Margaret I broke up with Dustin three months ago. We were watching Breaking Bad, and my phone buzzes, right? So I look over, and it’s a text from Dustin–a long response consoling me about this thing yesterday. But when I look over, he’s glued to the TV! There’s just no way he could have written it.
So, he eventually admits that he’s using some kind of auto-responder service to reply to my texts. Apparently he uploaded all of his past conversations to this website, and it learned to text like him.
Dustin said he only used it occasionally, when I talked about something boring. He said he didn’t want to ignore me, but he didn’t want to spend the time thinking about it, so he just turned on this autoresponder, and it wrote these nice, gentle, empathetic texts for him.
Anyway, we had a big fight and I said that if he didn’t even want to text me, then maybe he shouldn’t date me, and maybe he should go home, and maybe he should never talk to me again, and that was the end of that.
Laura I started Scribe in 2025. Our goal was simple: we wanted to personalize language models. From a technical standpoint, language models are pretty simple: they try to predict the next word in a sentence. For example, given the input “I hate waking up ____,” it would predict that the word “early” is most likely to follow.
Enthusiasts had been using these language models to write their emails for a few years, but you could always tell when someone went from writing their own emails to using a language model. One day they’d be writing two line emails full of weird slang and zingers, and the next day their emails would be ten generated paragraphs of polite Wikipedia English.
Our original goal at Scribe was to prevent that–to make generated emails that sound like you really wrote them.
After learning from the whole internet, we trained your personal language model on your text and email history. Once it learns what kind of writer you are, it can generate emails and texts in your voice. We call this personalized generator your Scribe.
Your Scribe learns all kinds of things about you during the training process. To respond to jokes, it learns whether you're a "lol" person or a "haha" person. To complete the phrase “I hate waking up ____,” it learns whether you’re the kind of person more likely to say “early” or “late.” To respond to your coworker asking about that late project, it learns whether you’re the kind of person more likely to start your response with “actually” or “sorry.” To respond to your boyfriend’s texts about his long day, it learns whether you’re more likely to respond with “have you tried” or “that must be tough.”
So it turns out that in the process of learning to imitate your writing style, our models absorb deep aspects of your personality. Are you more argumentative or apologetic? How are your sleeping habits? What kind of emotional support do you generally provide?
Our Scribes use this knowledge to suggest realistic responses to our clients, which ultimately helps them spend less time in their email and more time in their lives.
Margaret So after I broke up with Dustin, I figured I’d give dating another try. And I kind of noticed that people on dating apps were much more talkative than last year. Eventually I realized I was mostly talking to bots.
I looked it up, and sure enough there's this service called ScribeAI. I just uploaded my entire chat history to it, a bit creepy but whatever? I don't have anything to hide, you know? And now my Scribe talks like me.
When I match with someone on a dating app, I just attach my ScribeAI account to the chat. My Scribe talks with the person I matched with, simulating the kinds of things I would have said. I just check in on the conversations every once in a while. If it looks interesting, I schedule a date. Otherwise, I just unmatch. I used to spend hours replying to messages, and now I barely have to write any messages at all.
I ended up dating this guy named Vincent. He’s really talkative, which is a good fit. Often I want to talk with someone but don’t have anything in particular to say; he’s in a fraternity and has lots to talk about, so our conversations flow easily. And he really knows how to listen when I want to blow off steam. I think it’s a great fit.
Vincent Yeah, I’ve been using ScribeAI since I got to college.
Man, it’s crazy good for dating apps. I uploaded my entire chat history to their website, and now they take care of all my conversations. Last time I was single, I went on 20 or 30 dates before I found someone.
Some of those dates were terrible–my date talked the whole time about her favorite ‘70s rock bands, or even worse, never talked at all. Now that I’m using my Scribe, it’s like I’m having hundreds of conversations at once. I only bother meeting the people that my Scribes have good conversations with. I think my Scribe had something like 500 conversations before I found Margaret.
And let me tell you, ScribeAI isn’t just for dating. I’m a brother at Pi Kappa Phi, and we’re using it all over the place for rush this year. Normally, during the first few weeks of school, I have to talk to hundreds of freshmen. Then we have all these meetings, where we have to guess if a frosh is going to be a good brother at PKP.
This year, we tried something new. We asked all our brothers and all the frosh to get ScribeAI accounts. Then we simulated a group chat with the current fraternity and each new member.
If the recruit’s Scribe had good conversations with the brothers’ Scribes, we considered admitting them. Otherwise we dropped them. This saved us so much time.
Obviously it’s great for admissions to be faster, but there’s another really important piece of rush. We have to be really careful not to mess up the fraternity vibe. Right now we have this great laid-back vibe, but if we admit too many hard asses we might lose that culture.
This year, we noticed that a simulated group chat with new recruits spent too much time worrying about their assignments. We dropped one particularly talkative academic recruit, and when we re-ran the simulation, the vibe was back how we like it.
Another classic rush issue is trying to screen for stuff like sexism. People are polite during interviews but then really problematic once we admit them.
One brother had the idea to simulate one-on-one conversations with each new recruit, where the brother says something overtly sexist. Then, if the recruit goes along with it, we drop them. We caught a couple of recruits that way this year.
Even aside from the frat and dating stuff, I’m starting to use ScribeAI to find friends. I like history, but not too many dudes in my frat dig it. So I signed up for this college-wide Scribe pool.
The admins run hundreds of simulated conversations between you and all the other people in the pool. Then they send back a list of the people that you had the longest simulated history conversations with.
They recommended me a girl in the math department. She thinks a lot about history, and we have these crazy conversations about the Bronze Age collapse.
Laura We just hired Seth as a machine learning engineer at ScribeAI. He’s coming from a job in accounting, but he studied computer science in college and we’re very confident that he’s a good fit.
Our hiring process here is a bit unusual. Instead of interviewing, we subject our applicants’ scribes to simulated slack conversations. We hired Seth because his Scribe did particularly well in two tests: attention to detail and loyalty.
We test attention to detail by proposing a solution to a complicated technical question. If the candidate’s scribe responds by saying “actually, have you considered …?,” we consider it a success.
Although the candidate might not know anything about our technology, their predicted response shows that they’re the kind of person who often points out issues that others have overlooked. This focus and willingness to speak up is very important in our line of business.
Since we’ll spend a lot of time and energy training Seth as a machine learning engineer, we also care about his loyalty. In our loyalty test, we pose as a competitor’s recruiter and offer Seth a 50% raise to leave ScribeAI.
Seth’s Scribe said that our company had been generous to teach him so much, and that he wanted to pay it back before working at a competitor. That’s exactly the kind of response we look for.
When it comes down to it, we’re not hiring Seth for any of his particular skills. He knows next to nothing about the technology we work with. We’re hiring him because, at a fundamental level, he’s the kind of person we want to work with at Scribe.
Margaret Vincent’s really into this Scribe stuff. He’s used it for his frat, and his dating life, and even finding friends now? And last week we even tried this new long-term dating predictor.
It simulates thousands and thousands of conversations between two people and reports some statistics. We decided to try it, and it turned out that we weren’t that compatible. On average, in at least 77% of conversations, the texting thread eventually dies with us choosing to break up.
There’s a pretty consistent pattern, actually: it seems like when my Scribe feels neglected, Vincent’s Scribe tends to become annoyed that I’m not more ambitious. Then he spends more time with his work friends, and I feel even more neglected, I guess.
I don’t know, it’s like a sort of negative spiral embedded in our personalities? But now that we can see this pattern so clearly over so many conversations, I really want to keep dating and try to avoid it.
Some of Vincent’s friends say this is pointless, that we’re obviously going to break up and we should just get it over with. But I want to decide how to live my own life. That’s the difference between really thinking and just predicting likely words, you know?
Laura As a company, ScribeAI has a legal obligation to allow our clients to access their personal data.
We don’t keep our clients’ actual texting history around after we train their scribes. Instead, we compress all that information into a single 768-long vector. The 768 numbers in that vector encapsulate everything our model learned about them from their texts. In some sense, those numbers are all we need to reconstruct someone’s personality.
Everyone used to be concerned about privacy. But after the Supreme Court ruled on Garza v. Oregon, the courts have consistently treated your vectors as legally equivalent to your entire chat history. The government can still get them with a warrant, but they need probable cause. In practice, it’s more useful for the government to get your actual chat history instead. We get very few subpoenas these days.
Vincent That long-term dating simulator got me thinking about whether we can change our vectors. Obviously kids’ vectors change a lot as they grow up. But the vectors of teenagers and twenty-somethings? Do they change?
I want to go into politics, and I’ve been thinking, what if I could become a bit more similar to Obama? Could I become more like him by pulling my vector a bit closer to his?
I found a startup that helps you become more like someone you admire. Their main service is matching people. They know that when two people talk, their vectors gradually drift towards each other.
To take advantage of this, they match people whose vectors are different from Obama in opposite ways. Say my vector points to the right of Obama’s, then they’ll match me with someone whose vector points to the left. Then, when we talk, both of our vectors gradually move towards Obama’s, and we’ll both discover a new part of our personality.
My first match is great. He’s a bit less socially confident than I am, but he’s more down to earth and he speaks extremely clearly. We try to talk every day, and I can feel myself becoming more like him–and more like Obama–every time.
Laura We’ve had a few challenges as we’ve rolled this technology out. Early on, we had mobs trying to cancel people for things they were merely likely to say.
Some nights I wondered whether we were the bad guys, pulling society towards some dystopia where the government knows everything about you. Over time, I think the data has come to show our work in a positive light.
Severe loneliness, which quietly rose all throughout the early 2000s, is at all-time lows and still dropping quickly. Likewise, global happiness is up. And those gains have been remarkably evenly distributed between rich and poor countries; everyone benefits from having closer friends.
When I really think about it, I find giving people an honest mirror into who they are, and some people don’t like it. But I go to work increasingly confident that we’re pursuing the most impactful social project of the century.
This was the winning entry to MIT's Envisioning The Future of Computing Prize. Thanks to Cory, Caspar, and the rest of SERC for their generosity and attention in organizing the competition.
If performed, Laura is meant to come across as somewhat condescending and aloof. Margaret is meant to come across as authentic. Vincent is meant to be assertive and mildly overconfident.
The piece draws inspiration from both Terkel’s Working and Chiang’s Liking What You See, in the way it uses a series of interviews to tell a story of societal change. I read and liked both, neither recently.