Okay, So, you have addressed the point three, right. So about what your marks and detection systems you mentioned, that the provenance that is to say if it is generated AI, we should know which tool generated for example, but you also noted that if I as Audrey Tang, then ...

So please feel free to start covering either one or all three of these consensus topics. We will make a full transcript but you are free for 14 days after this meeting. After a transcript is made to co-edit so that you can also add relevant links or read act ...

And finally, for the AI tools providers for generating AI content, especially the visual ones or audio visual ones. We should label them with provenance signatures or watermarks, etc. But the goal is to indicate to viewers that they are AI generated and I will just stop here and begin ...

First, for interaction with generative AI people would like you to clearly mark them as such. And if there’s a citation or sources where they came from, people would strongly prefer to see citations and labeling of those information channels, at least the source where they came from, for users ...

And what does it mean concretely? We chose three consensus always overwhelming support after deliberation as the topic for the discussion today.

And finally, large platforms should be evaluated on their capability to manage and identify the session and adhering to cybersecurity best practices. This is worth exploring a little bit here, because as of this week, our ministry became the competent authority for large online advertising platforms. So we will begin ...

First, but the public should be continuously invited to participate in a policymaking formulates mechanism and standards, which means that our aim is to at least run this kind of alignment assemblies twice a year, or maybe more as situations corporate large platforms should analyze and identify deceptive AI use ...

In the end, we’re very happy that more than 450 participants, including 120 practitioners, they’re sorted in groups of 10 randomly stored hid, and AI facilitation robot aided by Stanford, allowed to tap into not just come to consensus after deliberation, but also raised important topics for us that policymakers ...

Session Two, we deliberated about how should platform help the user maintain the standard of information integrity. And how we should help practitioners testers, the creators of media to maintain their information integrity standards in their practice?

First, self regulation versus government regulation. How should AI help analyzing emerging threats?

Now, we use a democratically rigorous process. We sent to 200,000 people through the official Baldwin one SMS platform a survey. And then through stratified random sampling, we chose at least 400 citizens that represent that microcosm of our citizenry, and so many of them agreed to participate. And on ...

Fortunately for Taiwan, while there were some of those manipulation attacks during our January election, our information resilience ecosystem has countered these attacks pretty well. Leading to low, effective polarization following the election. However, there’s many, many other democratic countries going to be in the elections this year, and many ...

And so this direct democratic input has the ability to not just educate the public about how to instruction tune how to alignment tune a language model, but also let people have a say, in tuning those models. And, in addition to the ideas on alignment assemblies are the future ...

It’s just fine. So welcome. So now, if there’s no further questions, either from here or from remotes, let’s get to the main discussion topic, which is alignment assemblies. Now, alignment assemblies is a project that we jointly launched with the CIP collective intelligence project last year during the summit ...

Okay, great. And we see OpenAI online. Hello. Would you like to say hi to the rest of us so that we know your microphones? Right?

So, the short response to your question is yes, while the accuracy one is somewhat domestic the other one is working relationship with international counterparts, which is why in the slide, we say doubts by next year, we fully expect our emails meet international standards and if they also accept ...

And also, we work quite closely with the leading institutes domestically in Taiwan. So that’s the ITRI and that’s the NICS and those two together will also ensure for example, for the secure part, the cybersecurity related issues, then NICS already has capabilities related to cybersecurity evaluations. And for that ...

We have already visited NIST. To implement the AI RMF, they established at AI Safety Institute (AISI) at NIST AISI in the US, and we have already had a working relationship with this AISI; the NIST director have visited us here, I have visited her there and we’re now in ...