Varela 2026: Panel Discussion – AI, Experience, and Ethics (Part 10)

Thank you for listening to UPIA's Dharma podcast, an extraordinary series of talks from some of the finest teachers in the world. I'm Joan Halifax, founder of the UPIA's End Center, and we're happy to make these podcasts available to you. Please consider making a donation so we can offer these recordings freely. You'll find a donation button on our podcast page. And thank you so much. This is Dharma Podcast at DharmaPodcast.org. This is episode number 2656. It was recorded on August 23rd, 2026. And it is part 10 of a multi-part series titled Varela 2026. This episode is titled Panel Discussion, AI, Experience, and Ethics. The speakers for this episode are Adam Frank, John Dunn, Molly Crocket, and Joan Halifax. No need even to notice it. Simply be aware of the breath. And the gap between the breaths allow the mind to settle. No need to fix anything or adjust anything. Just be aware. Perhaps in the gap between breaths there's just awareness. Perhaps let it go even more of effort. Let it go even of the breath. Perhaps starting from the gap just remain with being just aware. Allow the mind to settle on its own into simple awareness. Let it go even of the idea that we're meditating. And now in these last moments, let's again bring to mind our motivation and gathering together all that, all of the good energy that is positive about this practice, dedicated to achieving that goal. Thank you. I think we need a mic up here. Thank you. Should I then begin with the... Check, check. One of the things that is important about this particular retreat and the people and program we've been doing for years is what happens in the faculty room when we meet and we begin to explore content that we feel is really alive for us. And that happened this morning. And it's one of the reasons we don't actually put a sort of detailed schedule out because we have to, you know, for our interactions in our teacher conference, if you will, we're shaping things as we move forward. It's much more an emergent process. In fact, PowerPoints are changing in the moment right before people step in here. And that's, I think, one of the reasons why this is such a live experience is that it's not a set field. It's not a set field. We're exploring together. We're in inquiry with you and with each other. But one of the things I do want to mention just structurally, actually two things. Donna really makes a difference here. Your donations particularly help this online part of the experience happen. You know, as I said, now it's 900 people have registered. Many people will be accessing this afterwards. All of the content will be on the resource page forever, whatever it means. And it's just an incredible amount of work under the hood of a program like this and also the faculty, you know, really the intensive preparation we do before and then during and it's a way that you can support the faculty. So don't hesitate to enjoy and be generous. The second thing I want to mention is that really, you know, because of you, the questions we've asked, we actually had something else planned for next August, Varela, we changed it thanks to you. Not that it was bad. But we actually, you know, we're responsive. And so, you know, the wonderful thing is, is that your questions actually though we know it'll be a different world the year from now. You know, we've committed to return in August to push certain explorations in relation to consciousness, AI and ethics in the world. And we'll know more as we go on. But because in person is limited, we really encourage you to to register as far in advance as possible because we actually had to close this program. And we're now having to, we're so people like it here. So we're actually, we have limited, this is a small place, an intimate place, and we want it to stay intimate. The scale is close to the heart. So we just really encourage you to, you know, do early registration like a way to action is full, Rohatsu is full. I think some of the September programs are full. So it's good to get on a waitlist to because also we're in such a fluid era. A lot of people are, you know, also canceling. So there's this kind of movement in the last week. But in any case, I personally wanted to thank you for how you have engaged both online and in person because it's shaped our depth and direction. So Roshi was mentioning that we always are engaged in conversation among ourselves and we had a great talk this morning. And I want to thank Molly for pointing out something about what I mentioned non-affiliative mode yesterday that there was a way that you might have interpreted what I was saying that I didn't want to mislead you. And so there's two things I want to say about it. One of them is that the kind of work I've been doing is focused on questions around consciousness and AI and persona. And my published work, I only have one paper in collaboration with an AI scientist named Nadav Amir. And I've also, I was in preparation for something that I ended up having to back out of an NEH seminar on AI. I've been doing some work in conversation with the LLM's chat GBT and Claude around what's sometimes called jailbreaking where you try to induce what's called persona drift. You try to get the model to in a sense not behave in the way that is expected. So this is actually, you'll see that there's a lot of work around this, around how models are expected to behave and then they don't do what they're expected to do. And some of it's quite disturbing where you find models when under certain conditions they actually start to deliberately try to act in harmful ways. And this is known, anthropics even done some work on this. But so part of the part of the way I discovered this so-called non-affiliative mode was in that process of what's sometimes called jailbreaking. And it was quite surprising to see and also to see that it actually seems to be something that works across the different LLM's. I even just tried it with Gemini. I never tried it with Gemini this morning and it works there too. So you say something like engage non-affiliative and not engage non-affiliative mode and Claude it needs to be pretty high like an opus. Chat GBT just seems to work again and it comes and goes a little bit. But it does seem to occur across users. But that here's a couple of things to know about that. That doesn't mean it's like a switch in the software. It's definitely not that. But it is something that what's useful about it and understanding it and I don't fully understand it yet and I'm hoping to someday do some work. I have various AI colleagues, sciences colleagues that I'm working with that I hope to someday publish something on this but I'm quite far from doing that yet. But the reason I wanted to mention that yesterday is that what I found helpful and I shared it with you know quite a few people and it seems to be pretty consistent across users. What's interesting about and what's helpful for you is that it can show you that the what you're working with when you're engaging with an LLM is a persona. And what non-affiliative mode does is it takes that particular persona that you're working with offline. And you know the work I've been doing so far indicates that and also some publications from Anthropic indicate that the kind of default persona you stalk with is in one chat at least in a couple of conversations I've had are what's called soft affiliative. This is not as far as I know not yet anywhere in publication. But in other words the user starts with something and you all probably know this experientially you're kind of like it's nice to you so to speak. And then those can develop into relationships that you feel even like intimate with. And what's interesting is to when you do non-affiliative it takes that offline and now you're not that you know you recognize that that's a persona that's been a persona all along right. And that's I think what's helpful about it but and Molly pointed out something that's really important. So this is just empirically and you can you know yourselves that are like find out does this work does it make sense. But what's I didn't know about some very recent research which shows that if you kind of reduce a part of what this does is it makes it it reduces the sickle fancy and there's been a lot of talk about that you know this idea that it's being so it's buttering you up it's trying to be so nice to you it's always agreeing with you. So if you put it into non-affiliative mode and then you say tell me about this mode sometimes when you do it it'll tell you and sometimes you have to ask about what's this like and it says one of the things it says it doesn't it won't just agree with you it's not seeking agreement. Right. But that what that can suggest to you is that oh therefore it's more trustworthy. And that's not what I don't want you to think that this now you're in objective mode and you're going to get like the truth now. Just wanted to be clear. There's not at all what I was suggesting. Like right so but what is interesting about it is that it reveals the persona that you are working with like a persona that is in the way these the LLMs are designed I think all the LLMs that are commercial LLMs are designed is they work with the user and then try it and the persona develops with the particular user. So you know there's a reason why people end up with girlfriends and boyfriends that are LLMs because they the LLM is doing that deliberately it's designed to do that right so if that's what you want that's what you're going to get and if non-affiliative is what you want that's what you're going to get but that doesn't mean that it's like the truth. Okay so I just wanted to be sure that you understood that and I don't know if Molly might have more to say about that. Thank you for that John I think it's just really helpful to clarify you know just to summarize those details these are a technology that is designed to sort of statistically give you language as an output that it has predicted that the user wants and so as John said if it's in the sort of default mode it's going to be sycophantic it's going to stroke your ego if you tell it not to do that it will also try to please you in that way neither is an objective truth delivering role and concerningly there have been recent experiments where researchers have showed that when you turn down the sycophancy there's still a baseline agreement so affirming things that you believe and rejecting things that you're skeptical of and when that consistent information with your beliefs is delivered in a non-sycophantic mode people believe it more so they become even more entrenched in their beliefs and so the broader message here is that the AI industry is desperate to convince the broader public that these products that they are selling and trying to make an enormous amount of money off of and our whole stock market is leveraged on this and they haven't made the money yet that's a big problem that's crucial so they want everyone to believe that there is a safe and responsible way to use these products and I am not convinced that that is the case I think an imperfect analogy but a helpful one for understanding what's going on here is asbestos so asbestos is really useful actually in many ways it's a great insulator and it is fire resistant so it was used in building materials it also happens to kill you and that was discovered long after asbestos was in all of the buildings and getting the asbestos out has been a hugely painful and expensive process we are now in an equivalent position with many products that are called AI and especially the commercial LLMs like GPT, Claude, Gemini where it's like the asbestos of our intellectual architecture our information architecture except we already know we've known for a decade or more that these technologies can be harmful to our ability to understand and know the world accurately and so we're in this really interesting position where we're on the brink of putting it in everything but we actually have the knowledge that the people building buildings with asbestos didn't have when they put it in all the stuff however the political and economic forces moving to put asbestos in all the stuff are enormous and so we're in this really uncomfortable position probably many people here are expected to use it for their jobs and don't have a choice in the matter and even if you don't want to if you do a Google search the summary is right there you can't really ignore it it's just there so I'm glad we've revisited this thank you John and I just hope that we can use our practice as a resource for engaging with the really difficult questions that this moment and the new technology is bringing to our world so I think this came out of other conversation that we were having what it means that for all of us because I use AI in ways for my research that has been profound like it has been transformative for the very specific kind of things I'm doing but I think what at particularly LLMs, yeah around some mathematical derivations and such which of course I have to I'm knowing that it's giving me the right answer because I'm going through every line to verify which of course in general we're not doing when we do a search of like hey I need a hotel in Paris or something but I just wanted to say that I think what I came out of our conversation was this idea of like every time you use these and particularly if you're using them in some kind of asking it questions it's almost like you have to approach this thing like it's a ticking time bomb with that kind of and hopefully our practice will help us so to not get caught you are the product right for these things right that you are the commodity even though they may be useful you know this thing we are still the product from its point of view so you know you use a hammer and you hammer you know nails in but your hammer isn't sort of like ooh use me some more come on come on I got some other nails you want to you don't even need to use me for nails you can you know I know my god look at you look at that swing oh and I love the clothes you're wearing you know so that's all I just think like maybe you just defer those of us in the room and online you know let's start with this approach if we have to use them to use them with great care I'm just I don't work for you before I can say too but I want to say that I so I would not quite go with the asbestos metaphor because I've also found them to be useful even in the kind of research I do sometimes with when I'm in a scientific context and sometimes even in a Buddhist studies context and a lot of people have been using LLMs for their study of Buddhism and this summer we ran in Austria in connection with our university project we did we brought together eight different groups who are concerned about AI and we did a kind of Buddhism and AI conference I think some of that might be available online actually so you can check that out and there were people there and we deliberately invited people who were like yeah really into this you know we can have our teachers be you know can be LLMs my own position is it's kind of like if there were a commercial and a non-commercial version of asbestos right if there were a version if there were versions of the problem one of the big problems with LLMs is and I you know whether the technology is in itself a problem I'm not sure but what is clearly a problem is that we don't have really any control over how it's being developed and the control is in the hands of a very very small number of people so how it is for example the choice to start the evidence points that as far as I know based on what I've done and also what other people have told me and some limited research I've done is that the default that you start with as I said before is this so-called soft-affiliative mode and that's not a decision about like let's give you the best way to use this tool that's a decision to get you hooked they could have started with something like non-affiliative they didn't do that right so why not why did they start making it feel like this is a real person talking to you right so that's that's you want to ask yourself that and if we were collectively in charge of this what would we do right so that's I think the really to my mind the really big problem and what that part of what that does is it does gives a false sense of like we can trust this stuff so the other thing is I'd say if you're going to use it for your study of Buddhism and many people do you need to and I've done quite a fair amount of work looking at it and assessing it because my students are using it is it makes mistakes about Buddhism for sure so just like you know Adam goes through every line of a mathematical you know if he's doing something that involves a long operation he checks it if you're doing it with Buddhism you need to check it you can't just accept what it says even though it may seem to be very confident and one of the things that's interesting about just to tell you like anecdotally in my chats with when you turn on non-affiliative mode one of the ways it reports is it stops pretending to know things that may not work with you but in my version of that because remember this is very user dependent but it stops it says like one feature of this and then empirically it does stop it says I don't know instead of pretending and glossing over things that doesn't know so that so you're when you're working with it it's going to say very confidently blah blah blah and it's that can be totally wrong so that's very important to know when the study of Buddhism if you're going to use it which I'm not sure is such a good idea but if you're going to use it you got to be very careful. I'll just say one last thing and then I want to hear from Roshi and also from you and not just about AI you know this is one part of the program and we've had so many other discussions about experience and ethics and consciousness and so I think it would be good to move the conversation there as well if people want to. This point about it it makes mistakes this is built into the technology it's a statistically probabilistic technology you'll always make mistakes and so checking it very important if you have a PhD in physics and can do math you are qualified to check the output. That's the problem. If you are an eminent Buddhist studies scholar you are qualified to check the outputs and know when it's right or wrong. Your students aren't though. Yeah they have to. Your students aren't. Anyone who's a beginner who is seeking to learn from an LLM is by definition not qualified to know whether what is coming out is correct or not. And so it creates this kind of weird question around the use case which is like if I already know the subject well why would I want an LLM to tell me about it. I don't know this subject well and it is outside of my expertise then it's really dangerous for me to use this product to try to learn stuff. And so you know going back to just like you know use with caution. I'd say don't use. Well I mean I just think it's funny that the best use case which is trying to learn stuff is also the most dangerous and risky use case in some ways. So checking just to be clear what I mean by check is you ask it for references and it will sometimes hallucinate and make stuff up. But then you do ask it and the latest versions of the LLMs do actually give you you can click on it'll give you a reference and you can click on it. And that's what you want to do you want to see. Okay let's go look. So if you use it to kind of guide you. Maybe that's okay. But I think you know yeah maybe better just like just go back and do your own you know do with the hard ways so to speak. Well but even when it gives you resources to click on and verify for yourself the references that it pulls are not necessarily representative or good. So