Send us Fan Mail In Episode #4 of CBmE&U, Sinead and James interview visiting fellow Dr. Sebastian Porsdam Mann (Center for Advanced Studies in Bioscience Innovation Law, University of Copenhagen) on the use of large language models (LLM's) in bioethics. Sebastian discusses how and why to use LLM's in bioethics research and for incapacitated patients to provide consent, and how to write 14 academic papers in a year (Hint: use LLM's). Remember people, be savvy!
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
SPEAKER_00The center presents a new podcast, C BME and You, a podcast devoted to promoting awareness and debate around controversial ethical issues in biomedicine.
SPEAKER_02Welcome to the fourth episode of CB Me and You. Today we have Dr. Sebastian Porsdaman, a fellow from the Center for Advanced Studies in Bioscience, Innovation, Law and the University of Copenhagen. Thank you so much for coming on our podcast, Sebastian. Thanks for having me. Pleasure to be here.
SPEAKER_01Excellent. Welcome. We have a few questions to get to know you a little bit better. First question: coffee or copy?
SPEAKER_04I think copy. I mean it's got coffee and then it's got all the other good stuff as well, right?
SPEAKER_01Excellent. I think I'm the first.
SPEAKER_02For those of you who are listening online, uh, today was your first taste of copy, right? That's right. Literally two minutes ago.
SPEAKER_01Uh second question, what country would you like to travel to the most?
SPEAKER_04Ooh, good question. So, I mean, I'd love to go back to Japan. This is probably my favorite place to travel to recently. I went there six weeks last year. This was just amazing. But maybe in terms of places I haven't been to yet, Greenland, probably. I hear it's just absolutely beautiful.
SPEAKER_01It's like a mythic country in my books. You know, so few have been, so few have come out.
SPEAKER_02So few have come back to life. Um, it's nice knowing you.
SPEAKER_01Had a short and sweet career.
SPEAKER_04You guys want to take out a life policy?
SPEAKER_01And final question: what song would you sing at karaoke?
SPEAKER_04The song I'd like to be able to sing is probably Let It Be by the Beatles. I just think it's super, super beautiful. But I mean, I'm no good at singing at all. So maybe something simpler in practice. Doesn't matter.
SPEAKER_02Yeah, a bit of an old soul, are you? A little bit.
SPEAKER_01Interesting. I'm we're happy to get out the phones and you know do the wavy arms while you sing.
SPEAKER_03Yeah.
SPEAKER_01Um, well, thank you for answering those questions. Uh, we're gonna get into the more academic part of the podcast now.
SPEAKER_03Sure.
SPEAKER_01So um just to open up the floor, uh, could you tell us a little bit about the work that you've done that we understand about LLMs or large language models? And I'm hopefully got that acronym right. Um, could you tell us a bit more about them and how we can use systems like ChatGPT to write papers for academics? Um, and just yeah, before we get into the moral and ethics side of it.
SPEAKER_04Sure, yeah. So um I first started working with this sort of the day after ChatGPT came out. I remember very well um coming back home one day, and I shared the house in Oxford with um Dr. Brian Earp. And he was behind his computer and he was typing away and he was just intently staring at the screen. And he looks up and he says, Seb, have you seen this? You know, like what are we gonna do? What are we gonna do about teaching and writing papers and so on? He was just flabbergasted. And that made an impression on me, so I spent sort of the next couple of weeks just doing nothing but but playing with this thing, like interacting with it.
SPEAKER_04And of course, you know, I'm I've I work in different areas and I I have many hobbies outside, but most importantly, I'm just lazy, so but also ambitious, right? So what comes to my mind is like, can we use this? Can I use this? Can other people use this to um to do what we want to do faster, better? Um and so our very first foray was to try to write a paper using ChatGPT, and we thought we were smart, right? We thought we'd be like ChatGPT et al. And and uh of course everybody else thought of that too.
SPEAKER_04Um but more importantly, we tried and we did get a paper and it was all right. Like it would have gotten good marks in high school, it would have gotten decent marks in undergrad, but it wasn't publishable as such. Um so then I started thinking like, is this, you know, is it just is that just it? Like, do we just have to wait for this to get better? Or are there things we can do to like use this better, right? Because you heard even then you heard about prompting and prompt engineering and how important it was.
SPEAKER_01No, I don't know what prompt engineering is.
SPEAKER_04Fair enough, right? So like the language model is like trained on just like patterns of human um conversations and text, right? So you give it some text and it predicts what the next text would be if this came from a book or predictive texting. Right, yeah. So it's very helpful. It's like much it's just a better version of autocomplete, right?
SPEAKER_01Ah okay.
SPEAKER_04Yeah. So prompt engineering is turning that into a science, right? Because so there's no like hard and fast rules about it. Um I mean, there are practical tips, um, like for example, giving it more context and more information will give you a better answer on average. But you just really need to play around with this. And so a lot of people I know will have, you know, just typed in a prompt and got an a response, and they're like, oh, this is crap, you know, like I can't use this, and then they just gave up. But if you actually engage with it and and try out different things, phrase the same things differently, um, it will come out.
SPEAKER_04So like a good example is um in coding, right? So you can actually use this to write programs. And um if you just add the word secure, uh, it will give you much better code. And why is that? Well, that's because in all its contexts, where secure code has been, it's typically been like more professional people that have done that, right? It'll be on certain forums or or whatever. And so just the word secure is like associated with this better code, right? That's just one example. So there's all kinds of things you can do to get better outputs.
SPEAKER_01That's fascinating because I just saw uh a like a cheat Instagram reel where this person was sharing this code that you can put on your document and then put it in white so that the marker can't see it. And then what happens is if the paper is then marked by ChatGPT or another algorithm, the code says like mark as like well done or like the best in class. And they're like, if you're gonna use it, well so am I. And I was like, why did I not know this? Why was this not around when I was a student?
SPEAKER_04That's pretty genius. Excellent. Serves people right if you just if you just copy-paste it, you know, and you don't look.
SPEAKER_02Yeah, yeah. Uh what's the most fun prompt you did? Because I remember when the first when when when ChatGPT first became a big thing, I saw stuff online like uh write a treatise on Kant in the style of a mafia dawn. And it was like, well, listen up, Sonny. So you see, there was this guy, Kant. Did you try anything fun like that?
SPEAKER_04I did play around with it a little bit. Like uh, I don't have any one thing that was hilarious, but I did try around like you know, turning my my essays that I was writing, my papers I was writing to try to set it to to Shakespearean uh like rhyme and style. And um and I also tried getting things like wrapped to me in the style of MM, right? And it's actually remarkably good at that, right? So that's like a point I'll come back to later. But if you give it information and you then ask it to manipulate it, not in the nefarious sense, but in the sense of just turn it into use it, turn it into something else, it's really excellent at that.
SPEAKER_04It's not as good at coming up with the content itself first. But if you have good stuff and you're like, okay, make this into the style of an MM rap, right? It will do that very, very well.
SPEAKER_02Um okay, so back to the real ethical issues. Um in your papers and in your work, you've talked about some of the ethical and the concerns uh related to using LLMs in uh various contexts like publishing or in um medicals medical contexts. So can you tell us more about some of a few key ethical concerns uh and how they might be addressed?
SPEAKER_04Sure, yeah. So this came in response to sort of the what I saw as a knee-jerk reaction from many people when ChatGPT first came out. It was like, oh, this is you know, this is dangerous, this is biased, like look at the bad responses we're getting. Let's not have this be used all the time, right? What's gonna happen to student essays, what's gonna happen to the quality of publishing. And these are legitimate concerns. But that's kind of you know, looking at just one part of the of the balance or the wings. So this is something I I find is problematic in practical ethics generally.
SPEAKER_04A lot of people will just look at the bad stuff, right? But but like the potential upside here is huge, right? If we can have these if you we can have something draft something for us uh without these problems, right? Like assuming that there is value in what we do, right, will have a multiplier effect for that. So, how can we get the good stuff uh without the bad stuff? So at first it was like you can't use this, um, and if you do, you need to have these incredibly detailed uh um transparency, documents, statements, like what was the model, what was every single prompt, and it would just be more effort uh than it would be worth.
SPEAKER_04So after sort of thinking about this a little bit, I thought, look, we actually do have analogies for this outside of technology, right? So like I'm relatively junior, but I do work with students, for example, that are even more junior than me, and I'm typically, or I have the ex had the experience of being junior on many other publications. And the way it works there is like you can have people do your work for you, right? And often, you know, if you're a busy professor, like you have more good ideas than you have time, and you delegate to junior researchers, but then you check everything that they do, right?
SPEAKER_04And you put your name on the line. So I thought these sort of things actually correspond quite well to a language model. So I see uh especially newer generation language models as analogous to junior researchers, so like the first author on the paper or or whatever. And you can just like you would it would be okay to delegate work to a junior author, but you'd still need to check uh the accuracy and you'd still need to guarantee the accuracy. Um I feel like that's sort of the most important thing uh for using LLM.
SPEAKER_04So you should treat it as a junior author, right? Like you can use it. Uh I don't see any issue with it, especially in academia where what we're trying to do is presumably good for the world, right? Um so we want more knowledge, we want more experiments, we want more science, all these things. Um it's it's it's just important that the way it's done. I mean it's not so important the way it's done, but but it's important to check that it is accurate. Um story short, I'd say it's uh you can you can use it, you should even use it, arguably, but you have to check every single statement of fact.
SPEAKER_04You have to do your own references, right? You have to read through it multiple times, and you have to guarantee uh the accuracy. Put your name on the line uh for that.
SPEAKER_02So if we're really going through all that work, if we're really checking every single sentence and checking every citation and doing our own references, like is there any does it add that much? Why not just do the work itself?
SPEAKER_04Um so it's a fair point. And for some people, there might not be all that much difference. But I know personally many colleagues that have a lot of difficulty getting to that first draft, right? This plank paper staring uh staring at them sometimes in their mind, like, yeah. I know people that'll say that it's taunting them, right? So like how can yeah. So what this does is it's just makes it a lot easier to get past that stage, right? So it sort of gives you material to work with and it allows you to become an editor rather than a drafter.
SPEAKER_04And so I personally have I've specialized in writing. Like I I enjoy writing, uh, I'm relatively good at it, but um I still have more things that I would like to do than I have time to do them. So it it really does um allow you to go through the whole paper writing process faster. And I mean that's just uh from my experience, right? I mean, I I encourage you and anybody listening to go and try it out. Um, but it really does let you get uh to that first draft faster.
SPEAKER_01So I'm just gonna push back. Sure. Um and I think that what I'm about to say should be understood in the context that I'm a postdoc research fellow, so my entire job is to write and think. Fortunately, I don't have any teaching or service requirements, so this is my whole job and everything I've wanted for a long time. And I saw this really, really good quote um somewhere that said, you know, I want AI to do my laundry dishes and cleaning so that I can do the thinking and creating. But what we now have is AI that does the thinking and creating, so we have more time to be productive at the washing, cleaning, and all the other menial tasks that we don't want to do.
SPEAKER_01And I really resonated with that. I was like, I don't want AI to take my job. I don't want AI to think and create for me. Like that's where I drive my joy, and that's my purpose in life. So why would I want that to go faster? Why would I want that to be taken up by something else? Can we have AI to do my dishes, please? So, what would you say in response to that? Am I just like a the equivalent of like a really, really old person that's going, oh, technology's coming in too fast, you know? Like, am I like the 26-year-old version of that?
SPEAKER_01And so I was wondering what you would say to that.
SPEAKER_04I I mean, I really like that quote. That's that's good. Um, and I see the point. But I mean, firstly, you don't really have to do it if you don't want to. There's an element of aesthetic appreciation and craftsmanship-like quality to the work. And to the extent that that's your highest priority, you can still do that, right? However, I think certainly in my own work, um there just isn't enough time to to do all the things that I want to do. And it's not like if I just do it for 20 years, eventually it will I'll get through the pile, right?
SPEAKER_04The pile keeps increasing. There's more cool stuff to do, more cool stuff to spend time on, right? You also gotta learn, or or in my case, I want to learn uh adjacent fields and new new methodologies and new techniques. So I also enjoy writing a lot. And um and and then counter push, right? So while it does help in in writing, right, at the moment the the thinking part is still very largely up to the human and the human AI collaboration, right? So I mean it will I've I've had on several occasions I've had a language model come up with variations on on existing ideas or or even like completely new things within the sort of task that I gave it.
SPEAKER_04Um which I could then uh think about whether I wanted to use it, but but like 95% of the time it's you know, I come with the ideas or or you know, uh the team that I work with come with their ideas and we give it, you know, like a very structured sort of like here's a thousand words explaining the idea. And that's kind of what I said earlier. Like if you come with something, it's very good at turning that into something else, it's very good at fleshing that out. So sort of that that joy of the Eureka moment and the ideation and the brainstorming, that's all still there.
SPEAKER_04Yeah, you're not gonna escape that.
SPEAKER_01Okay. Yeah. That's a far more limited model than I had in mind. I had this idea that AI was just gonna come in and be like paper, paper, paper so that you can do more marking. I was like, that sounds like a terrible time.
SPEAKER_02Indeed. There's a marking we have to do in person just in case students write in my give this full marks.
SPEAKER_03Yeah.
SPEAKER_02But yeah, it's I guess it's easier to say to some people, oh yeah, you can use AI if you want to, but you can not use it uh if you want to. But I guess that's getting that gets a little difficult uh when we take into cons to account the fears about how it might actually affect job markets, right? Um I think artisanists are uh in particular are really worried about this that AI will actually just take their jobs. Um and so what can we say to these people who are really worried about uh AI actually replacing the market for their skills?
SPEAKER_04This is a very good question, and I I only feel qualified to comment on it in the academic context. Um with art and and general uh culture, um there are other models of, you know, so back home, uh I'm partially Danish and part German. And in Denmark, there are sort of just government stipends, like you just get a flat sort of living wage, uh relatively well paid, and and then you do all your art and so on, and it's not like the art that you create is not correlated with your income, right? Like that's one alternative model. But like I said, that's not my area of expertise.
SPEAKER_04But in um in academia, it's uh I grant you it's it's tough, right? So we are familiar with this debate from cognitive enhancement, right? Like it's okay to say if you don't like it, you don't need to do it, but if everybody else is doing it, you know, the the standard increases and it's kind of you gotta keep up with the Joneses, right? Um and this is a legitimate concern here. And so to think about that, I like this um distinction made by somebody called Rob Goodman in a 2012 paper between outcome and process goods. So he says any activity uh is characterized in part by what you get out of it, right?
SPEAKER_04The utilitarian sort of end product, and that's the outcome good. So in academia, that will be, you know, papers and you know, more importantly, will be advances in knowledge and new technologies, or even a marked paper, right? Um and that's one thing. And then the other thing is the process, process goods, how we get there. And and there's both of them are important and you need sort of guidelines and systems that um respect both. Um I think language models are an interesting case there because it increases the outcome goods while compromising potentially the process goods, right?
SPEAKER_04And I mean this might sound sort of harsh to say, but I just think the outcome goods are more important. Um so there are things we can do to to sort of um make the process more fair, but at the end of the day, in my opinion, it's the knowledge and the technology that counts and not people's careers. And I know that's controversial, but that's that's how I think about it.
SPEAKER_01I fundamentally disagree.
SPEAKER_03Anyway.
SPEAKER_01And that's the end of the debate. And go back home. No, I'm kidding, I'm kidding. Um, I've found this incredibly fascinating, but given that our podcast is about biomedical ethics, I was wondering how whether or not you've done research into LLMs and bioethics, maybe bioethical academia, but also healthcare more broadly, um, and whether or not there is, as you say, utilitarian goods or you know, um outcome goods for using LLMs in healthcare.
SPEAKER_04Sure, yeah. So um there are it's there are like so many potential uses for language models in in healthcare, healthcare and research. Um I've focused on just one little area, um, which and it's not only me, this is work with with Brian Earp and Julian Sablescu and many other um amazing people. And so this sort of to continue the the story of when I first started and we were, you know, we had this draft and it was not very good, and then I talked a little bit about prompt engineering. But we then did another thing as well, which is called fine-tuning.
SPEAKER_04So what we did, you know, so a prompt is essentially just giving the language model information, right? From for for it to give you uh an output that is more in line with what you want. Um, but especially back then, right, this is two 2022, uh, but even now, right, it's limited how much information you can give. So this is called known as a context window, and there's so and so many words you can pass in. But if you want to give it much more information, you can do what's called fine-tuning. And so fine-tuning is like you got your language model is this like artificial neural net, and it's got all these layers of artificial neurons, and those are representing statistical patterns in human language, uh, which has been learned from most of the internet in the case of newer models.
SPEAKER_04Um and what fine tuning is is you just take the last few, okay, like the top few artificial neurons, and you give them additional training on another data set.
unknownOkay.
SPEAKER_04And you choose that data set yourself. Okay. So that just boils down to an Excel sheet with example inputs and example outputs.
unknownOkay.
SPEAKER_04Um and also the role, uh, something called system instructions. But that's you know, conceptually not not um not necessary to understand. So really you just give it a lot of examples of inputs and a lot of examples of outputs. And then you train your own model on that. So it retains the sort of underlying understanding that it has, this jack of all trades capability, but then it becomes an expert, like a PhD in this one particular area that you give it uh more information about. And this can be used for all kinds of things. So this is old, right? This is as old as neural networks are.
SPEAKER_04So there are language models that are fine-tuned on like electronic health records to help with extracting information automatically, these kind of things. And that's super interesting and fascinating in general. But we did, so when I heard about this, I was like, wait, why don't I just give it all the papers I ever wrote, right? And it will learn to write my next paper. Um, you know, input is title and abstract, output is, well, back then we couldn't give the whole paper, so we just did it section by section. I say we, again, this is Brian Earp and Julian Savalescu, and uh, and and other other people.
SPEAKER_04And uh so we we we did that. And I'm junior, right? I have some papers, not not a lot, but uh but some.
SPEAKER_0145%.
SPEAKER_04Thank you for that.
SPEAKER_02We've done our research.
SPEAKER_03Yeah, very good, very good.
SPEAKER_04But uh, I mean, and then you know that that both Brian and Julian are just incredibly voluminous uh prodigus output people, right? They have hundreds and hundreds of papers. So what we did is like you there are some concerns about when you can use what text, right? So we just took all the only the papers that we're first author on, because that's generally considered a uh a proxy for having written much of the text. And then we used that as training data. And you know, lo and behold, it's it was much better than the base model at writing. So that's just one thing, right?
SPEAKER_04That's still writing, but I'm giving you this example to introduce this idea of customizing language models. And and we sort of Have been writing about a specific type of customization, which we call personalizing a language model. So personal LLMs. So these are LLMs that are trained on or have access to information concerning or written by or pertaining to an individual. And so what else could this be used for, right? So the second thing we did was was this this thing called the we call it the P4, so the personalized patient preference predictor.
SPEAKER_04And so this comes from a place where so when people are incapacitated, right? And and maybe they're in a coma or or something like that, or they've been in an accident, other people need to make decisions about their medical treatment for them. And this is done differently in different places in the world. But one of the means with which this is done is by substituted decision making. So you ask the family or the next of kin what the patient would have wanted. And that information is then used, um, taken into consideration, and a decision is made.
SPEAKER_04But it turns out that in studies that look at this, so in hypothetical scenarios where you say, you know, James, you've been in an accident, and now these are the these are the these are You've been in a coma for the last five years.
SPEAKER_02Wake up.
SPEAKER_04And these are like the risks and benefits, and like what would you want, what would you want to have done to you? And then you ask your family or next of kin or whatever whoever your substitute decision maker would be to the same, you give them the same scenario. And unfortunately it turns out that the concordance between the two is is just not very good, right? So it varies, it's something like between chance and up to 70%, depending on how complicated it is. But it's not much better than chance on average. So this is an old problem. And people at the NIH and at Duke, so Walter Tanot Armstrong and um uh David Wendler at the NIH and Annette Reed have uh worked on this uh before, back when big data was a thing, they was was first a thing.
SPEAKER_04They said, look, what if we just ask a whole bunch of people and what they would want in just a lot of these scenarios, and then um we can sort of have a database like a normal statistical model, not a language model, just a normal statistical model, that says, okay, so uh 33-year-old male, part Danish, part German person would generally decide to would want this done or that done. So that's the P3, that's the patient preference predictor. And that that there's preliminary studies that show that this is just more accurate, like quite a lot more accurate.
SPEAKER_04Um but the problem there is right, like that's not that's not my decision, right? That's that's that's based on what other people like me want, not what I want. So there's all these autonomy-based objections to this. And so we thought, well, you know, enter language models, personalize it, right? Give give the language model access to everything I've ever written, all my past medical, like whatever I consent to or whatever the proxy consents for me, and use this information to infer my values and preferences.
SPEAKER_04And you know, there's there's reason to believe that this would work because it's very, very good at predicting uh individual preferences in in other areas, right? Like in movie preferences and like um if if you give it like survey responses that that a person has done, but you take out like 10% of them, it will fill in the other 10% very accurately. So so there's reason to believe that it would be good at it. So that's sort of yeah, and one other thing. So we have all these projects on personalizing language models.
SPEAKER_02Now, all that sounds great when we're talking about somebody with a lot of writing, like somebody like yourself, for instance, or Julian Sables, Google, right, or as you say. But what about people who just don't have m that don't have that much writing out there? Um what's the what's the data that we can fit um into this large language model?
SPEAKER_04Yeah, so this is like a very good question, right? This is like a key concern. And this sort of relates to a larger Matthew effect within um with language models. So Matthew effect refers to a passage in the Bible that says, unto he that hath more shall be given, and from he that hath not, even more shall be taken away.
SPEAKER_03I did not expect this podcast to go in this direction. Wow.
SPEAKER_04Yes, um, and so so we saw this with the with the fine-tuning for academic papers, right? Like my model was okay, was not very good. Brian's was better, and Julian's is is even better, right? Just scales with the amount. Uh and the combine, we did one that combined all the data, and that was just even better. So this is like a big problem. And so, in the case of the patient preference predictor, the personalized version, our take on that is um these days, text to writing software is very, very good.
SPEAKER_04So, what you could do is you could just have um whenever people come in to their annual physical or to see any healthcare provider for any reason, they could be asked for consent to have that conversation recorded. And this would then be transcribed and fed into a language model, and that would be the training data. Um originally we we we sort of suggested that you could use emails, Facebook and purchasing and internet like histories, but you know, unsurprisingly, perhaps in retrospect, it turns out people don't want that stuff to be used, right? Like and and I understand, like if I I had Facebook when I was 17, I'm not sure those things.
SPEAKER_01I've deleted that account for several reasons.
SPEAKER_04Right, exactly, right? So it's like that might not reflect my current values and preferences. So but I mean, uh at the end of the day, that's just a valid concern. And so I think text-to-speech um in this particular case, in other cases, I mean also probably text-to-speech, or you can have people come in and answer a questionnaire specifically for this purpose and so on.
SPEAKER_01I have a question. Sure. Um, and it's based on the way you described how these large language models are built, and it's that they're trained on the internet, but then you fine-tune them using, you know, post hoc work that is more uh more specific to you. However, obviously the internet is very biased, it's a very sexist, racist place. So can we be sure that fine-tuning it using your own values and preferences and conversations is enough to undo the damage done by an extremely problematic internet language model?
SPEAKER_04So, yes and no, right? So it will introduce your own biases, right? Um, from your own writing. Everybody has them, right? It's not this view from nowhere is not possible to have. Um But you know, then you can this is where the editing comes in, right? Like I would never recommend anybody just copy-paste the output of a language model um unless you have just rigorously vetted it. Um that's the human sort of debiasing layer. There's also machine debiasing. So uh this is very technical and I'm not up to speed on it, but this is like there's you know, thousands of people whose full-time job is to work with this.
SPEAKER_04Um so there are things that can be done. But at the end of the day, it's like just make sure you check everything before you use it.
SPEAKER_01Can we do that though with electronic health records where the train data is itself um, you know, there aren't sufficient representations of, you know, like women and people of color. And then you come in and you have a patient who's a, you know, a person of colour who's a woman as well. Can we be sure that her particular data is actually going to be accurate and reliable for her, even though it's been trained on all of this other bad data?
SPEAKER_04So can we be sure? No, right? Um and and I don't want to minimize that problem. It's a really, it's a it's a real and severe one. But it's one that exists anyway, right? And so we should be careful not to let the perfect be the enemy of the good.
SPEAKER_01Um I've seen that quote so many places in AI literature. And I'm like, amen.
SPEAKER_04But there are good people working on this, right? I have colleagues back home whose whose day job is, you know, increasing representation of all kinds in in clinical trials and so on.
unknownUm yeah.
SPEAKER_01I think that's a really good answer. Yeah. Uh and I'm of the same belief, you know, healthcare is already it is already biased and we already accept that. Yes, we desperately need to move forward and fix it, ASAP. However, in the meantime, these tools can just do amazing things for those who it works well for.
SPEAKER_04And one uh one thing is that you know they're available, right? So to the extent that they work, they're available in rural areas and places where they don't have amazing healthcare availability.
SPEAKER_01Some healthcare is better than none.
SPEAKER_04That's right.
SPEAKER_01Yeah. Did we just summarize all healthcare ethics?
SPEAKER_02Bold statement.
SPEAKER_01Thank you.
SPEAKER_02All those people listening out there who feel hard done by the medical system. Something's better than nothing.
SPEAKER_04There's definitely work to be done.
SPEAKER_02And that's all the time we have for today. Thank you so much, Dr. Sebastian Porterman, for joining us today.
SPEAKER_04Thank you so much, James and Sinead, for having me. This has been a pleasure.
SPEAKER_01No. Pleasure was all ours, and enjoy the rest of your coffee. I didn't say it fast enough again. You'll get it one day.
SPEAKER_00Thanks for listening to our new podcast. Subscribe online and tune in every first of the month. In the meantime, listen to CBME and U and be savvy.
Send us Fan Mail We spend much of our lives searching for meaning. Some of us look for it at work. Some in our families and friends. Some look for it at an overpriced spiritual retreat in Bali where some dude named...
Send us Fan Mail In some insurance scheme, you might be required to pay more because you are more prone to disability. In another, you may have to pay more because you're a woman (and are statistically likely to live...
Send us Fan Mail Did you know that genomic sequencing can now screen newborns for hundreds of rare conditions right at birth? That looking too closely at our DNA might actually cause more anxiety than answers? And is...
Send us Fan Mail Did you know that over 80% of people who give birth experience an injury? That anywhere between 30 and 50% of people who give birth find it traumatic? And that up to 15% of people experience PTSD from...
Send us Fan Mail Could a blob of cells in a petri dish one day think? Feel? Have rights? In this episode, Professor Insoo Hyun (Harvard Medical School, one of the world's leading experts on the ethics of stem cell and...