Behind the AI in the Newsroom: The Washington Post’s Vineet Khosla
Me, Myself, and AI · 5 May 2026
Source: https://sloanreview.mit.edu/audio/behind-the-ai-in-the-newsroom-the-washington-posts-vineet-khosla
Transcript: https://podcasts.opennash.com/shows/me-myself-and-ai/behind-the-ai-in-the-newsroom-the-washington-posts-vineet-khosla-e6e4e2/ (via OpenNash Podcasts, machine-generated)

How can AI help companies meet customers where they are, especially when their behaviors and needs evolve quickly? Find out how one news outlet turns this challenge into an opportunity on today's episode.

I'm Vineet Khosla from the Washington Post, and you're listening to me, myself and AI.

Welcome to Me, myself and AI, a podcast from MIT Sloan Management Review, exploring the future of artificial intelligence. I'm Sam Ransbotham, professor of analytics at Boston College. I've been researching data analytics and AI at MIT SMR since 2014, with research articles, annual industry reports, case studies, and now thirteen seasons of podcast episodes. In each episode, corporate leaders, cutting-edge researchers, and AI policymakers join us to break down what separates AI hype from AI success. Hi listeners, today we're joined by Vineet Khosla, Chief Technology Officer at the Washington Post.

The Post isn't just a newsroom, it's a giant technology machine that delivers journalism to millions of people around the world every day. And Vinny leads the teams that build those systems behind the breaking news and the audience experience and security and AI. We're hoping based on the discussion today. So we'll talk about how technology is shaping journalism and maybe a little bit about what audiences don't see behind the scenes and what the future of news might look like.

Vinny, thanks for being here.

Thanks for having me, Sam. Been listening to your podcast for a while, so it's a pleasure to be finally on the other side of it.

Maybe we can talk a little bit about what happens behind the scenes with the podcast. Let's start with something that many listeners feel. I think consuming news in our modern world can be pretty overwhelming and fragmented and tough to understand. And that may be especially true for younger audience who are more raised in a different digital world than I was. So, from your side, what's maybe currently broken about how we're experiencing news and what needs to change?

The way I view it is there is not something broken about news. And if we zoom out, we should think about journalism as a discipline, not a format. When you start to think about it solely as a format, it might seem broken to the younger audience. The difference is they're just consuming it very differently than you and me. And I use this example of we used to just read the news, then came radio, we heard the news, then came TV, we watched the news, then came AI, we started talking and asking to the news.

And in all of these changes, the consumption of news actually increased, the value of news in our society actually increased. We are just consuming it very differently at different times of the day.

Yeah, so that consumption is a big deal. I want to know only the news that I care about. I don't want to hear stuff I don't care about, but I want to be aware that the stuff I don't care about is happening. I don't want to be in a bubble. So other industries have really struggled with this. If you think about the streaming industry and retail and and music, what is personalized news going to be for the Washington Post?

That's a question I've grappled with for the last two and a half years. I'm not from the news industry, I come from outside. So when I landed here, I realized there are two things news does that is very important. One is it tells us what is important in the world, and then it tells us why it is important. That's the sense making, right? And the personalized aspect is taken over by social media. They already tell you what's important, so by the time they come to us, you know, there are very few things we are telling them different than they already know.

But the why that is the core value that we provide, and that's where I think we have to have a balance of personalized. You need to be data driven, but you need to use your data almost like a compass, not a GPS. It is still the onus of the newsroom, a responsible, ethical newsroom with journalistic standards, to make sure the news we give out to the people is not so personalized that it becomes an echo chamber and a reinforcement of their beliefs.

And it's a hard thing to balance, cause we understand looking at big tech outside, if you go deeply personalized, you will have audience, you will have clicks, you will have money, you will have revenue. And for our industry to balance both of these, you know, meet the consumer where they are, give them the news they actually need. Don't give them too much when they're not ready for it, but at the same time, make sure we are being very even and our perspective and our opinion is coming through is very important.

I think what you're describing is a really difficult Goldilocks problem, which is you want to do enough, but not too much. You know, it's the not too hot, not too soft, just right. We want to know about the whole wide world that's going on. But we also care about opinions that are closer to what my prior opinion was. And often, you know, I try to be pretty active about keeping new sources in my life that I dislike intensely.

How do you maintain a journalistic integrity in that process then, when you're choosing a lot about the kinds of things that you focus on and don't? And this, you know, been going on for years, so this is not a new problem.

So I think it's a multifaceted problem, right? Like first, it actually starts with the newsroom. I do believe our newsroom, you know, with its standards and the way they do reporting, they're trying to put a very fair perspective out. What you will see if you come to our application is there are actually many different ways to consume. You can read it, you can listen to it. We just started AI podcast, where the AI chooses some articles that you might be interested in and turns it into a podcast.

You have the option of going to the home page, which is edited by our editors, right? So this is the expert perspective on what the world is happening. You can go to the for you tab and just read personalized news. So from our side, what we ensure is we give you many options and we educate you with good product and design why these options exist. So hopefully, somewhere between that, you get out of your echo chamber.

Now we want to go beyond that too, right? If you go to our homepage, you will almost see like a old style ticker at the bottom of our Washington Post dot com. Where we are letting other news organizations what they're putting on their homepage almost run for free on our site to say, hey, these are other things that are happening. Because it's quite possible we are not going to cover everything in every perspective. And to keep extending the service to the nation, right?

Like, I really think we need to, as a news company, try and give value to everyone's life as much as possible. We recently started something called Ripple, so it's Washington Post dot com slash Ripple, where we are going to opinion sections across America and trying to bring their content and partnerships with them to our consumers, to our users. And it's just a hard problem. But you do need people who are solving it, and you also need people on the other side who want it to be solved, people like you.

That's a really fascinating idea, the idea of trying to surface those ripples from lots of different places. And you know, I mean, let's let's be frank, it's you're not going to be perfect at doing that. But I think that's inevitably part of the process. That the cost of not doing it is probably more extreme than the cost of making some algorithmic problems in there. Man, I know you've had trouble with that with the podcast in terms of personalization and trying to get that extreme personalization.

Can you share with us a bit about how that project has gone?

Yeah, we realized there is a market need in the middle of heavy curated editorial podcast. I almost view them as expert opinions, right? These are the experts of our company who are saying these are the important things you need to know versus sometimes things that are not important to the world, but they're important to me. And I'll give you one very good example, which, like, really made me a fan of this product. You remember, like, when the Texas redistricting fight was happening and there was a lot of court cases going on, and at the same time in India, there were elections happening in the state of Bihar.

So we covered these two stories, and somehow the podcast, given my interest, talked about the redistricting, the laws, and you know how the party in power over there is trying to hold on to the votes. And then it contrasted with the elections of Bihar, where some of this might have already happened in the past, and therefore the party that's winning is banking on the wins coming from those type of redistricting efforts. And it just like neurons fired in my brain, Sam, and I'm like, Whoa, this is so interesting.

I have seen this side in India and I see what's happening in Texas. I kind of don't like it, but thank you for, you know, showing me these two. Now, if you imagine an expert's view to ninety nine of population of America, that second story is not relevant, right? And even if they are interested in it, it's not really gonna fire the neurons in their brains the way this podcast did for me. And I think that is the gap we are trying to really hit with personalized podcasts, is because this is all based on our reporting, right?

Like, this is all factual stuff we did at Washington Post, we did it because we think this is important for the world to know. We work very closely with our newsroom, we tested it very well. And yes, it's not going to be perfect. It made a few mistakes, and once we launched it, and we made sure when we present it to our consumers, with our design, with the disclaimers, with the warning, they understand that this is a beta experimental product, right?

They understood that there would be mistakes that happen, and we were all, as a team, watching it very closely. In terms of technical, you know, one thing we realized was it has a lot of trouble when you have lot of third person references in an article. So let's just say it says, uh, Vinny said this, and Jennifer said that, and the following sentences are he and then she. To us, it's immediately clear who the he is and who the she is.

To AI, it might not be. And once we started figuring out those type of problems, we really went back, changed our scripts, changed our prompts, made sure that we didn't change the writing of the article. It is exactly what the newsroom wrote, so none of that is edited. We just made sure on the AI side, we have solving this problem. And the proof of that is we have published about over 100,000 personalized podcasts by now.

The completion rate of these podcasts, it's actually higher than the completion rate of normal podcasts that we publish.

That's a beautiful example because it's going to connect some things, it's going to miss some things, but maybe when it does, it's going to be amazing. And one of the kind of enduring themes of our show seems to be this exact idea of improvement. One of our early podcast guests mentioned the idea that the first day is the worst day. So when you put this experiment out, you're going to discover some stuff like the pronoun problem you mentioned and how it's obvious to us which story connects to which one.

But you're going to fix those, and it'll keep improving. What's your plan for this product, for this personalized podcast? 100,000 episodes? I'm already quite jealous. I think we're just over 100. And it's been exhausting.

Well, I don't think it replaces the experts. You know, one hundred is a lot of work, 100,000 is still a lot of work on the team who's building it, because we review problems that they come in. So the work happens, I guess, on the different side, you know, for us, it happens on the QA side. But I would just zoom out of personalized podcast and maybe talk more about the AI efforts we are doing over here, and then it would all make sense, right?

The way we are viewing AI in our company is we call it AI everywhere, right? It's an AI everywhere approach where we want it in the production of the news. There's so much gen AI can do. We have a tool called Haystacker, which can go through hours and hours of videos. You know, what would take people weeks? And now our journalist can go and say, I want to find that person with red cap, you know, and go through Gen six right videos and get that type of information.

You have probably heard all about how big data sets are now no longer a thing journalists fear anymore. You know, they don't have to manually read it, they can really ask it intelligent questions. So we are building a lot of tools internally for that side. So that's one big pillar is use AI to help the core mission we have of journalism. The second thing where we go is consumer facing, that's where AI podcast, ask the post, AI summaries, all of these products come in.

Where we have taken an approach that the audience in the case of the AI revolution, I feel like the audience moved before we moved, right? When there was an internet revolution, people had to go buy computers, they had to learn it, they had to get on the web browsers, and then the newsrooms moved to a website. In the world of AI, the audience went overnight.

I want to push back, you know, a little bit on this haystacker. I really like that name. What you're saying is, hey, you want to go through that haystack and do it with artificial intelligence and find all those needles. And it's certainly true that we've got a lot more content in the world to go through. You know, that it's staggering, the amount of things that are happening. We're getting a lot more content. Are there more needles in that content, or is there better discovery of the existing needles, or is there a lot of the hay that you're sifting through?

Just a lot of left tail junk. Does that make sense? When I think about a haystack, I think about, okay, let's grow the whole pile, and when we grow the whole pile, we'll have more needles because we've got more hay, but we may just be hiding those other needles better.

Yeah, both things are right. So let me start with just because I use the haystacker project. The name came. We are finding a needle in a haystack because we actually already had a haystack. Somebody gave a reporter a lot of videos, somebody gave a reporter a lot of data and said, Hey, something's going on over here. And it would take them two, three weeks to go through it. So we just help them, right? Like we are helping them find that needle instead of them watching it frame by frame.

So that's really the origination of this tool, and this is one of the many tools that a lot of news companies are building these tools. But going back to your bigger question of like, well, there is just a whole lot more data, most of it is not interesting, right? We don't think it is the job of AI to find all those interesting things and serve them to you without a journalist involved in the middle, right? So the journalist is usually with their instinct asking questions, trying to find more out of it.

And I'm sure you can get to a world where you have really curated data sources. You can take like department of labor reports out, right? And our journalists use those reports and they create stories out of it. So when you go to ask the post and you say, Hey, what was the unemployment rate in 2013 in agriculture sector, right? We may or may not have written about it in a news article, but if it this is one of those data sources that our journalists trust and use.

I I think it's fair to use it and give the answer to the question, right? But once again, there is a newsroom in the loop, you know, like that verification of data. And I think that makes us for a little bit higher quality than the general purpose internet, you know, hoovering ask engines. And they have their own place. I'm not taking a dig at them, I'm just saying there's a different place for that. And what we are trying to build over here in Washington Post is if you are in the market for trusted news and journalism, and you want some verified facts and have confidence, you should start with us.

Right? And I think let's tie back to how you started this process. You started talking about why, and right now that why has to be part of that. Otherwise, like you say, that's a sharp contrast between the useful search engines, which produce a list but do not produce the why. As I say that, though, I think about modern search technology, and it seems to be trying to use artificial intelligence to move towards more of a why and more explanation.

But you were pretty clear about the role of your journalist in this process. So maybe expand a little bit on that. Where are you automating? What absolutely requires human judgment? How are you figuring out where those lines are about what can you know? And we could talk about individual examples, but what's the process for figuring out how to decide?

So it goes back to AI governance and policies around how we are using AI in the company, and we broke it down into three parts, right? The easiest one I'll talk first is infosec. We got our infosec team involved and we said, like, listen, you need to tell us how to not mess it up really bad. You need to tell us what's happening on the bubble in terms of security and put a policy out, which is easier for us, cause we are using a LLM that we are hosting on a private instance.

Then comes the newsroom aspect, right? Like the newsroom and the journalist sat down and they've decided for themselves how they want AI to show up in the work they do, how they will use it, uh, how will they attribute to it? What are the do's and don'ts, right? And then the third aspect is the consumer, right? This is the tricky aspect, cause this is what you typically think of as a product. And the approach over here we have taken is using good design.

We want to always inform our consumers, our audience, what they are consuming, how much of this is from AI, and it's a spectrum, right? Let's take the example of summaries. We still label AI summaries that this is an AI summary. But the way I see people use it and the number of people who are actually looking at the disclaimer or giving us a thumbs down button on it because they didn't like it, it's moving down, right?

Like it's almost to the point that nobody is shocked that we have an AI summary, and none of the users are bothered about it. But I'm pretty sure if we put a full AI generated video, which we haven't done so far and we don't plan on, we would put stronger disclaimers. So at a product level, we want to lean on design and consumer behavior to make sure we are always informing them when they are using something is AI or not.

Let's jump forward though, you know, if we were sitting here together in a decade, you've got to be thinking about the direction that the news experience is going. And you've mentioned the read the news, listen to the news, watch the news progression that's happened. You've thought about this a lot. So tell me what you think is going to happen in the next decade or so.

Oh, if I was that smart, Sam,

You wouldn't be talking to me.

I would be somewhere in New York in the hedge fund business, making my bets.

Okay, let's do this back next month, then we can go shorter next month. Maybe you can give us a little hint about next month, and we can try to expand from there.

Yeah, I do sincerely believe the need for news and quality news has never been more. Journalism is a discipline, not just a format. We need to keep adapting our journalism to different formats, use technology where it can help us, and that's what we intend to keep doing at Washington Post. You will start to also probably hear, you know, the ideas around, like liquid content, right? Like, so think about the content the way we do. Typically, news lasts twenty four hours, right?

After twenty four hours, every newsroom will tell you the story dropped off. They take it off the homepage, people stop talking it. You do a deep investigative piece, maybe you go seven days, right? Like, we will pin it somewhere, people will share it, it will have longer legs, but no matter what, after that, it just drops off. I see a world going where people's curiosity drives the news. News can literally live in an infinite forms for a long period of time, because somebody could come back and start asking bunch of questions.

They could start asking questions, or they could say, Can you help me write up a report in the change in ice task tactics between DC and Minnesota? I really want to understand what was happening in the world at that time, that it became more violent than it used to be in the past. So I do think this unlocks more news. It actually grows the market more than initially the fear of shrinking, and that's always the fear, right?

Like when a new technology comes, first, there's a very genuine fear of shrinking. Like, I don't want to deny that, right? Like, honestly, as an engineer, you know, I see what cloud code has done in the last two months, and I'm like, Whoa, there goes my backup career choice. I guess I'm gonna not be a super short Java programmer anymore. But once you get past the fear, I think this grows. AI helps us grow as long as people and their curiosity and the need to get verified news, information, facts exist.

This is going to be good. So there's the bull bear, what do you call in stock market, the positive

There's bull and bear. You need to do that if you're going to switch to hedge fund. yeah, as Bull is positive, bear is negative.

as you're realizing, my future career choices are quite limited.

You better stick with Java.

I'll stick with Java. But I also do see, you know, there is risk around trust, right? And when I look at the future, the thing that worries me the most is the trust of consumers used to be with the mastheads, right? You would read a newspaper because you trusted that there were standards and procedures and professionals, right? And then in our lifetime, I see the trust move to creators. People started trusting creators more. They were more influenced by people on Twitter, they were more influenced by Instagram and TikTok people who were telling them the news.

And I thought about it and I'm like, you know, what's going on over here? One is like, our news did not adapt fast enough, right? That's true. We did not meet the consumer where they are, but we as humans just generally trust other humans. We trust voice, we trust language. You know, no matter what part of the world you are, if somebody speaks any other language, you know that you're in company of intelligence. In fact, if I could go back to my Apple days, we had this anecdote.

Like, when Siri came, it was the first voice, right? Like it was the first voice interaction with your machines, people could talk to it. And then Apple Maps came at the same time, and we had few incidents where we had wrong data and people would go on dirt roads and get stuck. The consistent complaint Sam we used to get is well, Siri told me to go there, and that's when we realized the Siri voice and Apple voice being the same voice was actually a problem because they were putting more trust in it than they should.

Their eyes were showing this road doesn't exist, but they would turn right because Siri told them so. So I think this is what happened to us is the trust move from mastheads to people, because naturally, as humans, we trust other humans a little bit more. What worries me is as these AIs become almost a better human than a creator, because they can talk back to you, they can be deeply personalized, they can understand you more than a creator does.

I fear the trust will move to the AIs even more than it was with the humans. Now, given that, what do we do, right? That's my hypothesis that trust to AI that people will have, the relationship we have will be very deep. I think the onus is on us in the news, in the journalism world, to build that equal type of experiences. So the consumer doesn't just get locked in with couple of big options that exist in the world outside, right?

And I feel hopeful when I see things like MCP protocol come out,

Model context protocol.

I see model context protocol, I see agent to agent conversations happening. I see enough companies out there, big tech, small tech startups, who are working down this path of saying, Hey, if my agent needs news, I want to connect it with your agent so it can get, like, the right verified news. So I'm hopeful also, but I'm also very worried about the trust. I want to just make sure it stays with people who deserve it.

Actually, there's four or five things that are pretty fascinating there. One, I had not really thought about that transfer of trust between the different Siri products and how that my gut reaction, my naive approach would have been to say, hey, that's good, that that trust transfers. But what you pointed out is that when you have two different products with different base levels of accuracy, that you might not want to transfer that trust. That's an interesting way of thinking about that that hadn't.

I naturally thought, hey, more trust is better, but you can actually signal this is something that should not be trusted with a more robotic voice, for example. You touched on Siri. Let's back up here and talk about how you have not always been at the Washington Post. Tell us a little bit about how you got to where you are there, and Siri as a part of that journey.

Well, I mean, back in my undergrad days, I got introduced to AI, and I kinda got seduced by the idea of machines doing all the work for me. So I was like, this is great. I'm gonna go get a master's in artificial intelligence, you know, so I can just sit back and relax. And that led to my first job in the mortgage industry. We used to do these AI models for loans. And if you remember, the year being 2007, when the great mortgage crisis and the financial collapse happened, my entire industry got wiped out.

Turns out, nobody was listening to AI when it came to loans, but that one door closed and a universe opened. I was contributing some open source code. The founders of Siri saw my code, they invited me to apply for an interview. So I went over to Silicon Valley, and then I spent like the next ten years working with them, building Siri. And we were the first voice driven AI for our time, and for the longest time, till Alexa came and Google Assistant came, and that whole universe opened up.

Spending about ten odd years, I took like a hard right turn and I went into Uber Maps. I ran the team that was building the routing algorithms, and it it was a whole lot of fun. It was graph search, it was hardcore computer science, right? Like this is as computer science as you get is graph search. So I really loved that stint. Uh, and after doing that for about four years, LLMs came on the scene, right?

And then I just was like, Okay, I'm going back to my old world of natural language processing, and I wanted to do something over there. So I took some time off from Uber. I thought I'm gonna reeducate myself. I bought some gardening tools. My wife got really worried. She's like, how long are you going to reeducate yourself? You have too many tools over here. But this Washington Post opportunity came and all the neurons in my brain fired.

I said, Listen, this revolution is all about language, it's all about knowledge. This is what newsrooms are. They are the repository of language. They are the masters, they are the experts, they have all the knowledge and information. And then I interviewed with Washington Post, they had a great team. I interviewed with Jeff Bezos, and, you know, that's finally. I was like, Yes, this is what I want to do as my next chapter in life.

There's a whole bunch of things to push on there. One part of that I want to pull on you glossed over very quickly was that you had made some open source contributions and people at Siri noticed it. And that led to being involved with Siri, which led to the Apple acquisition and your involvement there. I particularly like that because I'm a very big proponent in this idea of contributing things. And we think about the incentive for contribution.

That's a great story for how being interested, being curious about technology and working on something and providing evidence of that through an open source project or other. There are other ways besides open source projects, but that's one great way can cascade into a very interesting arc around how that developed.

No, that's true. And you know, I got lucky in a lot of ways because I was doing something that people were interested in and that opened up this opportunity. And you're very right, right? Like, I I do think like when you're early on in your career, you should dabble with things a whole lot more than become an expert in, because you don't know who is looking.

You say luck, though, and I do think that there's a big part of that luck, but luck is only combines well with working on something at the same time. I'll also make the snide comment that one part of the story I'd like to gloss over is your master's in artificial intelligence was from the University of Georgia. And I'm a Georgia Tech person, so I want to quickly gloss over that that you can have bad luck as well.

No, I actually do think it's an important one, and I have deep, deep respect for Georgia Tech. Of course, you have amazing computer science program, robotics program, AI program. What University of Georgia was offering uniquely at that time, and it still does, is it's an interdisciplinary program. So I studied language, I studied philosophy, I studied the theory of mind, I studied first order logic, and then I also studied all this statistical AI, which is basically 99.99 of the AI as people understand it now.

So congratulations, you guys won.

One other part of that was the you mentioned the graph based. Why do you think that the graph based approaches are so interesting? Why did that catch your eye?

Well, it was a classic routing problem, right? Like we are doing maps and routing, so you have to route over graphs and edges and nodes. Those algorithms, you studied them in school, right? That's what caught my interest. Now for Uber, there was a twist. The twist was this that routing for a transit is very different. When I say mass transit, I don't mean buses, I mean like taxis and Ubers is very different than personal routing. We settled on a metric which was ten meters or ten seconds.

If your map is wrong by ten meters, or your ETAs are wrong by ten seconds, you don't have a great experience. If your Uber stop ten meters further away than where you are, you are running to catch it. You're putting yourself in an unsafe situation, maybe you're crossing the street. If you didn't reach in time and your Uber is standing over there, maybe that guy's getting a ticket. The traffic's backed up, the cops are on their case.

So for us, the level of accuracy was actually way more than what Google and Apple do. And we had to scale not with linearly, right? Like with Apple and Google, the number of phones they sell is the number of map directions that will happen. While we are trying to balance a market. So for one rider, you would probably reach out to one hundred drivers to see when they can get to them, and similarly for one hundred drivers, you reach out to one hundred riders.

And it's possible that the driver that's closest to me is five minutes away, and the driver that's closest to you is one minute away. But I might switch the order of driver, so we both get a driver in two minutes, and then the market is balanced. Otherwise, I would have canceled it, cause mine was five minutes away, right? So the problems that once you start poking at them, you see this is a very different routing problem.

Of course, graph search and the routes and the Dijkstra is at the heart of it, but the layers we had to keep putting on it to get to, like a balanced marketplace, was just very exciting, right? Like no one had really done that before.

That seems fun, actually. You mentioned Dijkstra's algorithms and these things, it makes me happy to think that these core ideas still maintain. I mean, this matching problem you just described is a classic example of a generalized assignment problem. And so these are some root problems in operations research and in graph theory and mathematics that it's fun to see that not everything is statistically picking the next probable word. Glad to see some of these old school things come through and come back.

Vinny, this has been a fascinating look at where journalism and the technology behind it, I think, may be heading. And the future of news clearly seems more personalized and more AI powered in many ways, and more complicated in any ways. And I'm glad that you and others are working on it. Thanks so much for joining us today.

Thanks for having me, Sam.

Thanks for listening. On our next episode, I'll talk with Andrew Palmer, a journalist at The Economist, who'll learn how another news outlet is thinking about AI. Please join us.

Thanks for listening to me, myself and AI. Our show is able to continue in large part due to listener support. Your streams and downloads make a big difference. If you have a moment, please consider leaving us an Apple Podcast review or a rating on Spotify, and share our show with others you think might find it interesting and helpful.