Is AI Here To Replace Us Or Work With Us?

Dr. Niloufar Salehi, an assistant professor at UC Berkeley who works on human computer interaction and is chief product officer at Across AI, on why AI should support human judgment rather than replace it. She discusses AI in medicine, translation in hospitals, forensic evidence in the courtroom, synthetic data and how YouTube creators think about the algorithm. The hosts then discuss what struck them from the conversation.

Watch on YouTube

Transcript

Today we are tackling one of the most critical shifts in AI, agentic AI. It's a new approach that enhances human agency instead of replacing it. We'll explore how this technology is reshaping industries like healthcare, law, and the creative world and what it means for the future of human AI collaboration. We'll be talking to Dr. Niloufar Salehi, a leading researcher in human computer interaction or HCI and focuses on agentic AI.

We look at what people are doing in a system and we think okay the way to improve the system is to take what the person is doing and have a computer do that and stuff and that completely misidentifies where the opportunity actually is and what the strengths of people are and what AI is actually capable of.

Now I am delighted to be joined here in our DC studio by my colleagues Laila Rizvi and Stephen Horn. Pleasure to have you both here today.

Great to be here.

All right. Have we both heard of agentic AI?

Well, have now.

And our thoughts are.

Well, well, so what I'm particularly I'm I'm really looking forward to speaking to our our guests today because the human computer interaction piece here is is is kind of something that I've often been thinking about, which is it's fine to have the computers doing what they're doing, but how you have to interact with humans as part of the job that needs to be done. And some of the work that she's done in health care, some of the work that she's done in the legal system. I'm particularly interested some of the work that she's doing with restorative justice. So, so I think it's I think it'll be great to understand, you know, how we can improve AI to actually work with humans, which is what we're talking about today. Where are you on this level?

I I think it's so important. I think actually human AI interaction and how how easy it is for humans to interact with AI is the thing that kind of set everything off or the average person something like ChatGPT was was really easy to use and really intuitive. I think every technological revolution we've had has come with frankly just great user interface and usability. The iPhone it's just then it's the design. I think that's what's going to be be critical

when it comes to AI users and to really kind of the general population. Do you think that there is an understanding between AI that empowers versus AI that replaces?

Good question. That's a very good question.

I mean, do you think that people are able to kind of um discern the two? I I don't know because I don't know if there is AI which fully replaces because you always I mean you always have someone that is giving that is empowering that AI. Do you know what I mean? So I think I think we it hasn't emerged yet what is a framework I I think we will have a framework where you have to have a certain level of kind of training to deal with different levels of agentic AI.

Right, but I think it's what we talked about in previous episodes, right, which at the moment is AI replaces functions, it doesn't replace people, and maybe that's what will carry on or maybe we're just such at an early stage now that you know when when AI gets more agency as we're talking about today then it can replace people. At the moment it just does jobs and has very limited agency.

AI is still currently stuck in our lap and we haven't seen evolution in terms of hardware. So if you start having AI that can talk to you that's in a robotic body that can move physically. It's multimodal. It can see. It can hear. It can respond.

Well to some extent we have that, you know, with self-driving cars. I don't think that's a long way away where you you combine AI with a physical function. I don't think that's a long way away, but at the moment it's still about doing a specific task.

So with AI advancing at a breakneck pace, Dr. Niloufar Salehi's work highlights the urgent need for AI systems that empower users rather than automate them out of the decision-making process.

Such a pleasure to have you with us today.

Thank you for having me.

All right. So, I want to jump right into your research because it spans multiple fields. We're talking about education, health care, criminal justice, and social computing. What first drew you to human computer interaction, and how do you see it shaping AI's role in society?

I started off actually being much more interested in computers. So, um started programming, did competitive programming since high school. Um, and during my undergrad, actually, I learned more and more about the impact that computers had on society. And that just made me so much more interested in that intersection. Uh, I'd always been interested in programming and the computers themselves and artificial intelligence. And HCI gave me an area where I could really start to understand what the effects of that are on people and how can we design them to have more positive societal impact.

Um, so that's what drew me to HCI and um, one of the things I love about the field is that it's very interdisciplinary because I I don't believe that you can really deeply study those problems or really have a positive impact unless you bring lots of different disciplines including uh, social sciences and humanities together with um, computing uh, to be able to actually really study, understand and and design for better futures. So that's what brought me to the field.

And then um when I started as a professor at UC Berkeley, I um really went after the questions that sounded interesting and impactful to me at a societal level and and then started to look at what is the impact of computing there. So um we looked at school assignment algorithms for instance across the United States um school districts use matching algorithms to assign students to public schools. That has massive impacts on where hundreds of thousands of kids receive an education and how they receive that education and it's using computers. Uh we looked at the criminal legal system and how um data is showing up much more and more in the courtroom. Um and so those are the types of questions that really excite me and interest me are when there's a human and societal uh aspect to it as well as a computational one.

Well, thank you again for for joining us. It's it's really great to talk to you and this whole sort of HCI area is is, you know, it's really really fascinating. We don't always give a lot of thought, do we, to how how humans and and and computers actually interact. We talk an awful lot about the sort of capabilities of computers to do this or to do that but and so what do you think are some of the biggest misconceptions about AI's role in the workplace today or or decision-m

I think the biggest misconception and stemming from the fact that most of the people who build the systems don't know a whole lot about humans

Or humans don't know an awful lot about the systems as we found out today.

That's completely true. Um is that we we look at what people are doing in a system and we think okay the way to improve the system is to take what the person is doing and have a computer do that instead and that completely, you know, misidentifies where the opportunity actually is and what the strengths of people are and what AI is actually capable of and this is not something new. In fact the field of HCI in part, um, at least some waves of it come from this many in the 60s and 70s.

Um so there's this story of um Xerox PARC was doing really really well. They were selling Xerox machine everywhere in the world and um they had, you know, a few computer scientists including one that later won the Turing Award think about how do we make the Xerox machine simpler to use because it was actually really complicated. I don't know if you have any experience with a Xerox machine, but I remember when I was in school, it was very there were so many buttons. And so the these AI researchers were thinking if we could just classify people's intent and predict what they're trying to accomplish with the machine, then the machine should be able to do it itself with just one button. And so this whole project around let's make a Xerox machine that has just one button um was born.

And they spent many many uh years working on trying to get this to actually work. And and in theory it should because unlike self-driving cars where there's just so many different parameters, Xerox machine, there's not actually that many possible states. And so they started to draw out here's all the possible things that someone might want to do with a Xerox machine. Here's their starting point. Here's where they're going. Here's how we identify their intent. Maybe they have a simple conversation with the machine. And the more they draw drew this out, the more it didn't work. And and the the more they realize that actually intent classification is really difficult in part because people don't always know their intent and then when there is an error or a bug repairing that error or bug is really difficult.

Um Lucy Suchman who was a PhD student in anthropology at UC Berkeley at the time was doing an internship at Xerox PARC and she did a project around this one button Xerox machine that turned into a book. It's called Plans and Situated Actions where she talks about how in the AI researchers' minds people came to the Xerox machine with a plan and they wanted to implement that plan and as long as the AI knew what that plan is, it could do it for them. But really the way people work is that they come in and they have some idea of where they want to end up. She uses the example of you're going to cross a river with a canoe and you're on one side of it and you don't really like plan out every every step of the way. You have an idea of where you're headed and you move forward a little bit and then the current takes you sideways so you like revert and go back and forth and within the situation that you're in you take the next step that gets you closer to your goal and because of that messiness of how people work it was basically impossible for the AI to replicate what they were trying to do.

Um so that that book that I mentioned, Plans and Situated Actions, is one of the foundational books of the field of HCI. There's been many many studies over the years of HCI in the workplace, many failed attempts said, uh, you know, automation and um putting in computers in different kinds of workplaces that HCI has then studied why and how did that work or not work. So I I I told you this little story to tell you that this is not something that's new. Um but the misconception has remained where now we go in and we say okay what is the work that the person is doing? How do I build an AI model that replicates that? And I think that's the biggest misconception.

Sure. And I I want to move on to kind of what the industry focuses on right now in AI which is very much increasing efficiency, productivity, automation and kind of replacement as well. Whereas your work emphasizes AI that enhances human agency. Um what are some examples of AI supporting rather than replacing human decision-m? I

I would refer to the field of medicine. Um so over at le at at least for a decade models have existed that um diagnose patients and physicians have been very resistant to them for various reasons. Um this is actually one of the things where I think you see a really clear example of the people developing the AI just look at what is the physician doing and try to replicate it with a model and that just doesn't work within the ecosystem of doctors and nurses and many signals that are not in the data.

So um I'll give you an example. Um there's been a lot of work done on taking a photo from a a heart scan. So you know the heart scan that goes up and down, taking that photo and converting that into a diagnosis

that is taking it to sort of an extreme where you actually can get the signal right from the machine but because what the doctor does is look at the photo, that print out from the machine and diagnose, the model is also trying to do exactly that. Look at the photo and diagnose. Whereas you could, you know, use lot a lot more signals if you just thought about what is the actual problem I want to solve. Um, so these diagnosis models haven't really seen a lot of use.

Um, but the moment where you had transcription models that could be in the office and transcribe the the session with the doctor that just took off in less than a year. Uh, you I've been to multiple doctor's appointments since then where the meeting is being recorded and the transcription automatically gets uploaded. Um, that's something that doctors didn't enjoy doing. Patients didn't want their doctor staring at a computer and typing the whole time. Um, and because it was so useful within that workflow with the patient and the physician, it it took off and it's being used a lot now.

Uh, that's not to say that I don't think there's a place for diagnosis models. I've actually worked on especially models that diagnose rare conditions because that's another place where I think the model can actually have a big impact. If you take what the people are trying to accomplish and see what are the the gaps and places where these models can be really effective and usually it's something that's either tedious no one wants to do or it's something that requires understanding patterns from a lot of data that is really hard for people to do. Those are the places where you can really impact and enhance what the person is trying to accomplish. But the work that people do is so complicated, so messy, requires so much context um that it's actually really challenging and that's why you see so many AI startups take off and then just flounder right after because they see so much churn because people come back and say, "Hey, it's not working."

A lot of your work has been on uh AI in in health settings as you said, you know, and a lot of times at the moment AI is used for translation. Uh how dangerous can that be in a health setting and how can you put safeguards in place to make sure that, you know, uh it it doesn't go wrong?

Back when I started doing this work um around 2020, the this was before ChatGPT, um Google Translate used in UCSF hospital to translate English to Chinese had about 20% error rate uh without 8% with about 8% of that being errors that are clinically significant, that can cause clinically significant harm. The models have gotten much better and part of my research has always been about this question of how do we know that they're good enough and how do we create them in a way that they're actually usable in a reliable way and so I I'm sure if we ran the study again now it the error rates would be much lower so some of that is just the models get better over time but to me the the big problem is that we had no idea.

So um right now translation models are being used across the United States at least as far as I know, probably worldwide, in really high stakes settings like hospitals, courtrooms and there isn't any systematic evaluation being done, any benchmarking being done so that we at least can point to something and say if the model is this good, if it passes these qualifications then it's okay to use it and then here is who is responsible if they used it and and it was wrong because that's the other side of that is the liability. Right now because there is no official guidelines or guidance um ultimately it's up to the physician who has to decide in that moment whether whether they're going to circum, you know, try to at least uh talk to their patient in a way that they can understand quickly um or, you know, not do that. And there's liability either way.

Um and so we need better systematic um reviews. We need agencies. Um I I often think of the FDA as an example of that. Uh we need ways that these models are constantly benchmarked and evaluated and given standards. Now the the challenging part of that is the non-deterministic nature of the models where they learn from the benchmark. Um so unlike running a clinical trial where if you run it today it's going to be the same as if you run it in three years, more or less it will have the same impact, that drug will have the same impact on those people's bodies. Um, these models are really tricky to evaluate because they learn from the benchmark and you constantly have to change your benchmark and so I think we need to put a lot of resources into making sure those evaluations are done.

I want to pivot from the clinical setting to the courtroom for just a second because um, you've also, you know, worked with the criminal justice system and you've studied how public defenders actually have a problem and struggle to challenge AI generated forensic evidence in court. One, why is it such a a problem to challenge that type of evidence? And two, what does this tell us about uh the transparency needed when it comes to these types of high-stakes situations?

There's also almost no evaluation done on uh models used in the courtroom. So uh when a judge sees that for instance this forensic evidence says that this person was the person who committed the crime, um it's usually coming from a model that gives some kind of statistic analysis. It's very difficult for the judge to know how reliable is that really. There are some standards that all of these um, you know, people who provide these kinds of systems have to pass. What those requirements require you to do are different from what would actually be needed to to be able to tell that this is a reliable way to judge someone.

Um for instance, you might have tested your DNA evidence forensic software on a certain population but then run it on a different population. Um so that is not in the standards and that would completely change the reliability of your tool. Um another example is gunshot detection models. They're usually trained in suburban areas where loud noises are infrequent, but when they're tested or used in an urban area, they might mistake a lot of different noises for gunshot. So, this is again the exact same problem of evaluation, which is a really, really difficult one, uh, which is we need standardized ways to evaluate.

Now, when we don't have those standardized ways to evaluate, and you're a public defender, you're up um against certain claims are being made based off of certain models. And over the past decade, courtrooms have increasingly moved away from eyewitnesses and accepting that as evidence towards just looking at data and thinking that anything that is relevant to data is more objective, more reliable. So now you have um a prosecutor with access to a lot of data. Most companies work directly with police and prosecutions. Sometimes they even have a direct phone line for prosecutors to call them and request data. Um the prosecutor also has a lot more resources to run analyses on that data and come up with some sort of conclusion. But when you're on the other side of that and trying to defend that person, you have very little resources. Um you have very little time because you're doing this for a lot of different people. And at the end of the day, it's less about who has the data. U technically you're supposed to share the data um in discovery, but more about who has the resources to tell a story with that data. Um and that usually ends up being the person who has the more resources. So there there was a lot of challenges there in uh really identifying and making sure that due process is actually happening.

I want to get into the weeds a little bit more now and talk about synthetic data and how it's being used to improve AI performance and reduce bias, but it also raises some ethical concerns. What do you see as being the biggest risks and potential benefits of data synthesis um in AI development?

I think the biggest benefit is that you can make the models much better with much less. Um, so if you look at how DeepSeek was trained, it was at a fraction of what the other models were trained, but in part it actually used what other models already had done. So it used data from other models to train itself.

So you can get much better much faster with less resources. The challenge is that because you're not getting any signals from the real world, um you risk getting further and further away from what makes sense to people. Um and we see this happen a lot like if you train a vision, an image generation model to generate people and then uses outputs to generate, to create another model and then you keep doing, after a while you have not real people. Um so you get further and further away from the real world.

Um there was this really interesting experiment that they ran at Meta where they had um these language models play the Diplomacy game together. The Diplomacy game is a board game where you are given certain territories on the board and then you have to go to war together and get more territory. So they had these models play the game together and they let it run for quite a while and after a while the models were still playing the game but they weren't speaking to each other in English any anymore and were basically sending gibberish tokens to each other. And so that's the one of the dangers is we get further and further away from a thing where people have any can have, you know, influence and visibility into what the system is actually doing and sort of steering it more and more towards what makes sense to people.

Um, so I think what that's one of the challenges. There are ways to use that as you said to mitigate some biases because when I say bringing closer to what people makes sense to people has an upside which is it makes sense to people. The downside is that people have a lot of bias and any people generated data can have lots of problems of its own. Um, and so there are ways to use that synthetic data to to mitigate some of that.

Okay. I want to get into something really fascinating and your research on YouTube creators and how you have identified three different personas. Is we have the agent, the gatekeeper and the drug dealer. One, can you start by explaining the characteristics of those three different personas? And then also um what does this reveal to you about the impact of AI-driven platforms on creative work?

So this is a study that um two of my students did a few years ago. Uh they started to interview people who were doing content creation on YouTube and tried to understand how are they viewing the algorithm and how the algorithm, so that this is the algorithm that creates the the feed for people, right,

how is the how do they view that and its impact on their work and the more we talked to people the more we realized that they were really assigning personas to the algorithm which we were very surprised by. Um they were almost describing it as they would describe another human and they had these certain personas. Different people used the different personas. Some people used multiple different personas, but three of the main ones that we saw were agents where they saw the algorithm as as almost their own personal personal talent agent where they would say things like the algorithm helps me find viewers. The algorithm, you know, shows my videos to the people who would be interested. So really thinking of it as as almost their own talent agent.

Um, the other one was gatekeeper where they saw it as sort of an adversary that was, you know, gatekeeping the views and so they had to work with or against it to get to those views. Um, and then the third one, the drug dealer was really the nefarious view of the algorithm is trying to get the viewers hooked on watching more and more videos. Um, and so sometimes again they they thought about strategies to work with or against that persona. Um, maybe making my videos shorter so it'll think that more people are going to get hooked on them, you know, all these ideas that they had that they were sort of trying to figure out how to work with this algorithm.

This was like almost I think six years ago. So it was a long time before we were talking about agents or anything like that. But it was an early glimpse into people whose livelihoods depends on how an algorithm acts. How do they start to figure out what the algorithm is? What are their perceptions of it? And then how do they work with or against it? Um, and so those personas really helped us understand that.

Um, I think the flip side of that is that I I personally think that we shouldn't anthropomorphize algorithms. I think that we should think of them as systems. Um, but people were doing that anyway. They were definitely talking about them as if they were people within an ecosystem. And so that creates a little bit of a dilemma of how how should we design these systems? Should we lean into the fact that people assign them personas anyway or should we, you know, really stop them from being anthropomorphized because there are downsides to that as well. For instance, then when the model does something wrong or perceived as bad, then they react to it the way they would a person doing that without realizing that this is just a model or an algorithm. Um, so that's an overview of what that work was.

Um, so I think we're going to see more and more of that collaboration almost or that work with and against the models. Um, I think a lot of creatives now are using image generation models, other kinds of models. They're also a lot who are very much opposed to their work being used to um, train models and I I completely understand that. I think that's a very valid view. So I think we'll see the same back and forth of assigning a role or persona to the algorithm within this ecosystem of producing and and acquiring and uh, you know, viewing art. Um and then some some that will try to work with it and some that will try to work against it.

Yeah. I I also just want to jump on that um and ask, obviously the difficulty is often understanding and working with the algorithm and trying to trying to outsmart it. Do you think creatives will start to use AI to effectively better understand the algorithms and leverage the AI as their tech and their translator for basically other AI AI models?

Yeah, I I I think they will. We already were seeing that with the YouTube algorithm. So we um saw people sort of A/B test the algorithm to see what it was actually reacting to. So they would put up a video with multiple different titles or thumbnails and then see which one the algorithm was showing more to people. Um they would share that information with other people there. Um there's this, at least at the time there was a a convention for YouTubers every year and we looked and every year there was at least one session just about the algorithm um and trying to understand it and work with it. So, um I I think that definitely happens. You know, once you have a a thing that impacts you and your work, you become very invested in understanding and figuring out how to work with and against it.

That'd be the session I'd be in, trying to understand the algorithm.

Let me take you back to HCI just for a a second. And uh, you know, what we talk about a lot is is roles that uh, you know, computers are better at doing. We talk about functions or roles that humans are better at doing. Though not everybody agrees which is which, but but to your mind what do you think where we are at the moment that uh not only that but which roles are computers doing best, which roles are humans doing best, but also how do we create the interface, how do we create something where we're actually working to both strengths as opposed to pitting one against the other?

Rather than think about what are computers better at and what are people better at, I tend to think about what is it that computers can do because there is a a whole lot of things that people can do that are basically I think impossible to have a model or a computer do. Um we have to understand that even when we think ChatGPT is making sense or understanding what it's saying, it's not really. Um all it's doing is finding patterns in language created by other people and replicating those patterns and spitting them out. So it's not really understanding anything. And so any work requires sense making and understanding and knowing what the context is and what the goals are and how to achieve them. And that's something that only people can do. But these models can be used as tools that are very useful. Um the things that I think they're very good at is finding patterns in lots and lots of data, right?

Um this is something that is basically impossible for a person to do at the scale of data that we're thinking about. Uh so whenever you have lots of data and you're trying to find a pattern in it, when you're looking at lots of um health data and trying to find correlations between different things in a certain medical condition, when you're looking at lot lots of different ways to fold proteins and trying to figure out what is the best one that will have the desired effect that I want, that's something that I think the models can uniquely do.

And I think that interface is a really interesting question. If you think about, I I would always center the person and try to figure out what is it that they're trying to do and then what is, where are they doing that and then that's the right interface for the model to show up. So um for knowledge work I think a lot of that is we're in the browser a lot, we're in meetings a lot, um I think those are the right places for the models to show up just in time when they're needed and doing the thing that is actually relevant and useful. In a lot of manufacturing and other types of work, for instance in the hospital, that there are more um more more surfaces where the models and algorithms could start to show up actually where people are. Um, so for instance, when smart glasses first came out, I immediately thought this would be so useful for people in factories. Unfortunately, it didn't really end up there and they they tried to make it work for regular everyday use and there just wasn't that much of a use for it. But there are many surfaces where if you look at what is the job to be done and where are people when they're doing that job that will, you know, just take you straight to what is the interface where I need to show the outputs of the model.

I want to move on to more kind of high stakes decision making from hiring to loan approvals to policing which AI is increasingly being used um used in. What do you think we should be most concerned about when AI kind of replaces human judgment in these particular sectors and where do you think it's absolutely critical for humans to stay in the loop?

Yeah, I mean I would say people should always be in the loop for all critical decision-making. Um, you know, there are certain rules and regulations that we've all as a society agreed upon, like due process. Um, and it's impossible to say that we are achieving that if all we're doing is running a model that basically takes patterns from the past and tries to use them to predict the future. Um, there's this really great book that recently came out called AI Snake Oil that goes into a lot of detail into why is it that predicting the future is actually not always mathematically possible. But if you create a model that does that or at least says that it does that, it will always come up with an answer. And so it looks like it it did it. It said how likely is this person to, you know, return the loan and things like that. And there's some certain signal there for sure. But there's this other value that, you know, equal opportunity to loans for instance or due process when you're thinking about the criminal legal system. Um and for those to actually be achieved you have absolutely have to have a person.

Um now there are people who would say, you know, okay well maybe the model can give guidance to the person and the person makes the final decision. I think that's a bit of a gray area because at the end of the day the person will be largely influenced by what the model said. Um there is little evidence that people will, you know, systematically overrule what the model said and so we have to think and do a lot more work over what are the benefits to something like that and what are the risks and harms.

Would it make sense to design models so that they give you various options which might be slightly contradictory but allow you to make a informed decision for the final kind?

I I think that's one of the paths forward. U I agree with that. There's actually this uh paper that came out of UCLA that I loved which was um thinking about medical diagnosis and they had done this study of how do physicians describe or, you know, talk about their own diagnosis with other physicians and it's usually that they will say, you know, here's the, I think that this is the right diagnosis and here's the evidence or the clinical trials that I've seen where this has worked. Um and so they they tried to replicate that dynamic with a model where given the patient's current conditions, the model would come up with multiple different possible diagnosis. And then because it's a model and it has no ego, it also came up with evidence supporting and also rejecting its own hypothesis. Um, and I thought that was brilliant because then you you're giving the person all of the information, at least to the extent the model understands it, of why different things might be happening and then you're putting the person in a position where they're really going back to their contextual understanding, um their analysis of having seen the patient, their experiences and comparing that against everything that the model found to come up with a judgment. It's much harder to do that as a person if you just, you're just given one output and then you have to decide whether to use it or not.

All right, final question for you today. How do you think our relationship with AI will evolve over the next five years? And what do you think the biggest challenges are going to be?

I think we will be faced with a lot of challenges around hype and expectation setting. Um, you know, I think that keeps going up and down where we expect the model's going to do everything, then it doesn't do everything, then we're disappointed, then it starts to do some things. Um, I think as people we're just really, really bad at understanding what AI can and cannot do. Um, even I sometimes get surprised, and I've been working on this for 10 years, that ChatGPT does something or does not do something because our brains are just not wired to understand billions and billions of parameters and so we're going to have a lot of trial and error in understanding how is this useful and where is it potentially harmful. So we're going to have a lot of surprises there, I think a lot of disappointments as well.

I think the next 5 years are going to be very critical in AI entering the workforce actually. I think we're going to see that happen a lot and with it there will be, um, I think one of the things that was mentioned earlier that I thought was really interesting and I hadn't really thought about was um all those people who worked in companies that only knew how to do that, they were the only people who knew how to work the Xerox machine. Um and we're going to see the same thing with AI. It's going to be, you know, people, they might be in IT, they might be in in various functions of a company that know how to get the models to do the right thing. Um, and I think we're going to see new jobs shape around that. There are going to be people managing the AIs, making sure they have the right data, right? Uh, you know, putting guard rails around them. I think there's going to be a host of new jobs as more and more of the AI actually gets deployed in the workforce. But it's it's going to be um a learning uh effort for everyone and it's going to come with a lot of trials and tribulations and a lot of growth as well.

We can expect growing pains.

Yeah,

exactly. Expect growing pains.

All right. Well, thank you so much for your time today. Thanks for joining us and for sharing um all your insights and your research. We really appreciate it.

Yes, thank you. Thank you so much for great. Thank you. Bye.

I thought that, you know, the the translation component of AI and some of the problems that that can bring to the table as well as, you know, the criminal justice, um, you know, trying to challenge AI materials in court. That's something that I hadn't even thought of when it comes to all the different applications of AI. You know,

for sure. I didn't even realize AI was being used in

I didn't either. I didn't realize it was being used in courtrooms.

Yeah. I I had no idea. And I also think it it speaks to her point about the fact that the people creating these technologies do not understand the workflows of these disciplines that they've never had any experience in, like the courtroom.

Right. Well, it's the point that she's was making and that we've been making today which is that what are the safeguards? So yes, it happens in the courtroom all the time, particularly looking at pathological evidence, you know, where you're looking at data and this is what these models are good for, as she says, you know, understanding, but then somebody has to be able to pop in there and say hey, hold on a minute, you know, uh in this instance I don't think it's right. I think I think she was absolutely fascinating and I love the fact that she's talking about the sort of interaction of where people are and what the job to be done is and how things can work together and and just just just thought that she was awesome. The one thing I would say is, you know, she's still though basing on a world where we do agree what due process looks like.

Yeah. And that we're committed to due process.

I'd love that to be the case.

I also think it was really interesting when I asked her about synthetic data and she was talking about the paper from Meta where where these AIs just started talking in their own language. Is in that just if you leave it running long enough, you actually create this digital world and ecosystem that we have no insight into and we don't understand. And I think that to me was just

just an inch of what it could. Yeah. That you have this like compounding and snowball effect where the the line's meant to be going this direction, but you turn the angle and it actually goes.

What I'd have been really interested in there, and I don't know if you know the answer to it, I, but, you know, was to how well they actually played the game of, they were playing Diplomacy, which I've played for many years, and how well were they actually playing in not speaking English to each other.

Yeah.

Final question. If you were an influencer, if you were an influencer, do you think you would be an agent, a gatekeeper or a drug dealer based on the characteristics that she laid out?

I think I would be the drug dealer.

I also think I'd be the drug dealer. Yeah. I would be like, I'm on this platform to generate as many views as possible, so how can I keep doing that?

Yeah. But I don't think that's a good incentive and like societal benefit, but I do think that's where most people lie, which is the more because the the returns are massive.

It's a dopamine hit. Yeah,

I agree. All right. Fascinating conversation. Certainly our thanks to her today. Uh that does it for today's episode of Agents of Tech. If you liked our program today, be sure to like, subscribe, and share as that helps us to curate specific content just for you. Until next time, goodbye. Thanks for joining us.

Bye.

Agents of Tech is brought to you by WebsEdge. A special thanks today to our studio engineer Madison Bert, our editor Abby Harris, and our producer Kath Sheen.