Today’s episode is about thinking practically about the AI revolution. Yes, it may one day usher in some now unthinkable utopia or dystopia. But in the meantime, our imperfect world exists, and your imperfect job exists, and you face a forced choice: Should you use this technology? And if so, how do you make it work for you? Kevin Roose, a tech columnist for The New York Times and the host of the podcast Hard Fork, talks about how generative AI tools are already changing his job and others, including in medicine, consulting, and software development.
If you have questions, observations, or ideas for future episodes, email us at PlainEnglish@Spotify.com.
In the following excerpt, Kevin Roose explains to Derek all of the ways he’s currently using generative AI—and all of the tasks it’s particularly good and bad at performing.
Derek Thompson: ChatGPT is 1 year old this month. We have had what feels like nine different hype cycles. There’s been a flood of investment into anything that calls itself artificial intelligence. The White House wants to regulate it. Artists want to sue it. Marc Andreessen says it’s going to save the world. Effective altruists say it’s going to destroy the world. I have a much more prosaic place that I want to start with you. I want to know how to use this technology. I want to know how people are actually using this technology—not how they might use it five, 10 years from now, [but] how it’s actually being used right now. So let’s start with you, Kevin. Out of the suite of tools that calls itself generative artificial intelligence, what’s useful? How are you using it?
Kevin Roose: I use it for so many different things, and I guess the way that I’ll start is with work. So I am a journalist and a podcaster. I have experimented with using ChatGPT, Bing, Claude, etc. in many ways in my journalism. I’ll start with what I have found it is not useful for: writing my column. That is a thing that I am glad that it is not good at, but these tools are not good at producing high-quality outputs when it comes to things like newspaper columns. That is just something that maybe I have resisted turning over to it because I have some pride in authorship, but it’s also pretty generic. It sounds like a Wikipedia article sometimes. So unless you give it very detailed instructions, it’s probably not going to spit out a high-quality column about something new that is happening in the world of tech.
So that is what I’m not using it for, but what I have been using this stuff for is almost everything else. So I have a podcast. We interview guests. I start almost every brainstorming process about what to talk about with a guest by asking ChatGPT, “What are some questions that this person might be able to answer?” I have used it to remind myself of things. The other day, we did an episode about AI wearables, these gadgets that you can put on your lapel that’ll record everything and use AI to analyze it. And I was thinking to myself, “I remember reading some story, some science-fiction story, about one of these devices,” but I couldn’t remember what it was. So I just asked ChatGPT, “What is the science-fiction story about a wearable AI device that records everything?” And it came back and it told me, “Well, that was Ted Chiang’s The Truth of Fact, the Truth of Feeling.” And that was, in fact, the story I had been thinking of.
So I’ll use it for brainstorming. I’ll use it for research, for summarizing white papers. There’s a feature that’s been in Claude, which is the chatbot from Anthropic, for several months now where you can upload a PDF and say, “Summarize this for me.” So I use that a lot. I’ve also used it for things like, “Show me some empirical research about the effects of Airbnb and other short-term rental sites on urban housing prices.” That’s the kind of thing that these models are very good at. So I found it less useful as a way to write and more useful as a way to come up with the creative grist for writing.
Thompson: A lot of people’s interactions with the stories of generative AI circulate around this concept of hallucinations. The first thing they think of when they think of ChatGPT is, “Oh, it often lies.” So you’re a New York Times journalist. What are your fact-checking protocols when you’re using ChatGPT or Anthropic’s Claude?
Roose: Yeah. So I always, if I’m going to use something that one of these chatbots tells me in a podcast or an article, I always check it against reality. That’s definitely a shortcoming of these models. And I found that they’re best at tasks that don’t require a high degree of precision or factual accuracy. But I also think that this hallucination problem is not as severe as maybe some people make it out to be. For most tasks, in most disciplines, it actually is pretty good at giving you the right answer or something approximating the right answer.
Thompson: I think it’s really interesting that you point out that it’s not so much about truth as it [is about] idea generation, about stirring creativity, about getting you unstuck. And I also totally agree that while it’s not particularly good at replacing humans and offering the final product of something like a New York Times column or an Atlantic column, it’s very good at being a copilot for getting you up to that final stage of writing the final product, whether it’s summarizing a body of research, summarizing a research paper, or surfacing research papers about a certain topic. You mentioned that you use it both in work and also outside of work. How do you use generative AI, ChatGPT, etc. outside of work?
Roose: All kinds of things. I’ve started trying to experiment with it as almost like a conversational practice coach. So if I have to have a tough talk with a friend or I’m nervous about some speech I have to give or something, I will practice on the AI first and have it pretend or role-play as my friend or the audience of a talk and just give me feedback. I’ve used it to come up with fitness plans, for meal prepping. I’m also using it to teach myself things. One of my goals for this year is to learn how to code in Python. And so I just made a custom GPT the other day that is my Python teacher, and it’s giving me lessons, and I can actually learn how to code using this chatbot as my personalized tutor.
Thompson: How have those difficult conversations with friends gone after you began with ChatGPT to help coach, rehearse your way through these future conversations?
Roose: Really well. Part of it is that it’s just useful to ask questions like, “How would a person with this personality type respond to this kind of conversation?” and just have it practice that interaction. I don’t use this all the time, every day, for every social interaction. I don’t want to give that impression. But I have tried it out on that kind of thing. And I would say it’s been fairly helpful. There’s a lot of talk right now about AI as a therapist, and people are trying to use it for that purpose. And I would say right now, it is probably not as good as the best human therapists and certainly doesn’t know as much about you, doesn’t have the history that a really seasoned therapist would have. But for basic stuff, for like, “I’m feeling really stressed at work; help me understand why that might be,” that kind of thing I’ve found that AI can actually be pretty good at, especially as an addendum to traditional therapy or just talking with a friend or a loved one.
Host: Derek Thompson
Guest: Kevin Roose
Producer: Devon Manze