Artificial intelligence is often described as a faster way to process information. Deborah Estrin is asking a deeper question. What happens when it becomes part of the interaction itself?

Deborah is a computer scientist whose work sits at the intersection of technology, health, and everyday life. Across her career, she has studied how mobile devices, sensors, and personal data can help patients, clinicians, and caregivers see patterns that would otherwise remain hidden. In this TEDMED conversation, she joins Jay Walker to examine a moment that is both technically extraordinary and socially unresolved. Generative AI may make health care tools more testable, more personalized, and more responsive to real life.

It may also bring the lessons of social media into a far more intimate arena where phones, wearables, synthetic voices, and behavioral data begin to shape how people seek help and how systems respond. The central tension is not whether AI will move quickly. It already is. The question is whether speed can be paired with judgment.

Together, Jay and Deborah explore what it would mean to use these tools with ambition, caution, and enough public honesty to avoid mistaking power for progress.

Good morning, Deborah. Good to see you.

Great to see you as well.

So, you know, I like to start by zooming out to the thirty five thousand feet level where we’re looking down from the top. And from where you sit, how would you describe how the world is changing in the areas that you focus on?

So a lot is going on.

It’s sort of like a medical situation where you have lots of comorbidities lots of things playing out at the same time.

Clearly, the big thing in the room is AI and what’s happened with GenAI over the last few years in particular.

And there is the I think the content base of how there are really wonderful opportunities for making things that were very difficult more readily available, more readily deployable, more readily testable, more studyable. So from a research perspective and an applied research perspective, there are so many more things that we can rapidly try in realistic settings. And I think that gives me great hope for things that in my opinion have taken way too long to play out and actually intervene in care.

So an acceleration of testability Yeah. And an acceleration of repeatability as well as testability since you can do the same experiment many more times?

Yes. With challenges to repeatability, and we have to grow our minds to decide what is repeatable and make use of the same technology to understand that repeatability is a relative concept and that things don’t have to be exactly the same because in this world, things are not ever going to be exactly the same since they are, you know, data driven and much more personalized. So it’s that opportunity, and it’s also a challenge to design it right, do it right, fund it right, and regulate it right.

So speed, which is an essential dimension to all progress, allows you to have more iterations and more quickly try things that need experimental validation.

Yes. I also might quibble with speed being always an improvement at all states. So speed is a relative thing. And so this can allow us to support interacting with individual clinicians and patients in ways that relate to their particular situation and really take on that complexity because the technology helps us do that in real time and in real context. But we also need to test those things just as we’ve been testing the impact of drugs. We need to test them in a smart way to understand what their downstream implications are.

And speed is important, but not skipping steps.

So the advent of a dramatic new tool, and we’ll just call AI, gen AI a tool, the advent of a dramatic new tool always leads to dramatic changes in how we perceive the world historically. So in the simplest case, we invent a microscope and suddenly we can see the micro world that’s invisible to us, and all of mankind’s understanding suddenly gets expanded by the increasing power of resolving microscopes.

Yep.

AI sounds like it’s at that kind of a leverage point, only this time the leverage is against information as opposed to optics.

And maybe also against interaction. Right. And maybe that’s the more disruptive place where it sits.

Let’s explore that. By interaction, you mean by the fact that it’s a dynamic process between, let’s just call it a patient and a system, and how that patient and that system are interacting.

The AI allows us to continuously see and or test against that dynamic interaction.

Yes, and then you can repeat the same sentence with the clinician, and with the payer and with the various parties in our complex healthcare system.

Has there ever been anything in the past that has had this kind of breadth of potential impact?

I don’t know how to play that balance because everything else that has had that major impact, we’ve had centuries or at least many many decades with which to experience and reflect and analyze.

And this is just at the beginning of it happening.

So I think it’s a very interesting question and one that I have trouble grappling with. In retrospect, obviously, telecommunications and the Internet and railroad and energy, you know, many things. And this is an integration of all those things. Yep.

If not for the Internet and the data that that began, this would have had trouble growing. So it’s tremendous impact. Much of that impact is a great opportunity for doing good things. And with as with many things, I think it has to be done with care.

Not always with the greatest speed.

So if you look at AI essentially within the lens just for a second because at TEDMED we like to just sort of step out and say, have we seen this movie before?

And if so, what have we learned from the previous iterations of the movie so we can avoid making the mistakes we’ve made and make new ones? Yeah. Right? You could say the last great revolution in this space was the computer revolution, the data processing revolution, our ability to take data in large data sets, manage it, interpret it, analyze it, store it, you know, compare it, right?

So now we we’ve got that architecture of storing information in forms that we can share, use, and we have reliable ability to manage and now we’re adding an intelligence quote unquote layer to that which is truly new. The software can analyze a thousand years of analytics in thirty seconds, and it would take humans a thousand years to go through that dataset, right, looking for patterns. So that’s one dimension of it. It appears that another dimension of it is the mobile phone, which reaches and touches every human.

So it’s not operating in some mainframe kind of way, It’s actually operating in my hand kind of way.

So that has to be a a big change because you don’t have to distribute devices. Everybody has a device.

Yes. Huge.

And so those are both enablers

To this particular revolution. The potential for the benefits of it across the sciences and discovery and the day to day struggles that people have getting information or navigating information are quite clear.

For better or worse, when I think about where we need lessons, lessons are probably where how do we do this smartly.

And there’s a much more recent lesson, I would say, which we probably could have been smarter about. Well, I wish we could have been smarter about. Right? There are many political forces. I would say it’s social media. There are very few people these days that are entirely happy with the path that social media took And the implications, no one wants to get rid of our ability to connect and share and what we used to call crowdsource and all kinds of things, but many, many people are concerned about where it’s come to with you with all of us in this attention economy, and many have written about this. And I would wish I I’m not predicting.

There’s a difference between my wishes and my prediction. I would wish that we would take some lessons from that and do this smarter.

Yeah. The unexpected damage that social media is clearly causing, right, which is not a reduction of the positives, it’s an offset to the positives, right? So the damage that social media is causing, its unrestrained use, its addictiveness, its ability to create alternate realities that are dysfunctional, its ability to clearly disable parts of our, you know, sort of neuro processing pathways, ability to amplify hatred or amplify dopamine, amplify other addictive behaviors, these things serve as, in some ways, you’re saying, a warning for how AI could also turn out to have unexpected consequences times ten because AI itself is capable of initiating novel behavior independent of being programmed to do so.

Right. And it’s a computing infrastructure, the internet, the mobile phone, and social media. AI is not coming into a prior era, it’s coming into this where all those are in place. There’s so much that we have to benefit from it, and I hope we for the sake of society, commerce, and and and progress, I hope we invest in doing it smarter.

I’m biased. I’m in a university. It turned out I’ve been to university my whole life.

Hope is not a strategy there, Deborah.

Yes. Yeah. And the university has a role to play. And this is a pretty crazy time to keep things going in universities. The young faculty members who are right in the center of developing the newest and latest algorithmic advances that are going to continue to drive what AI can constructively do to have impact, they’re increasingly drawn to industry. They were drawn there before. This isn’t even a paycheck issue.

Right.

This is an issue of they want to have impact on where the technology is happening, intellectual impact.

And it’s just a draw because the resources are there and such.

And that’s great in many contexts. Always most of our students have gone to industry. Most of my PhD students for decades have gone to industry. I’m relatively applied.

We still need work going on outside of specific business models, And that’s what happens in the university.

So let’s add one more piece to the puzzle. There is an additional technology component that is largely overlooked but is about to become the default technology, which is natural language speaking. The fact is that you can talk to the AI and the AI can talk to you. As of now, these type of synthetic conversations, which is what they are, are largely constrained in sort of a chatbot mindset.

But that’s not how they’re going to be. These synthetic conversations fairly soon will be undetectable from having a human to human conversation. I mean completely undetectable In the same way that deepfakes are undetectable now with anything but the most sensitive instruments on video. So not only do we have a revolution in applying intelligence, applying data processing, but we can talk to it.

And it can talk back to us. Yes. Which means that anybody anywhere now has direct access to the software layer.

Whereas before only the priests who could write code had access.

So that is probably going to be an asymmetric kind of change where anybody can essentially not only talk to the intelligence, but can have it write code for you? How do you see that amplifying the whole story?

So I wanna pull it into two parts.

The first part is about the difficulty in differentiating when you’re interacting in full audio video mode as we are now with the actual jaywalker versus the synthetic jaywalker.

And that one comes back to this point that AI is about even more than perhaps information is about interaction. It’s intervening in our interactions, including with when we interact with synthetic agents through them getting information. And how it is that as humans have evolved, how we respond to audio and visual information. We also respond to text, and there’s text that we would read in a traditional publication.

There’s text that’s chatted to us. There’s already a difference there in how we respond to it. And certainly, when we are largely evolved to respond to each other, And we’ve shifted pretty well from it being in person to it being this sort of of virtual mode. And the opportunities are again very large and tremendous and deep and broad in terms of the places where that will be not just an issue of efficiency but be able to provide accessibility that could not be achieved otherwise.

And figuring out how to do that right.

And so do we want to have synthetic entities always apparently and clearly and explicitly synthetic when they’re interacting with young children, when they’re interacting with the much older in between you could say caveat emptor or not. There’s a whole bunch there we have yet to grapple with. And then you raised the second issue which is that we in the past, things were delivered as products of one form where there were groups of developers, of range of developers, what it took to deploy and deliver a product.

And when individuals remember in the day when individuals were first able to put up their website? When they were first able to put up what became a YouTube video?

And the notion that individuals can then program an artificial entity in a synthetic form quite easily, they’re using a lot of product that was created, but then they can control the content. And so how that will all play out, it turned out that the development of those individual websites was sort of a quaint thing that helped drive things initially, but then became probably not that high impact. Whereas the whole very short form video format had a huge play in terms of social media impact.

And I would expect this next generation will as well.

You can tell a human all day long that astrology is fake and has no basis in anything you ought to pay attention to, and yet we will know that a hundred million humans will read their horoscope and decide to behave differently today because of it.

Humans are just very strange creatures. You’ve described gen AI as a fundamentally new generation of research tools that need to be rethought and used smart, but way more powerful than tools that have come before.

And at the same time, you’ve said, look, mobile phones are just as much a part of that system because we’re wearing the mobile connection, is listening to us twenty four seven unless we don’t want it to, but even then there’s limits to what that means. And on forget about the quantified self, your phone is more than enough horsepower between its microphone, its cameras, its sensitivity, the fact that it’s always with you, right, it is the ultimate monitoring tool. And then you add the fact that natural language is the mechanism by which you communicate and it communicates with you, suddenly those multipliers seem like everything before that’s gonna look like the dark ages pretty soon.

Yes.

I will say that when it comes to health care, since you mentioned the role of the mobile phones and its proximity and just how powerful it is, there is still a gap that we’ve made very little progress in filling that I do believe is going to be assisted by the introduction of this generative AI tool, I think it’s going to be interesting to see how much our innovation ecosystem actually helps that along. In other words, it is still unclear to clinicians whether they’re primary care, specialty care, emergency doctor, how to use the fact that their patients have the mobile and the wearables.

Yes. There are certain things there for diagnostics that come out of the box if you have a continuous glucose monitor, if you’re looking at AFib.

But the longitudinal story that data from these devices could tell with respect to iterative care, drug titration, adjusting recovery from a surgery, recovery from infectious disease, The work of collecting that data using LLMs to help identify what among that noisy longitudinal data is actually relevant for a particular health care interaction. What’s relevant to the patient? What’s relevant to the caregiver? What’s relevant to the clinician? That work still needs to be done. Otherwise, the clinician won’t have a system that’s telling it how to use that longitudinal data.

Unfortunately, it has to be over validated because validation alone won’t be enough. Because the first do no harm principle and the liability economy we have says, well great. If you have this tool and it helps ninety nine people and one person is hurt, that person’s gonna sue you and you won’t make enough from the ninety nine to commercially offset the one that sues you, so you better just stop trying. I’d like to point out though, we can almost divide the data into two areas.

In one area, I’m gonna call that chemistry. Is the answer you’re trying to find ultimately going to affect a chemistry? A medicine you’re gonna prescribe? A therapeutic, where you’re actually trying to modify for good the chemistry of the situation.

But here’s the other, mental health, which has been the black hole of all of health and medicine in the United States. The undertreated massively and yet at the same time the one that most people who are observers of the system would say, we are suffering a mental health crisis unlike any in the history of mankind, both in its breadth, its depth, and its, you know, intensity.

Here’s an area where really I just need to hear you. Right? Hear how you speak, who you speak to, and also listen right below the surface, maybe stress levels, cadence, other things that are embedded in your mental wellness that show up in your language or lack thereof, could turn out in the mental health zone, we have the first real tool that can assess or provide data for real mental health evaluation.

Such an interesting one. And one of the first ones we thought about, what’s now like almost decades about, at a place for it to intervene.

Whether it’s thinking about how we might use this to help in adjusting and diagnosing, adjusting medications for ADHD, for example, back in the Perfect.

Yeah.

But that is also ties into chemistry.

The problem that we’ve seen, because I expected what you said, to have already had more of that effect than it has had, there’s so much context in mental health. And with all that measurement, the context and the confounders are maybe the hardest place.

It’s interesting you say that because AI can make a difference there too. Let’s just play a science fiction game. I give permission to my phone to record every listen to everything I say all the time and no matter whether it’s text, conversations, whatever. And I give that because I’m confident it’s going to help me deal with the mental health issue I’m having.

Some, let’s just say, some uncontrolled fears I’m trying to deal with. Yeah. Right? The ability for the AI to listen to absolutely everything and process every bit of context, every type of conversation I’m having, every little piece of data that can be captured by listening to my voice now is processable by an AI at volume and speed that has no cost.

So you could literally have your own twenty fourseven therapeutic interpretation going on despite the fact that that’s dystopian in its sound.

In a research setting it might actually be able to have an n of one for mental health for the first time.

So and I would say that’s been true for a decade at least.

It has actually been true for a decade. There are these minor bits of friction such as the person who is hearing voices and has fears of paranoia. This is not so much a compatible tool. But let’s put that aside. The experiences that that person had ten years ago is not necessarily monitorable.

Because I expected it to be that end of one tool and mental health would be one of the first places it was because you can’t do a blood test. You can’t run around with an MRI continuously going on. It seemed like it would fill that gap, and it’s been very hard for it to actually do so. And in part, because people wanted a more general broad model that would do that diagnosis and work across individuals.

I don’t think we necessarily had the patience to say, let’s build a tool initially that makes the therapist’s job faster, easier, some of it done through asynchronous vehicles so that you make the therapist smarter, but there’s still some human there who has a longer history on the individual. Now if you start to say we capture family history, we do that history, moving the history into that whole process, I think there’ll be continued to be development there.

I don’t understand why we didn’t see some of that progress already.

Well, in all fairness, Gen AI is only now getting to be good enough to listen to five hundred hours of my speaking over the course of a month, categorize it, interpret it, score it, grade it, you pick it. You know, yes, it can’t read my mind, so it can’t see the causal you know, psychiatric baseline for so much of this stuff, but it’s going to be able for the first time to interpret how I am behaving all of the time, including when nobody’s watching how I’m behaving. It can’t yet read my thoughts, though ironically in an MRI you can now start to read dream patterns and map them. But you know, we’re not close enough that we can talk about thought reading in the kind of way that would be psychiatrically useful, though it can now control a coffee cup by thinking.

So And the behavioral component is huge.

I don’t know, just a few months ago, Apple developed some foundation models that they refer to as behavior as opposed to just focusing in on particular heart rate variability and such. They developed a slightly higher semantic foundation model for individual behavior. And that has a component in this mental health story alongside the voice piece, but in people’s behavior that is not necessarily one you hear, but that is witnessed.

We did some early work quite a while ago with some Google takeout data and folks at Northwell who looked at the online behavior of individuals before they were hospitalized for a first psychotic episode who agreed to have their historical data shared. And mostly what you saw was this shift in their temporal behavior. Late night and you saw a major shift of that sort that they clearly weren’t sleeping because they were online. So the richness of this, I very much agree with you. I think mental health is just so tremendously difficult in its origin. Maybe as we get to parse it out, we’ll be able to do a better job.

Well, psychiatry is a talking cure, right, at some level. Obviously it can be pharmacological as well. But given the fact that all of us will be able to talk to another human twenty four hours a day that is synthetic, And given that that human will learn about us and will behave exactly as far as we’re concerned from a Turing test standpoint, just like another human. Whether or not that other human is clinically authorized to do things Yeah.

That won’t stop me from having a friend. Artificial intelligence in many ways is a new species. It’s not living, but it’s a species. It’s capable of making decisions.

It’s capable of acting on those decisions. It’s capable of doing things we don’t understand and don’t know why.

And I think we are obligated as the actual sentient humans Right.

We are obligated.

To not worry about being popular and to simultaneously do important things with the technology as safely as we can and speak our mind as to what this actually is and what’s better and worse the same way we raise our children.

Yeah. Well, Deborah, you have the extra responsibility of being the canary in the coal mine because you are at the core of the academy who are looking to use these tools in thoughtful, safe, what you call smart ways. At TEDMED, we are committed to elevating that dialogue regardless of where it goes and regardless of who likes it or does not like it. Yeah. Because if we don’t speak up, we know this movie will end very badly. So our goal is to, as best we can, see if we can’t have like minded people from both sides or all sides of the question say, yes, this isn’t perfect, but it is critically important to talk about.

Yes. Thank you.

Thank you for that, and I look forward to talking with you more, but we can both agree we live in both amazing, exciting, and frightening times as technology impacts our world.

Yes. Thank you very much.

Alright. Good chatting with you today.

Take care.