Latest / Health Ethics Podcast / Against AI Exceptionalism: Healthcare for the Common Good

Transcript

Against AI Exceptionalism: Healthcare for the Common Good · Health Ethics Podcast

• 0:09
  Speaker 1: Welcome to Health Ethics, a series of conversations with leading
• 0:13
  Speaker 1: experts about ethical considerations associated with health and healthcare, both
• 0:17
  Speaker 1: locally and globally. I am Brian Pilkington, Professor of Bioethics
• 0:21
  Speaker 1: in the Department of Medical Sciences at the Hackensack Meridan
• 0:24
  Speaker 1: School of Medicine. This series, like its predecessor, which focused
• 0:29
  Speaker 1: on the COVID nineteen pandemic, relies on the idea that
• 0:32
  Speaker 1: challenging ethical issues are best addressed by folks from diverse
• 0:36
  Speaker 1: backgrounds practically reasoning together, and complements the Health Ethics series
• 0:41
  Speaker 1: with its panel of leading experts from medicine, nursing, politics, economics,
• 0:46
  Speaker 1: philosophy and the law. Thank you and enjoy the episode.
• 0:58
  Speaker 2: It is my great pleasure to to speak with Alex
• 1:01
  Speaker 2: John London about AI, ethics and health. Professor London is
• 1:06
  Speaker 2: the kN L. Gates Professor of Ethics and Computational Technologies
• 1:10
  Speaker 2: and co lead of the K and L. Gates Initiative
• 1:14
  Speaker 2: in Ethics and Computational Technologies at Carnegie Mellon University. An
• 1:19
  Speaker 2: elected Fellow of the Hastings Center, Professor London's work focuses
• 1:22
  Speaker 2: on ethical and policy issues surrounding the development and deployment
• 1:27
  Speaker 2: of novel technologies in medicine, biotechnology and artificial intelligence, on
• 1:32
  Speaker 2: methodological issues in theoretical and practical ethics, and on cross
• 1:37
  Speaker 2: national issues of justice and fairness. His book For the
• 1:40
  Speaker 2: Common Good Philosophical Foundations of Research Ethics is available on
• 1:44
  Speaker 2: hard copy from Oxford University Press and available in PDF
• 1:50
  Speaker 2: on his website. His papers have appeared in Mind, The
• 1:54
  Speaker 2: Philosopher's Imprint, Science, Jamma, The Lands and BMJ Plus, Medicine,
• 2:02
  Speaker 2: Statistics in Medicine, the Hasting Center Report, and numerous other
• 2:05
  Speaker 2: journals and collections. He is also co editor of Ethical
• 2:09
  Speaker 2: Issues in Modern Medicine, which is one of the most
• 2:11
  Speaker 2: widely used textbooks in medical ethics.
• 2:14
  Speaker 3: Thanks Ian Alex for joining me. Among the host of
• 2:17
  Speaker 3: topics you've written on and you're amazingly prolific, are AI
• 2:20
  Speaker 3: and research ethics. Let's start with the latter, and I
• 2:24
  Speaker 3: think the best book on Research ethics that I've read
• 2:27
  Speaker 3: is your book For the Common Good Philosophical.
• 2:29
  Speaker 2: Foundations of Research Ethics.
• 2:31
  Speaker 3: In case listeners haven't heard, haven't read it yet, haven't
• 2:35
  Speaker 3: heard you on it, Let's take a step back and
• 2:37
  Speaker 3: could you just briefly share what the main contention of
• 2:40
  Speaker 3: the book is.
• 2:41
  Speaker 4: Yeah, great, well, first of all, thanks for reading it.
• 2:44
  Speaker 4: The main idea of the book is that research ethics
• 2:47
  Speaker 4: has been an incredibly successful area of practical or applied
• 2:52
  Speaker 4: ethics because after the scandals of the nineteen seventy into
• 2:57
  Speaker 4: the revelation of Tuskegee in particular, it wasn't us that
• 3:00
  Speaker 4: people became sensitized to a set of ethical issues. It
• 3:03
  Speaker 4: was that the nation was moved to create institutions and
• 3:08
  Speaker 4: structures that would translate those ethical issues into regulation and
• 3:13
  Speaker 4: oversight that would.
• 3:14
  Speaker 5: Have some teeth to it.
• 3:17
  Speaker 4: And it's been an incredibly successful area in the sense
• 3:20
  Speaker 4: that you went from a research enterprise that was primarily
• 3:24
  Speaker 4: carried out in prisoners, people who were not free and
• 3:28
  Speaker 4: who were subjected to science, to use Susan Letterer's term,
• 3:32
  Speaker 4: you know, so to go from this enterprise that is
• 3:35
  Speaker 4: primarily carried out in prisons to one in which three people,
• 3:40
  Speaker 4: yes they have you know, sickness and disease, but not everybody.
• 3:43
  Speaker 5: Sometimes it's healthy volunteers.
• 3:45
  Speaker 4: Three people are willing to participate in order to generate
• 3:48
  Speaker 4: this the you know, the knowledge that science needs in
• 3:51
  Speaker 4: order to progress. So, on the one hand, it's been
• 3:53
  Speaker 4: incredibly successful, but it was also myopic and parochial in
• 3:59
  Speaker 4: ways that I think limit what it can do in
• 4:03
  Speaker 4: the future. It was very much designed to prevent the
• 4:08
  Speaker 4: kinds of problems that we had seen in the past,
• 4:11
  Speaker 4: which was sort of domination and exploitation and cruelty, and
• 4:17
  Speaker 4: it was there was an assumption built in in a
• 4:19
  Speaker 4: sense that you know, we didn't need to talk about
• 4:21
  Speaker 4: aligning medical research with the common good, We didn't need
• 4:25
  Speaker 4: to talk about aligning research with social value and making
• 4:28
  Speaker 4: sure that the research that we carry out, you know,
• 4:31
  Speaker 4: is scientifically and socially important, because there was an assumption
• 4:35
  Speaker 4: that like well that would be taken care of, researchers
• 4:37
  Speaker 4: would bring that And you know, I think one of
• 4:39
  Speaker 4: the places where we are now is we have a
• 4:42
  Speaker 4: high volume of research, but just about everybody is concerned
• 4:45
  Speaker 4: about the degree to which that high volume of research
• 4:47
  Speaker 4: is equally high value. And search ethics doesn't really have
• 4:52
  Speaker 4: the tools to talk about that because it's primarily you know,
• 4:57
  Speaker 4: participant protection. So the goal of the book is to
• 5:00
  Speaker 4: try to rethink the foundations of research ethics, reconnect it
• 5:04
  Speaker 4: to these larger social purposes, you know, move away from
• 5:07
  Speaker 4: this what I call the IRB triangle, this idea that
• 5:10
  Speaker 4: the point of research ethics is to insert the IRB
• 5:15
  Speaker 4: in between researchers and study participants to protect their interests.
• 5:19
  Speaker 5: That's an important role.
• 5:21
  Speaker 4: I'm not saying we shouldn't be doing that, but that's
• 5:24
  Speaker 4: like a little tiny part of a much larger set
• 5:28
  Speaker 4: of social purposes that research ethics ought to be concerned with.
• 5:32
  Speaker 4: And so the book is trying to articulate that foundation
• 5:36
  Speaker 4: and then show why. Concern about social value concern about
• 5:41
  Speaker 4: the way institutions function, healthcare institutions, public health institutions, because
• 5:47
  Speaker 4: research generates the knowledge that goes into propagates back into
• 5:51
  Speaker 4: those institutions. So we need the knowledge from research to
• 5:55
  Speaker 4: understand novel diseases, know how to treat them, understand, you know,
• 6:00
  Speaker 4: for diseases that impair people's ability to function, what are
• 6:04
  Speaker 4: the best treatments. There's a huge amount of diversity often
• 6:08
  Speaker 4: in the approaches and treatments, some of which might work,
• 6:11
  Speaker 4: some of which might not.
• 6:12
  Speaker 5: How do we know that?
• 6:14
  Speaker 4: That's the goal of research, and so making research aligned
• 6:18
  Speaker 4: with these larger social purposes and helping institutions fill their
• 6:23
  Speaker 4: their social function that's a major sort of reorientation of
• 6:28
  Speaker 4: the book.
• 6:29
  Speaker 3: Yeah, and the book's great, and we won't have times
• 6:32
  Speaker 3: sort of go through argument after argument because we want
• 6:34
  Speaker 3: to get to AI too. And what you've just laid out,
• 6:37
  Speaker 3: I think connects up really nicely to the points that
• 6:39
  Speaker 3: you make about a more just social order. Could you
• 6:42
  Speaker 3: give for listeners who may be less familiar with the
• 6:45
  Speaker 3: philosophical content, could you just give us, sort of briefly
• 6:48
  Speaker 3: in a couple of sentences, what's your take on the
• 6:50
  Speaker 3: common good? I mean, what's your conception of the common good?
• 6:53
  Speaker 3: And how and then sort of how it operates in
• 6:55
  Speaker 3: your thinking about research ethics.
• 6:58
  Speaker 5: Yeah, I think so.
• 6:59
  Speaker 4: A big part of the book is tracing out the
• 7:03
  Speaker 4: common good is a concept that people basically ran away
• 7:08
  Speaker 4: from after World War Two in the United States, partly
• 7:12
  Speaker 4: because you know, at Nuremberg the Nazi doctors who were
• 7:16
  Speaker 4: tried for war crimes effectively said, what's the difference between
• 7:20
  Speaker 4: sending the flower of youth to run at machine guns
• 7:25
  Speaker 4: to advance the interests of the state. What's the difference
• 7:28
  Speaker 4: between that and conscripting people to be in medical research
• 7:32
  Speaker 4: to generate things that are going to cure people?
• 7:35
  Speaker 5: You know, the knowledge that's going to cure people.
• 7:37
  Speaker 4: In both cases you have to sacrifice people for the
• 7:40
  Speaker 4: greater good, and so on the philosophical side, there's this
• 7:44
  Speaker 4: utilitarian conception of the greater good as this sum, this aggregate,
• 7:50
  Speaker 4: and that's very often a very dehumanizing concept because the
• 7:55
  Speaker 4: sum and the aggregate is this giant thing and individual
• 7:59
  Speaker 4: people are just one little tiny data point. They're like
• 8:03
  Speaker 4: one little digit in that, and so it's very easy
• 8:05
  Speaker 4: to run roughshot over their rights. And so part of
• 8:09
  Speaker 4: the book is to try to reorient things to see
• 8:13
  Speaker 4: the common good as actually the set of interests that
• 8:17
  Speaker 4: all humans, all people share in being able to develop
• 8:22
  Speaker 4: the capabilities that they need to pursue some life plan.
• 8:26
  Speaker 4: So we all differ in terms of the life plans
• 8:28
  Speaker 4: that we want to pursue. You know, some of us
• 8:30
  Speaker 4: want to be doctors, some of us want to be lawyers.
• 8:32
  Speaker 4: Some of us want to be dancers, artists, you know, monks, priests.
• 8:38
  Speaker 4: You know, like there's there's a huge diversity in the
• 8:42
  Speaker 4: life plan that people want to pursue. But we all
• 8:45
  Speaker 4: share this second order interest in being able to pursue
• 8:49
  Speaker 4: some life plan, and so we want to be able
• 8:52
  Speaker 4: to figure out what life plan that should be you
• 8:54
  Speaker 4: have the freedom to pursue it. And part of what
• 8:57
  Speaker 4: makes sickness and injury and disease so morally problematic is
• 9:01
  Speaker 4: the extent to which they hibit inhibit our ability to
• 9:05
  Speaker 4: do just that. And so once you sort of reorient
• 9:08
  Speaker 4: the common good now to sort of the set of
• 9:11
  Speaker 4: basic interests that everybody.
• 9:13
  Speaker 5: Has, then that helps.
• 9:15
  Speaker 4: To dilute this idea that linking research or the common
• 9:20
  Speaker 4: good is going to empower researchers to run roughshod over people.
• 9:26
  Speaker 4: Because the part of the goal of the book is
• 9:28
  Speaker 4: to show that there's no conflict between a productive research
• 9:33
  Speaker 4: enterprise that generates information that institutions need in order to
• 9:38
  Speaker 4: help people function, help people function those very basic ways.
• 9:42
  Speaker 4: There's no conflict between that and respect for the basic
• 9:46
  Speaker 4: interest and rights of the individual.
• 9:49
  Speaker 5: And so that's a conflict.
• 9:51
  Speaker 4: I think that research ethics skirted a little bit, and
• 9:54
  Speaker 4: it skirted it by basically saying we're not going to
• 9:57
  Speaker 4: talk about the common good. We're not going to talk
• 9:58
  Speaker 4: about these larger social purposes is because of fear that
• 10:01
  Speaker 4: they'll run rough shot over the individual. And I'm trying
• 10:04
  Speaker 4: to say, now that was a mistake. We can talk
• 10:07
  Speaker 4: about those larger social purposes. We can talk about the
• 10:09
  Speaker 4: common good in a way that supports, rather than undermines,
• 10:14
  Speaker 4: respect for individuals and their ability to function in really
• 10:18
  Speaker 4: important ways. And so then what that that also internalizes
• 10:22
  Speaker 4: the idea that a healthy and robust research enterprise should
• 10:26
  Speaker 4: be one that is respectful of people, so that free
• 10:31
  Speaker 4: and equal people in a democratic society can see research
• 10:35
  Speaker 4: as a way of contributing to something valuable and not
• 10:38
  Speaker 4: as something to be afraid of where they're by agreeing
• 10:41
  Speaker 4: to participate, they might be giving someone permission to run
• 10:46
  Speaker 4: rough shot over their rights or can script them into
• 10:50
  Speaker 4: as canon fodder for the greater good.
• 10:52
  Speaker 3: I think you were writing the preface of the book
• 10:55
  Speaker 3: around when COVID popped up COVID nineteen, right, because that's
• 10:58
  Speaker 3: sort of noticed. I think we never if we're not
• 11:01
  Speaker 3: thinking about free and equal persons. I mean thinking about
• 11:04
  Speaker 3: challenge trials. There are lots of folks sort of volunteering
• 11:07
  Speaker 3: in that way. Wouldn't be there great answer, and I
• 11:10
  Speaker 3: really appreciate. I think there's a host of applications, especially
• 11:13
  Speaker 3: for the students will be listening to this. I'm thinking
• 11:16
  Speaker 3: about sort of the many and individuals and respecting rights.
• 11:20
  Speaker 3: Since you mentioned Nuremberg and some of the medical research
• 11:24
  Speaker 3: atrocities in the US, I wanted to just ask about
• 11:27
  Speaker 3: the current socio political context across the globe. So, I mean,
• 11:31
  Speaker 3: given considerations of justice and equality, you have that great
• 11:34
  Speaker 3: part to the title of research among equals or something
• 11:37
  Speaker 3: like that right, which I think is thought provoking. I'm
• 11:39
  Speaker 3: curious what you think about research in the current climate.
• 11:43
  Speaker 3: Has it changed what you want to say about research ethics?
• 11:46
  Speaker 3: Has it changed me? The book still a few years old,
• 11:48
  Speaker 3: How do you think about research ethics now?
• 11:50
  Speaker 4: So, I actually think that there's a lot of destruction
• 11:55
  Speaker 4: that's happening in terms of a lack of support.
• 11:58
  Speaker 5: And a reigning in and moving away.
• 12:00
  Speaker 4: From some of the institutions that were part of the
• 12:03
  Speaker 4: success of research ethics. And there's a retrenchment happening from
• 12:07
  Speaker 4: a kind of globalism where the idea that we want
• 12:11
  Speaker 4: to share knowledge and information globally so that smart people
• 12:15
  Speaker 4: everywhere can contribute to the solution of problems that affect everyone.
• 12:21
  Speaker 4: And I don't think that those are beneficial trends, but
• 12:26
  Speaker 4: I think we can make benefit happen from them. In
• 12:30
  Speaker 4: the sense that the book is a criticism of research ethics,
• 12:37
  Speaker 4: but also a little bit of a roadmap for how
• 12:39
  Speaker 4: we could do better. And I think that one of
• 12:42
  Speaker 4: the beneficial byproducts of some of the changes that are
• 12:47
  Speaker 4: being implemented now is that the ground is being cleared,
• 12:53
  Speaker 4: maybe not the way that we would have wanted to
• 12:55
  Speaker 4: go about reform, but for us to then rethink, what
• 12:59
  Speaker 4: should the institutions of research ethics look like? What should
• 13:05
  Speaker 4: research oversight look like? I do think it's incredibly important.
• 13:10
  Speaker 4: I think there is multiple factors from national security and
• 13:16
  Speaker 4: international competition and other factors that are leading various people
• 13:22
  Speaker 4: to not want to share scientific information because it gives
• 13:25
  Speaker 4: you a competitive advantage basically, and I think that that's
• 13:30
  Speaker 4: shortsighted and so and I think the there's also it
• 13:35
  Speaker 4: would be a mistake to also not look back and
• 13:38
  Speaker 4: see the ways in which that global sharing wasn't always
• 13:42
  Speaker 4: equal sharing. But there was a lot of extraction right
• 13:46
  Speaker 4: that high income world research was outsourced to low and
• 13:52
  Speaker 4: middle income countries to generate knowledge that was primarily going
• 13:57
  Speaker 4: to be of benefit to populations in high end income countries.
• 14:02
  Speaker 4: So you were sort of exporting risk but importing, you know,
• 14:05
  Speaker 4: the benefit. I don't think that that's great either and
• 14:10
  Speaker 4: so the part of the not the idea of research
• 14:13
  Speaker 4: among equals, is a roadmap for saying, here's how you
• 14:16
  Speaker 4: can have international collaborations where your international partners value the
• 14:21
  Speaker 4: collaboration for the same reasons that the more powerful, wealthier,
• 14:27
  Speaker 4: more enfranchised countries value the collaboration. And this is basically
• 14:32
  Speaker 4: about trying to make research generate information that works to
• 14:35
  Speaker 4: the benefit of everybody.
• 14:37
  Speaker 5: And I think we will need new new institutions to
• 14:39
  Speaker 5: do that.
• 14:40
  Speaker 4: And COVID was a reminder that disease doesn't know borders, boundaries,
• 14:47
  Speaker 4: it doesn't know ethnicity, nationality, you know, and it's it's
• 14:53
  Speaker 4: not just a threat to members of one political party
• 14:57
  Speaker 4: or members of one you know, nation state. And one
• 15:00
  Speaker 4: of the radical things about ethics is, you know, those
• 15:03
  Speaker 4: basic interests that I have and that you have in
• 15:06
  Speaker 4: being able to form and pursue some life plan. There
• 15:10
  Speaker 4: isn't a person on the planet who doesn't have that interest.
• 15:12
  Speaker 4: I mean, we might want to pursue different life plans,
• 15:14
  Speaker 4: but we all have that higher interest in being able
• 15:19
  Speaker 4: to formulate, revise and pursue a life plan a value
• 15:24
  Speaker 4: to us. And that's true no matter where we are,
• 15:26
  Speaker 4: what our form of government is and what our political
• 15:28
  Speaker 4: allegiances are. And so in that sense, I think it's
• 15:32
  Speaker 4: good for medicine and for research and research ethics and
• 15:36
  Speaker 4: science in general to transcend some of those local identities
• 15:41
  Speaker 4: because all humans share this fundamental Our well being is
• 15:45
  Speaker 4: all tied up with being able to pursue a life plan,
• 15:47
  Speaker 4: and we share that at a deeper level than any
• 15:51
  Speaker 4: of our more parochial identities.
• 15:54
  Speaker 3: Yeah, I mean really helpful, and I think we'll resonate
• 15:57
  Speaker 3: with alsots folks who listen to the podcast, especially some
• 16:00
  Speaker 3: of the future physicians who are going, you know, into
• 16:04
  Speaker 3: medicine for a lot of reasons connected up nicely with
• 16:06
  Speaker 3: the common good in the background. So I want to
• 16:09
  Speaker 3: shift us a little bit from thinking about research ethics
• 16:11
  Speaker 3: to thinking about artificial intelligence in medicine. So you were
• 16:16
  Speaker 3: out front, at least as far as I saw, over
• 16:18
  Speaker 3: the past few years on a lot of conversations around
• 16:21
  Speaker 3: ethics and AI, and as it's rollout, really ramped up
• 16:24
  Speaker 3: in the last few years, and you've got the big
• 16:26
  Speaker 3: name chair in ethics and computation Technologies at CMU. So
• 16:31
  Speaker 3: now into the AI phase of the conversation. What for you,
• 16:33
  Speaker 3: what are the most important connections between research ethics and
• 16:37
  Speaker 3: artificial intelligence that you're well situated to think about.
• 16:39
  Speaker 5: But do you think.
• 16:41
  Speaker 3: Medical AI should be assessed using research ethics frameworks? Some
• 16:45
  Speaker 3: folks have argued that should be thought about differently. What's
• 16:48
  Speaker 3: your take?
• 16:48
  Speaker 4: Yeah, I think let me step back for a second,
• 16:52
  Speaker 4: like a little bit more of a kind of general
• 16:55
  Speaker 4: level and just say what it was that why does
• 16:59
  Speaker 4: somebody like me who writes about research ethics also read
• 17:02
  Speaker 4: about AI? Like what's the connection? And the connection is
• 17:07
  Speaker 4: that in both cases there's uncertainty. There's uncertainty about what
• 17:12
  Speaker 4: something does. So if you think about a drug, you know,
• 17:16
  Speaker 4: the goal is how can we take this substance and
• 17:20
  Speaker 4: figure out how to use it in order to create
• 17:23
  Speaker 4: some beneficial effect inside of a person, And so that
• 17:27
  Speaker 4: we have to figure out all kinds of stuff, like
• 17:30
  Speaker 4: what's the drug's action, how does the body respond to it?
• 17:34
  Speaker 4: What's a dose that is active not toxic? Right, what's
• 17:38
  Speaker 4: a schedule for giving that drug that it maintains the
• 17:42
  Speaker 4: right kind of concentrations? What are the contraindications Like people
• 17:46
  Speaker 4: will have some you know, comorbidity or other thing that
• 17:50
  Speaker 4: we need to treat, or there's a side effect that
• 17:52
  Speaker 4: we need to manage. There's all kinds of things that
• 17:54
  Speaker 4: we need to know in order to produce a therapy,
• 17:56
  Speaker 4: and it's incredibly difficult and we're all often wrong. So
• 18:01
  Speaker 4: drug development is super interesting to a philosopher because the
• 18:06
  Speaker 4: very often, most of the time, we're wrong. Nine out
• 18:09
  Speaker 4: of every ten things that we try that goes into
• 18:11
  Speaker 4: development doesn't work, and nobody spends the colossal amounts of
• 18:17
  Speaker 4: money to bring a drug to market because they don't
• 18:19
  Speaker 4: believe in it. Right, everybody believes in their pipeline and
• 18:23
  Speaker 4: almost nothing comes out of the pipeline. So that's just
• 18:26
  Speaker 4: a sign. That's not to say anything take anything away
• 18:28
  Speaker 4: from everybody. You have smart people, it's just incredibly hard.
• 18:31
  Speaker 4: The science is immature. Okay, Why am I saying that
• 18:35
  Speaker 4: when we're in the AI part of the conversation, Because
• 18:37
  Speaker 4: to me, AI is very similar to that. We have
• 18:40
  Speaker 4: these AI systems that are they're not deterministic in the
• 18:42
  Speaker 4: way that a traditional computer program is deterministic where you
• 18:46
  Speaker 4: just have a set of rules. If you have this input,
• 18:48
  Speaker 4: then deterministically you have this output. You have these very
• 18:52
  Speaker 4: complex stochastic systems, and they're impossible for us to get
• 19:00
  Speaker 4: so we know how they work because we designed them,
• 19:03
  Speaker 4: like you know, you know, very smart people in computer
• 19:06
  Speaker 4: science understand the architecture that make these things work.
• 19:10
  Speaker 5: That's not what I'm saying.
• 19:11
  Speaker 4: What I'm saying is if a model has a trillion parameters,
• 19:16
  Speaker 4: it's your mind can't wrap around what a trillion parameters is.
• 19:20
  Speaker 4: If the model has been trained on large portions of
• 19:23
  Speaker 4: the Internet, you can't wrap your mind around all of
• 19:26
  Speaker 4: the things that it is seen. And in that sense,
• 19:29
  Speaker 4: then it becomes very difficult to know is this not?
• 19:32
  Speaker 4: Is like, so this thing is producing texts that I
• 19:35
  Speaker 4: can understand. I can ask it a question. It gives
• 19:38
  Speaker 4: me a well formed English sentence in response, Is it understanding?
• 19:43
  Speaker 5: Does it know anything? You know?
• 19:46
  Speaker 4: It gives me the right answer to something, so does
• 19:49
  Speaker 4: it understand? And these are all the same sorts of questions.
• 19:54
  Speaker 4: So basically, I think we have to look at AI
• 19:57
  Speaker 4: systems like a drug at is there a task that
• 20:01
  Speaker 4: we can use these AI systems for that will improve
• 20:05
  Speaker 4: our ability to prevent disease, treat disease, help people understand
• 20:10
  Speaker 4: their health conditions, avoid what's bad, promote what's good. That's
• 20:14
  Speaker 4: what we want to use AI for. It's uncertain whether
• 20:18
  Speaker 4: we can and we have to develop the knowledge. That's
• 20:23
  Speaker 4: very similar to developing a drug. Under what conditions is
• 20:26
  Speaker 4: this AI going to work? You know, what are things
• 20:29
  Speaker 4: where if we prompt it, it's going to go, it's
• 20:32
  Speaker 4: gonna hallucinate, It's going to do something that's bad. Right,
• 20:35
  Speaker 4: what's our tolerable risk? Like how many hallucinations are we
• 20:38
  Speaker 4: willing to tolerate? Are there tasks that we shouldn't use
• 20:41
  Speaker 4: it for because it's going to hallucinate?
• 20:43
  Speaker 5: So for me, I came.
• 20:45
  Speaker 4: To this as seeing, yeah, many of that same uncertainty
• 20:49
  Speaker 4: appears in both places. And then the thing that bothered
• 20:53
  Speaker 4: me was the degree to which people sort of felt
• 20:56
  Speaker 4: like they could just look into the matrix and understand
• 20:59
  Speaker 4: how these things work and predict how they were going
• 21:01
  Speaker 4: to perform. And they were pretty credulous about taking what
• 21:05
  Speaker 4: they said, you know, the output of the model and
• 21:09
  Speaker 4: assuming that it had the same underlying cognitive grasp of
• 21:13
  Speaker 4: the problem that you would have if you said the
• 21:15
  Speaker 4: same thing to me. And and you know, there's lots
• 21:19
  Speaker 4: of evidence that that's just not the case. These things
• 21:22
  Speaker 4: produce text in they will produce the same output that
• 21:26
  Speaker 4: you might say to me, but your thinking and your
• 21:29
  Speaker 4: cognitive grasp of the underlying structure of the problem is
• 21:33
  Speaker 4: not something that is present in the model that it
• 21:35
  Speaker 4: produces its outputs in a very different way because of
• 21:40
  Speaker 4: statistical correlations between letter pairs or letter triples. And it's
• 21:44
  Speaker 4: just incredibly difficult to understand, Well, how can that be? Like,
• 21:48
  Speaker 4: how can something that just has statistical correlations between these
• 21:52
  Speaker 4: tokens produce these patterns that are so coherent.
• 21:56
  Speaker 5: But improving the methodology by.
• 21:58
  Speaker 4: Which we test and evaluate AI systems, to me, is
• 22:02
  Speaker 4: the analog to making sure that we are our tests
• 22:06
  Speaker 4: of drugs controls for confounding right. So you know, the
• 22:12
  Speaker 4: example that I like to use with my students is like,
• 22:15
  Speaker 4: tell me what the curve of a headache looks like?
• 22:18
  Speaker 4: If you get a headache, is it just a straight
• 22:20
  Speaker 4: line through space where you know, it's like just increasing
• 22:23
  Speaker 4: in intensity forever such that if then I give you
• 22:26
  Speaker 4: a drug and the headache goes away, then I'd know
• 22:28
  Speaker 4: that it was the drug that got rid of the headache. No,
• 22:30
  Speaker 4: the world doesn't like that. Like, if you get a
• 22:32
  Speaker 4: headache and we don't do anything for you, very often
• 22:36
  Speaker 4: the headache just goes away. Like headaches don't tend to
• 22:39
  Speaker 4: last forever, and so that you know, a headache is
• 22:43
  Speaker 4: this you know, rising curve that then goes down back
• 22:49
  Speaker 4: to baseline. And when do people decide to take something
• 22:53
  Speaker 4: for their headache, you know, usually it's like, oh, I'm
• 22:55
  Speaker 4: getting a headache, you know, and now it's really bad. Okay,
• 22:59
  Speaker 4: now I should do something about it. And the counterfactual
• 23:01
  Speaker 4: is like, what if you didn't do anything about it,
• 23:03
  Speaker 4: Like would it have just kept going forever or would
• 23:06
  Speaker 4: it have eventually subsided and gone away?
• 23:08
  Speaker 5: And if it would have.
• 23:09
  Speaker 4: Subsided and gone away, then I could give you anything.
• 23:12
  Speaker 4: I give you a sugar pill, I give you pez.
• 23:14
  Speaker 4: You know, I can be like, oh, you stepped outside.
• 23:17
  Speaker 4: That cured your Like all those things happened at the
• 23:19
  Speaker 4: same time that cured your headache, And so I think
• 23:21
  Speaker 4: you get some of the same thing on the AI side,
• 23:23
  Speaker 4: Like if the AI memorizes the benchmark that you're using
• 23:26
  Speaker 4: to evaluate it, you can't. You're a smart person, but
• 23:31
  Speaker 4: you can't memorize every board test that's been given in
• 23:35
  Speaker 4: the history of board tests, right, But a model with
• 23:37
  Speaker 4: a trillion parameters can memorize every board test that's been given.
• 23:41
  Speaker 4: If it's been shown every board test, and so that
• 23:44
  Speaker 4: may mean then they can guess the right answer by
• 23:47
  Speaker 4: memorizing or by seeing correlations between which distractors. You know,
• 23:53
  Speaker 4: so in multiple choice, you know, we often choose if
• 23:57
  Speaker 4: this is the right answer, then we choose a particular
• 23:59
  Speaker 4: time distractor.
• 24:01
  Speaker 5: Why, Well, because it's close enough.
• 24:04
  Speaker 4: Then we think that that's the sort of thing that
• 24:06
  Speaker 4: you would it might come to your mind because of
• 24:09
  Speaker 4: something else we said and the prompt. So there are
• 24:11
  Speaker 4: these underlying patterns. The people who make these tests, they
• 24:14
  Speaker 4: choose distractors, and the model can know that it can
• 24:19
  Speaker 4: ascertain like if this is the right Like, given the
• 24:21
  Speaker 4: pattern that's here, this is probably the right answer, not
• 24:25
  Speaker 4: because it knows anything about what you're asking, but just
• 24:27
  Speaker 4: because given the distractors that are there, it's probably this
• 24:31
  Speaker 4: is the right answer. And we're not used to that.
• 24:34
  Speaker 4: People don't work that way. So the uncertainty this is
• 24:38
  Speaker 4: all just by way of saying what makes me attracted
• 24:40
  Speaker 4: to it? And the reason why I think they're sort
• 24:42
  Speaker 4: of similar is that we don't know how the thing works,
• 24:45
  Speaker 4: and it's a struggle to figure out how to use
• 24:47
  Speaker 4: it for clinical benefit.
• 24:48
  Speaker 5: That's true for drugs, and it's true for AI that was.
• 24:52
  Speaker 3: Great and really helpful, and I had not haven't heard
• 24:55
  Speaker 3: of the headache analogy in this way. That's I think
• 24:57
  Speaker 3: it's illustrative. So when students and I talk about this,
• 25:02
  Speaker 3: when we do this in class, there's two very common questions.
• 25:05
  Speaker 3: I think the first one clearly connects up nicely with
• 25:08
  Speaker 3: both research ethics and questions about medical AI, and it's
• 25:12
  Speaker 3: around informed consent. So students will say thing, well, ask
• 25:16
  Speaker 3: about informed consent or some sort of less technical legalistic
• 25:20
  Speaker 3: idea around shared decision making, and they're concerned about what
• 25:24
  Speaker 3: they don't know and if they rely on something else
• 25:27
  Speaker 3: to guide them when in the future they're going to
• 25:29
  Speaker 3: be advising and helping patients. So any thought I mean,
• 25:34
  Speaker 3: I say something what I think about later, but any
• 25:36
  Speaker 3: thoughts on informed consent, shared decision making. I mean, this
• 25:39
  Speaker 3: is all patient centered care. These students are training on
• 25:42
  Speaker 3: the social determinants of health. They really really want to
• 25:44
  Speaker 3: help folks, and they want it to be the patient's decision.
• 25:51
  Speaker 3: But they're going to rely on something that they don't
• 25:53
  Speaker 3: fully understand, So any guidance thoughts what should they do?
• 26:00
  Speaker 4: I think that this is probably one of the most
• 26:02
  Speaker 4: confusing issues for so many people. And part of the
• 26:07
  Speaker 4: confusion is just trying to figure out where you are
• 26:12
  Speaker 4: in a map. It's like being disoriented, like before the
• 26:15
  Speaker 4: days of Google. So your students may be young enough
• 26:20
  Speaker 4: that they've never actually had to struggle with a paper map,
• 26:24
  Speaker 4: you know, like because Google maps are Apple maps, so
• 26:27
  Speaker 4: they put you, they tell you where you are, like
• 26:30
  Speaker 4: that's part of their value. They're like, you're here in
• 26:33
  Speaker 4: this space, and this is north. You're facing north or
• 26:36
  Speaker 4: south or so you can orient yourself. But in the
• 26:39
  Speaker 4: old days with a paper map, you know, you'd be like, where.
• 26:42
  Speaker 5: Am I in this space?
• 26:44
  Speaker 3: Is this?
• 26:46
  Speaker 4: You know? Is this feature the feat on the map?
• 26:50
  Speaker 4: The feature I'm looking at? And that's the way I
• 26:54
  Speaker 4: think about this, because what it's difficult for doctors in particular,
• 27:00
  Speaker 4: So uncertainty is the foe of doctors. Doctors are find
• 27:05
  Speaker 4: it very difficult to deal with uncertainty, partly because patients
• 27:10
  Speaker 4: find it very difficult to deal with uncertainty. You know,
• 27:14
  Speaker 4: when you're sick, you want answers, and you go to
• 27:16
  Speaker 4: the doctor for answers, and a lot of doctors want
• 27:19
  Speaker 4: to be able to provide answers and patients get frustrated
• 27:22
  Speaker 4: and doctors get frustrated if you can't that makes it,
• 27:26
  Speaker 4: you know, that makes people want to say, well, I
• 27:28
  Speaker 4: need certainty from the AI, or or at least I
• 27:31
  Speaker 4: need to understand how the AI works. And I've written
• 27:34
  Speaker 4: about how that isn't quite right because for a long time,
• 27:38
  Speaker 4: if the patient came to you with a headache, you
• 27:41
  Speaker 4: would say, well, take an aspirin. And if the patient
• 27:44
  Speaker 4: said why, you would say something like because it's going
• 27:48
  Speaker 4: to make your headache go away.
• 27:49
  Speaker 5: And if they.
• 27:50
  Speaker 4: Said how would you know that, you would say because
• 27:53
  Speaker 4: when we give it to people, their headaches tend to
• 27:55
  Speaker 4: go away. And if they said, can you explain how
• 27:58
  Speaker 4: that works, then you would have to say no, because
• 28:02
  Speaker 4: for the hundred years of using aspirin, we.
• 28:06
  Speaker 5: Don't know the mechanism by which it works.
• 28:09
  Speaker 4: And medicine is unlike lots of other areas because we
• 28:14
  Speaker 4: know first that something works and we figure out the
• 28:18
  Speaker 4: mechanism later, and when we try to use mechanism to
• 28:22
  Speaker 4: make things work, we're often wrong. So radical mastectomy was
• 28:26
  Speaker 4: an example where mechanism drove the idea that was like, well,
• 28:30
  Speaker 4: we can stop tumor recurrence if we remove more tissue
• 28:34
  Speaker 4: and then we ran you know, they had to run
• 28:37
  Speaker 4: a large randomized controlled trial between mastactiny and lumpectomy and
• 28:43
  Speaker 4: then measure survival benefit in order to see, oh, the
• 28:47
  Speaker 4: underlying theory wasn't true. So it's common in medicine that
• 28:52
  Speaker 4: we and part of what a randomized controlled trial?
• 28:55
  Speaker 5: What many research methods are there for.
• 28:57
  Speaker 4: They don't tell you how something works, They just tell
• 29:00
  Speaker 4: you whether it works or not. And so on my view,
• 29:06
  Speaker 4: it's it's not a problem that AI is a black box.
• 29:11
  Speaker 4: That isn't the problem because aspirin's a black box. Lithium
• 29:16
  Speaker 4: as a mood stabilizer. We know it stabilizes mood. We
• 29:19
  Speaker 4: don't know the mechanism. That's also a black box. But
• 29:23
  Speaker 4: you have to know that it works, and you have
• 29:25
  Speaker 4: to know what dose to give it and who not
• 29:28
  Speaker 4: to give it to and when you know So there's
• 29:31
  Speaker 4: all those things going back what we call the intervention ensemble.
• 29:34
  Speaker 4: It's ensemble of things you have to know about something
• 29:37
  Speaker 4: in order to unlock its clinical benefit. Right, you know, what,
• 29:41
  Speaker 4: what what's the what's the indication? Who?
• 29:43
  Speaker 5: Who can I give this to it? What does at
• 29:45
  Speaker 5: what schedule?
• 29:46
  Speaker 2: Right?
• 29:46
  Speaker 5: So I need to know that for the AI. So
• 29:50
  Speaker 5: there's lots of AI.
• 29:52
  Speaker 4: There's lots of examples where it's like it will work
• 29:54
  Speaker 4: very well if you're a white male. How well does
• 29:57
  Speaker 4: it work on the rest of the populations that I see.
• 30:01
  Speaker 5: I need to know that.
• 30:02
  Speaker 4: I don't know that by understanding how the model works,
• 30:05
  Speaker 4: how the model takes inputs and produces outputs. I know
• 30:08
  Speaker 4: that by evaluating how well it works in these different populations.
• 30:14
  Speaker 4: And to me, I'm okay using a black box if
• 30:19
  Speaker 4: I've evaluated how it works under realistic, real world conditions
• 30:24
  Speaker 4: that control for confounding, because that's if you're not comfortable
• 30:28
  Speaker 4: with that, Medicine's not the right discipline for you, because
• 30:32
  Speaker 4: how else are you going to prescribe many of the
• 30:36
  Speaker 4: things where we know the effect, you know the efficacy,
• 30:40
  Speaker 4: but not the mechanism.
• 30:41
  Speaker 5: So what does that mean now?
• 30:42
  Speaker 4: In terms of that, what that doesn't mean is like
• 30:45
  Speaker 4: you could just use AI for anything, because that's the
• 30:48
  Speaker 4: point you have a whole bunch of people. The oldest profession,
• 30:52
  Speaker 4: in my view, is the oldest profession is the profession
• 30:56
  Speaker 4: that says I can sell you this to make whatever
• 30:59
  Speaker 4: it is that's ailing you go away, right and that's
• 31:03
  Speaker 4: just like forever there have been people that are like, here,
• 31:05
  Speaker 4: suck on this pebble. It's going to make you know
• 31:07
  Speaker 4: your feet are herting here, this will get rid of it.
• 31:09
  Speaker 5: Or let's do this right, And that's super cheap and easy.
• 31:12
  Speaker 4: Right, and because most things go away, then it also
• 31:14
  Speaker 4: you look great, right, you're like here, you know, let
• 31:16
  Speaker 4: me rub your temple. Wow, your headache went away? See like,
• 31:19
  Speaker 4: look how great I am. Well, that's ai. There's just
• 31:21
  Speaker 4: there's lots of purveyors out there who want to sell
• 31:24
  Speaker 4: you something with the idea that it's magic, it'll do
• 31:27
  Speaker 4: whatever you want.
• 31:28
  Speaker 5: And all I want to see is a little.
• 31:30
  Speaker 4: Bit more incredulity and a little bit more sense of
• 31:32
  Speaker 4: what are what's the evidence that it does something of value?
• 31:36
  Speaker 4: And what's the how rigorously has that evidence been generated?
• 31:42
  Speaker 4: So have we controlled for other explanations about how that
• 31:45
  Speaker 4: thing works?
• 31:46
  Speaker 3: Super helpful. I think something that additional thing through will
• 31:51
  Speaker 3: will repay. So I want to get you out on
• 31:54
  Speaker 3: a policy question. Given the speed with which emerging technologies
• 31:57
  Speaker 3: are rot do you have any advice for folks who
• 32:00
  Speaker 3: are interested in policies? And we have a bunch of
• 32:01
  Speaker 3: students are super interested in health policies and sort of
• 32:03
  Speaker 3: implications of practice, Any advice on creating lasting but malleable
• 32:08
  Speaker 3: policies because there's some nimbleness that needs to be built in.
• 32:12
  Speaker 3: But if we want clear guidelines to make sure that
• 32:15
  Speaker 3: folks who know less right are able to still practice well,
• 32:19
  Speaker 3: so you know, from a thirty thousand foot viewer, or
• 32:23
  Speaker 3: really any advice for folks interested in policy and how
• 32:26
  Speaker 3: to create policies around AI or even around research which
• 32:29
  Speaker 3: are both malleable but also clear. Yeah.
• 32:33
  Speaker 4: I so this is one of the things that I've
• 32:38
  Speaker 4: been in various meetings where developers or people vendors very
• 32:43
  Speaker 4: often will say things like, we can't wait for regulation
• 32:46
  Speaker 4: because things are moving too quickly. Patients are desperate, you know,
• 32:50
  Speaker 4: patients want you know, want effective treatments or cures, and
• 32:53
  Speaker 4: you know, so we shouldn't stand in the way of
• 32:56
  Speaker 4: using things that work. And I think it would be
• 32:59
  Speaker 4: horrendous wrong to think that what ethics is about is
• 33:03
  Speaker 4: standing in the way of people having access to things
• 33:06
  Speaker 4: that work. So, which is something that's often said about
• 33:10
  Speaker 4: regulatory you know, like the FDA. That's why let me
• 33:13
  Speaker 4: just reiterate that nine out of every ten things that
• 33:16
  Speaker 4: we try doesn't work for anything in drug development, and
• 33:19
  Speaker 4: so very often what we're not standing in the way
• 33:22
  Speaker 4: of is people getting access to things that work. Part
• 33:25
  Speaker 4: of what clinical trials show is that they shouldn't get
• 33:29
  Speaker 4: access to this thing that you were developing because it
• 33:31
  Speaker 4: doesn't work. And you know, giving people access to things
• 33:35
  Speaker 4: that don't work is bad for everybody. We spend scarce
• 33:38
  Speaker 4: resources on things that don't work. Your time is dedicated
• 33:41
  Speaker 4: to things that don't work. Patients opportunity costs that either
• 33:44
  Speaker 4: it's either ineffective or harmful, and they have the opportunity
• 33:47
  Speaker 4: costs that they didn't get.
• 33:48
  Speaker 5: Something that would work for them, and we don't.
• 33:51
  Speaker 4: You know, we already spend the amount of money that
• 33:53
  Speaker 4: the United States spends on healthcare. If you made that
• 33:56
  Speaker 4: the GDP of a country, it would be the largest
• 33:59
  Speaker 4: GDP outside of China. That's just what we spend on healthcare.
• 34:03
  Speaker 4: So it's not like we've got more money to spend
• 34:05
  Speaker 4: on healthcare. So in that sense, what I would say,
• 34:09
  Speaker 4: I'd like to turn the question back around and say
• 34:12
  Speaker 4: when I go into the hospital, I go into hospitals.
• 34:14
  Speaker 4: When I look at medicine, I still see fax machines,
• 34:19
  Speaker 4: Like my students don't know what a fact is, and
• 34:24
  Speaker 4: that's just because you know. So it's been a little
• 34:28
  Speaker 4: bit ironic to see medicine, which is so traditional, be
• 34:33
  Speaker 4: at the cutting edge of you know, of AI and
• 34:37
  Speaker 4: AI deployments. And I think it's not just hard to
• 34:41
  Speaker 4: make policy, it would be hard to implement something where
• 34:45
  Speaker 4: when you know the next update, the model that you
• 34:48
  Speaker 4: were using is going to be obsolete. Like if the
• 34:50
  Speaker 4: pace of turnover and churn in the technology right now
• 34:54
  Speaker 4: is so fast and it takes so long just to
• 34:58
  Speaker 4: implement a technology and act will practice, then how do
• 35:02
  Speaker 4: you make sure that you're implementing something that actually works
• 35:05
  Speaker 4: under conditions that actually work. That's not about slowing anything down,
• 35:10
  Speaker 4: that's just about how do you solve the problem of
• 35:12
  Speaker 4: rapid obsolescence, not even from the regulatory standpoint, but from
• 35:17
  Speaker 4: like the effective and safe implementation standpoint. So I do
• 35:21
  Speaker 4: think policy is a you know, like we need to
• 35:24
  Speaker 4: develop nimble policies, but I also think are the cornerstone
• 35:31
  Speaker 4: of policies has to be a credible evidence that AI
• 35:36
  Speaker 4: systems produce a genuine clinical benefit or some genuine operational benefit,
• 35:41
  Speaker 4: because where a lot of AI is actually being used
• 35:45
  Speaker 4: probably to the greatest effect in medicine. Is is medicine
• 35:49
  Speaker 4: on the business side, right on the organizational running of
• 35:53
  Speaker 4: you know, the of the industry as opposed to patient
• 35:56
  Speaker 4: facing side. But I you know, I don't think it's
• 35:59
  Speaker 4: on reasonable to have a policy that says, show me
• 36:02
  Speaker 4: credible evidence that this thing produces a net benefit. And
• 36:08
  Speaker 4: I want that evidence to be such that I'm willing
• 36:11
  Speaker 4: to believe it. And if you just show me that
• 36:13
  Speaker 4: evidence because you've trained, you know, by by showing the
• 36:17
  Speaker 4: performance on an established benchmark, or showing me what it
• 36:22
  Speaker 4: does on the holdout set of you know, when we
• 36:26
  Speaker 4: train and validate machine learning models, you have often what
• 36:31
  Speaker 4: you do is you take one large data set, you
• 36:34
  Speaker 4: develop the model on some portion of that, say half
• 36:36
  Speaker 4: of it, and then you have a holdout set that
• 36:39
  Speaker 4: you then validate the model on. That's an important part
• 36:43
  Speaker 4: of model validation, but that is not sufficient to then
• 36:46
  Speaker 4: justify deployment into the clinic. So so here's the thing
• 36:50
  Speaker 4: that's missing in drug development. There's it's problematic, but at
• 36:56
  Speaker 4: least you know, there will say here's a phase one study. Here,
• 37:00
  Speaker 4: like here's an animal model study, here's a proof of concept. Right,
• 37:03
  Speaker 4: there's a proof of concept. Here's a phase one, here's
• 37:05
  Speaker 4: a phase two, here's a large confirmatory phase three that
• 37:09
  Speaker 4: controls for various types of confounding. So you can see
• 37:11
  Speaker 4: what's how mature is the knowledge that I'm you know,
• 37:14
  Speaker 4: how mature is this technology? We need something like that
• 37:17
  Speaker 4: in AI to be able to distinguish when is something
• 37:20
  Speaker 4: proof of concept? When is it been validated in a
• 37:23
  Speaker 4: way like in silico, but that's unlikely to translate into
• 37:26
  Speaker 4: clinical practice. When has there been a silent trial? When
• 37:30
  Speaker 4: has there been an actual, you know, uh, controlled trial?
• 37:36
  Speaker 4: Because the evidence we get from these different types of
• 37:38
  Speaker 4: validation is different. And like I said before, it's just
• 37:42
  Speaker 4: very common in medicine that people say, we need to
• 37:46
  Speaker 4: give this to people it works, and then closer scrutiny
• 37:49
  Speaker 4: reveals that it doesn't. And you know, this happened during
• 37:53
  Speaker 4: the pandemic. And one of the papers that I'm very
• 37:56
  Speaker 4: proud of that we published in Science during the pandemic
• 37:59
  Speaker 4: was called a against Pandemic Research exceptionalism because you had
• 38:03
  Speaker 4: a lot of people that were like, there's a pandemic,
• 38:05
  Speaker 4: we can't afford the time that it takes to run
• 38:09
  Speaker 4: well controlled studies, and it's like, look at what happened
• 38:13
  Speaker 4: to us, you know, hydroxy chloroquine went, you know, like
• 38:18
  Speaker 4: everybody wanted this right, it got out because of unpoorly designed,
• 38:23
  Speaker 4: uncontrolled studies in the beginning that are basically the pandemic
• 38:28
  Speaker 4: analog to the headache example. The majority of the people
• 38:33
  Speaker 4: who got COVID didn't die. The majority of the people
• 38:36
  Speaker 4: that you gave hydroxychlorocine too, who got COVID didn't die,
• 38:40
  Speaker 4: and that doesn't mean that it was the hydroxychloroquine that
• 38:43
  Speaker 4: saved them. And so lots of time effort, like the
• 38:47
  Speaker 4: you know, the pandemic was really a horror show of
• 38:52
  Speaker 4: redundant efforts, poor quality studies, studies that largely from inception
• 38:59
  Speaker 4: you knew, weren't to answer the questions that needed to
• 39:03
  Speaker 4: be answered, and we want to avoid that that wastes
• 39:07
  Speaker 4: everybody's time, effort, energy around our common goal, which is
• 39:11
  Speaker 4: to make healthcare better and be able to better meet
• 39:14
  Speaker 4: the needs of patients, and so requiring clear and convincing
• 39:20
  Speaker 4: evidence of safety and efficacy for AI to me is
• 39:24
  Speaker 4: like sort of should be the gateway to entry, so
• 39:28
  Speaker 4: that we spend our scarce resources on it we dedicate
• 39:31
  Speaker 4: our time to the right things. To me, the I
• 39:36
  Speaker 4: can boil down sort of the quintessence of research ethics.
• 39:41
  Speaker 4: You know, at least a core part of it is
• 39:45
  Speaker 4: finding the methods and the organizations that will allow us
• 39:50
  Speaker 4: to figure out what things work, what things don't. Shift
• 39:54
  Speaker 4: our practice to the things that do work, Shift our
• 39:57
  Speaker 4: practice away from the things that don't work, so that
• 39:59
  Speaker 4: we can move the ratchet in the right direction, so
• 40:02
  Speaker 4: that we're improving our ability to prevent disease, treat you know,
• 40:06
  Speaker 4: treat disease, improve the lives of patients, improve public health
• 40:11
  Speaker 4: with the and the good shepherds of the scarce resources
• 40:16
  Speaker 4: that society uses, you know, invest in medicine to do
• 40:19
  Speaker 4: those things.
• 40:20
  Speaker 3: So well said, and I think a powerful case for
• 40:23
  Speaker 3: good research, good oversight, both in research and in thinking
• 40:27
  Speaker 3: about AI and medical II. As alex A'll let you,
• 40:30
  Speaker 3: I said the last question before, here's the very last question.
• 40:32
  Speaker 3: For folks who enjoyed the conversation or interested, they're going
• 40:36
  Speaker 3: to go out and get the book. We'll put that
• 40:37
  Speaker 3: in the show notes. But any upcoming book, paper project,
• 40:41
  Speaker 3: big talk you're going to give that people could tune into.
• 40:44
  Speaker 3: Where can we point them if they're interested in more stuff.
• 40:46
  Speaker 4: I'm working on another book, but uh, it's going to
• 40:49
  Speaker 4: take a while. I'm writing it to be something like
• 40:52
  Speaker 4: an introduction to ethics textbook, mostly out of the frustration
• 40:57
  Speaker 4: that I experience when I read the way that people
• 41:00
  Speaker 4: in AI in particular tend to talk about ethics. So
• 41:04
  Speaker 4: some some teasers, lots of people will say, well, there's
• 41:07
  Speaker 4: there's deontology, there's there's consequentialism and virtue ethics, and those
• 41:13
  Speaker 4: are sort of the three traditions. And I think that
• 41:15
  Speaker 4: that's fundamentally wrong and unhelpful, because there's as much diversity
• 41:20
  Speaker 4: inside each of those boxes as there is across those boxes.
• 41:26
  Speaker 4: I also you have to understand, like I, I write
• 41:30
  Speaker 4: papers on Aristotle and Aristotle's ethics, and I, you know,
• 41:34
  Speaker 4: this will sound like heresy to a lot of people,
• 41:36
  Speaker 4: but I don't think he's actually a virtue theorist. Martha
• 41:40
  Speaker 4: Nosebaum has a really nice paper about that. So I'm
• 41:43
  Speaker 4: not like the only one who who thinks that. But
• 41:46
  Speaker 4: so the book is a way, you know, the next
• 41:49
  Speaker 4: book is just a way to try to say, look,
• 41:52
  Speaker 4: if if that isn't the way to think about these things,
• 41:55
  Speaker 4: then what is the right way to think about these things?
• 41:58
  Speaker 4: And so it takes some work I've already done in
• 42:01
  Speaker 4: some of my papers about the independence of practical ethics
• 42:06
  Speaker 4: and actually about Aristotle's rhetoric and some other things. And
• 42:10
  Speaker 4: so that's that's what's occupying me right now. But I'm
• 42:13
  Speaker 4: but I'm always, you know, thinking about the challenges that
• 42:18
  Speaker 4: AI poses to us. And you know, I have a
• 42:22
  Speaker 4: couple of papers that I'm working on that would be
• 42:25
  Speaker 4: relevant to the audience. I don't want to scoop them,
• 42:28
  Speaker 4: but you know, one is that I want as.
• 42:32
  Speaker 3: To scoop them. I can always pester you after and
• 42:34
  Speaker 3: we can stick them in the show page once they're published,
• 42:36
  Speaker 3: But give us one one idea and then I'll let
• 42:38
  Speaker 3: you out here.
• 42:39
  Speaker 4: I mean, part of the teaser is, I actually do
• 42:42
  Speaker 4: think that AI is going to change medicine pretty fundamentally.
• 42:46
  Speaker 5: But I think I.
• 42:47
  Speaker 4: Disagree with most people about how that's going to happen.
• 42:51
  Speaker 4: I don't think that it's going to happen the way
• 42:53
  Speaker 4: that say Jeff Hinton did when he said, Hey, i's
• 42:56
  Speaker 4: gonna we're not going to need radiologists, you know, like
• 43:00
  Speaker 4: seven or eight years ago now. He said in five years,
• 43:03
  Speaker 4: we won't need radiologists. It's absolutely clear. You know, AI
• 43:07
  Speaker 4: is going to replace all those people, which which is
• 43:09
  Speaker 4: there's definitely hasn't happened. But I don't think that in
• 43:13
  Speaker 4: order to succeed, something has to work better than the alternative.
• 43:16
  Speaker 4: There's going to be other reasons why people want to
• 43:19
  Speaker 4: implement AI. And it's not even going to be doctors.
• 43:21
  Speaker 4: It's going to be patients and other parties and actually
• 43:25
  Speaker 4: and doctors them themselves, not that they want to be replaced,
• 43:28
  Speaker 4: but the grind that they experience that already leads them
• 43:31
  Speaker 4: to want to use ambient systems, ambient listening so they
• 43:36
  Speaker 4: can have FaceTime and reduce pajama time. I think some
• 43:39
  Speaker 4: of these things are going to conspire to change medicine
• 43:41
  Speaker 4: in a way that people haven't thought about.
• 43:43
  Speaker 3: So that's right on and when that comes out, we'll
• 43:47
  Speaker 3: have you back on. So Alex, thanks so much. I
• 43:49
  Speaker 3: really appreciate the time.
• 43:51
  Speaker 5: Yeah, thanks a lot, it's been great.
• 43:52
  Speaker 1: Series is a production of the Health Ethics Lab, and
• 43:55
  Speaker 1: I am grateful to colleagues in the Department of Medical Sciences,
• 43:58
  Speaker 1: the IHIS Library, Hackensack Ready in the School of Medicine
• 44:02
  Speaker 1: for making this collaboration possible. To learn more about our
• 44:05
  Speaker 1: work or to engage with the series content, reach out
• 44:08
  Speaker 1: to me Brian Pilkington at Brian dot Pilkington at hm
• 44:12
  Speaker 1: h N dot org.