Transcript: ‘She Turned Her Whole Life Into Training Data—For an AI Baby’

The transcript of AI & I with Sarah Rose Siskind is below. Watch on X or YouTube, or listen on Spotify or Apple Podcasts.

Timestamps

1. Introduction: 00:01:54
2. How Siskind is running an experiment between her unborn child and an LLM: 00:02:03
3. A demo of Siskind’s FetusGPT: 00:07:34
4. Siskind’s pick for the funniest LLM: 00:15:16
5. How Siskind uses AI in her comedy writing: 00:17:12
6. Dan and Siskind use ChatGPT to write a joke together live on the show: 00:24:41
7. Why AI is useful even when you don’t use its output directly: 00:37:21
8. How Siskind used a ChatGPT project to biohack her energy levels: 00:44:15
9. A question we fundamentally couldn’t have asked in pre-ChatGPT times: 00:57:09
10. How ChatGPT is a source of emotional support for Siskind in pregnancy: 01:05:29

Transcript

(00:00:00)

Dan Shipper

So FetusGPT—remind me. It’s basically— You’re recording all of the things that your fetus has heard and then using that to train an AI, right?

Sarah Rose Siskind

An AI that has been trained on nothing other than the MP3 files of the past five months and text files of the past five months. So it’s a one-to-one experiment and it’s not a company—it’s just an insane performance art thing that I’m doing because I think it’s insane. It’s a GPT–2 unweighted-like model architecture. So it really is babbling—it says nonsense—because it’s training itself to learn English based off of only what I expose it to.

Dan Shipper

And so like what would be an example of what’s in the training data?

Sarah Rose Siskind

I was watching a YouTube video on lemur self-medication. And Dan happened to open up a random file.

Dan Shipper

And just for people listening, who’s Dan?

Sarah Rose Siskind

Oh, sorry. Dan Shiffman, who is different from Dan Shipper. You must be distantly related. You guys are bizarro world versions of each other and you have to collab. Dan Shiffman of the Coding Train is a coding legend and also my AI baby daddy because he created the model architecture for our joint FetusGPT product.

What was I saying about him? He opened up a transcript and he usually doesn’t read the data, the transcript of my life, but he was like, Sarah, are you looking at lemur self-medication? And I was like, it was a YouTube video I was watching.

Dan Shipper

What is Lemur soft medication? How are they self-medicating?

Sarah Rose Siskind

Well, it’s brilliant. They use ants as insecticide. They’ll like use a certain type of ant and eat it and rub it on themselves and the saliva of it to work as an insecticide. And very few animals use tools other than us—chimps and stuff. So it was like a big deal.

Dan Shipper

That is really interesting. Okay, so that’s in the training data. What else?

Sarah Rose Siskind

Seinfeld. A lot of like troubleshooting microphone issues—you know, 5 percent of all of our days snoring, eight hours out of every day. It’s being trained on snoring. This is my Plaud Notepin

That is actually the doing all the recording. And so it’s a lot of snoring. It’s a lot of silence. It’s a lot of me checking in with my partner about chores. Fighting. Seinfeld. YouTube videos about animals. And then work calls—nonstop work calls.

Dan Shipper

And what have the results been so far?

Sarah Rose Siskind

Well, one time FetusGPT referenced pedophilia, and now I stopped talking about pedophilia so much because, so I’ve stopped swearing so much.

Dan Shipper

How much were you talking about it before?

Sarah Rose Siskind

Well, Epstein’s been in the news and I don’t talk that much, but when a toddler learns to swear and you have nobody to blame but yourself. That’s the big thing I want people to come away from this experiment with is that AI is what we make it just like children. Just like the way a child, if it swears, it’s kind of the parent’s fault because they swore in front of the kid, so now it knows that word. Now FetusGPT references pedophilia, and I try not to swear in front of it. I’ve actually stopped swearing as much because of this fricking thing.

Dan Shipper

That’s really interesting. And do you have a demo? Can we see a demo?

Sarah Rose Siskind

Yeah. should I share my screen or paste in the chat?

Dan Shipper

Share your screen for sure.

Sarah Rose Siskind

Okay. And I just want to warn you. You’ve got to lower your expectations because it’s really bad. It’s only a couple months old. So it’s really good for that amount because my fetus can’t say anything.

Dan Shipper

Wait, actually yeah. Before, we even go there and we’ll say what’s on the screen in a second. Do you feel like this is a truly fair experiment because your fetus is also experiencing—it’s getting fed with food. There’s probably touch. It has some sort of sensation. So there’s probably data that that FetusGPT has not gotten yet.

Sarah Rose Siskind

Of course. Absolutely. And the experiment’s going to basically go to shit as soon as the, as soon as it gets exposed to visual data because like so much of our brain is trained for visual data, but for right now it’s really close one-to-one. So yeah, FetusGPT does not eat, does not smell it. Can’t see right now.

Dan Shipper

Not yet. It does not eat yet.

Sarah Rose Siskind

Well, maybe we’ll see how things go. But apparently the human fetus can see light. It’s actually kind of this red glow because it’s looking at light through my human tissue. So even then it has some visual data and it can open its eyes fuse. GPT doesn’t have that, doesn’t have taste, doesn’t have smell and it’s weights. You could make a metaphor between like the emotions that the fetus feels vs. weighting those certain weights in the model architecture of FetusGPT. So there’s kind of a similar one-to-one there. But the experiment’s going to get more complicated as soon as the fetus becomes earth side. What do you think?

Dan Shipper

So for people who are listening, we’re now looking at FetusGPT, which is a chat. It just looks like a regular chat window, more or less with with infant emojis. At the top it says FetusGPT. It’s a little pinkish. And Sarah is typing in, “what do you think of Dan Shipper?” Which, I mean, I guess I’m probably in the training data somewhere.

Sarah Rose Siskind

Yeah, because we talked at the Mets game—the OpenAI Mets game—so you, my fetus, both of them, my FetusGPT and fetus have heard your voice.

Dan Shipper

So what is the plural of fetus?

Sarah Rose Siskind

Feti? I obviously don’t know. Okay, let’s see what it says. Oh, she—it’s already misgendering. Oh, she’s a huge—. But it’ll be a really—like a certain sexualized—God damnit!

Dan Shipper

Wow. Interesting.

Sarah Rose Siskind

Now you get a sense of what I talk about in front of my fetus. That’s the common stuff that, like he was talking about New York. This has no sense of grammar. It has not been trained on enough data. So this has been trained off of like 1 million, 2 million words roughly. And that’s about I think it’s like 15 megabytes of text files compared to the petabytes that GPT–5 has been trained on. So it’s doing pretty well considering how little data it has, just like a fetus does. But one of the things that’s great about this experiment is like how it shows actually how efficient humans are with like very little data.

Dan Shipper

That is really interesting. What have you learned so far? What have you taken away from this experiment as you’ve done it?

Sarah Rose Siskind

So, I’ve taken away like a lot of things equally about AI and about my own fear of becoming a parent. I have learned that we are really efficient with data. Like, I said, like we’re incredible at extrapolating from a little amount of data a lot of knowledge, as opposed to AI, which requires just tons and tons of knowledge, and that’s a very like bottom up architecture. So we come in with a lot of priors and then I’ve also learned we’re way more energy efficient than AI. The human brain essentially uses up the electricity for a dim light bulb all the time, as opposed to training a GPT–5, which is like, just so many gigawatts of energy. So, we’re pretty cool—human beings. Big fan. Hope we stick around. And then I’ve also learned a lot of stuff about like being a parent. So like one of the things was talking with Dan Shiffman about his two children and he said that like, this experiment made him realize how he misses their hallucinations. So I miss when ChatGPT was bad or when—

Dan Shipper

Got to take him off the antipsychotics.

Sarah Rose Siskind

And then they go right back. Easy fix. Easy fix. Boom. He was like, his daughter used to say “libble it” instead of “little bit.” And now she’s like a teenager and she has no speech impediment. But he like misses those things when she thought believed in the tooth fairy. And mispronounced words. And I miss when DALL-E couldn’t do fingers. That was a fun time and now it’s gone because it’s good at this stuff. And the hallucinations are great because they’re like little kid creativity.

Dan Shipper

Yeah. I think there’s definitely so much overlap between child development and AI training and all that kind of stuff. My nephew, he’s three and he says direction instead of construction. So he loves, like whenever he’s walking around in New York, he’s like, “direction.” And it’s stuff like that where you’re just like—

(00:10:00)

Sarah Rose Siskind

I love that. My nephew, he’s like, “Sarah, I’m a boy, so I do karate and I do makeup.” And he’s just—those are the things he’s been exposed to and he knows he’s a boy. And so for him, he just extrapolates that all boys do karate and makeup and those are the two things that define the category. And he’s so confident. He’s hallucinating very confidently. And I love that.

There was a great episode of “This American Life” where this woman was talking about how she caught her dad putting money under her pillow with the tooth, and she was like, instead of saying, “Oh my gosh, the tooth fairy isn’t real,” she was like, “Oh my god, my dad is the tooth fairy.” Which, when you think about it, is a much, much more logical conclusion than everyone’s lying to you about this insane, fucked up body swap. That’s good, magician energy. You know. And sometimes children are more logical than the real world. So I just love that.

Because I think that reminds me in some ways of how AI… One time somebody asked Gemini, “How do I get the cheese to stick to my pizza?” And early Gemini was like, “Oh yeah, just add an eighth of a cup of non-toxic glue.” And it’s like, yeah, that’s an inaccurate answer to that question.

Dan Shipper

One method.

Sarah Rose Siskind

You’re correct. That will help it stick.

Dan Shipper

Yeah. That’s one of the things that earlier—and there probably still are people like this—but earlier in the LLM era, there were people that were just, “Oh, that’s so stupid. Look, I asked it what makes cheese stick on pizza and it said Elmer’s glue, or it can’t do this basic math problem or whatever. So therefore it’s just not smart and it’s never going to be smart. We need something else to actually make it good.” And I’m just, have you met a child? Have you ever spent any time around a toddler?

Sarah Rose Siskind

This makes me so mad. You know when people will be—because I’m a professional comedian—and people will say, “Oh, I asked ChatGPT to tell me a joke. And it told me a terrible joke.” It’s usually a pun. And I’m like, ask the world’s greatest comedian, after you just met in a completely quick interaction, to tell you a joke and then tell me whether that joke is funny. Jokes don’t work like that. You’re not going to be instantly crying laughing with a human being. Why would you expect AI to do any better?

People are just out with an ax to grind about this stuff and then they have no patience. And so my number one piece of advice—because I also do some AI tool consulting for creatives—is: iterate. Don’t give up. Don’t just give it one turn and then be, “This sucks,” and move on.

Dan Shipper

This is actually a great segue. So do you use AI in your comedy writing? And if so, how?

Sarah Rose Siskind

Oh my gosh, absolutely. Nonstop.

Dan Shipper

So, what’s the funniest model?

Sarah Rose Siskind

Oh boy. The funniest model was probably DaVinci because that shit was off the rails. That shit was crazy. And then they had to put a lid on it. This was—God, what do you remember? DaVinci was four years, three years ago, three or four years ago or something. So they had to put it back in its cage. But that was rad as hell.

Dan Shipper

Do you remember any specific DaVinci things that it said that you heard? I was not given access.

Sarah Rose Siskind

I don’t think it was public. I just heard about it from Simon Rich, who’s one of my absolute favorite comedy writers, who was just like, this beautiful off-the-rails model that had totally no safety parameters and was fantastic. So there’s a great podcast he did about it.

But in terms of modern stuff, it’s a great assistant. It’s a great assistant. It’s not funny itself. It’s okay. AI used to be funny because we would laugh at it. As a matter of fact, I brought some show and tell—here’s a book that could not be written anymore: I Forced a Bot to Write This Book by Keaton Patty.

There was a whole genre of humor about AI because it was not good at fingers or Nothing, Forever was a Seinfeld episode created by AI. And it was funny because it was bad, because it was a hodgepodge cliché and it wasn’t very good. But that just doesn’t work anymore. And the SNL sketches about AI are very dated right now because they all rest on the premise that AI is not good. But it is good. It’s not as good as a top-tier professional comedian, but it’s an incredibly good assistant.

So, in a long-winded way of answering your question—how do I use it? There are two parts of the process: there’s divergent thinking and convergent thinking. There are these awesome researchers at Wharton who have written a lot about AI and creativity, Gideon Nave and—you’re a Penn guy, aren’t you?

Dan Shipper

I am, yeah.

Sarah Rose Siskind

Indeed. Well, you probably wouldn’t know them.

Dan Shipper

I don’t associate with Wharton people.

Sarah Rose Siskind

Oh, they’re beneath you.

Dan Shipper

I’m just kidding, Gideon.

Sarah Rose Siskind

I actually went to your alma mater for the first time last week to give a talk about AI and creativity, and I got to write the word “dicks” on the whiteboard. And I was told that was a first for Wharton.

Dan Shipper

I never did that in four years, so, yeah. Wow. You’re—wow. So, boom.

Sarah Rose Siskind

I went out of my way to write “dicks” on the whiteboard, just to have that in a bio somewhere.

Dan Shipper

I mean, I think you own the University of Pennsylvania now. I think that’s how it works.

Sarah Rose Siskind

In the bylaws.

Dan Shipper

Benjamin Franklin wrote it in the bylaws.

Sarah Rose Siskind

He would. He had a good sense of humor. Okay, so these guys—Stefano Tonini and Ramon Salazar.

Anyway, he wrote a paper about divergent and convergent thinking. And what the difference is, is divergent thinking is the brainstorm phase, where you’re trying to think in a diverse way. You’re creatively brainstorming, you want a lot of diversity, you want a lot of off-the-wall thinking. This is the point in the writer’s— I run a lot of writers’ rooms. This is the point in the writers’ room where you give everybody Play-Doh and pizza and you’re just, be a child. Think weird.

And then you select an idea. Let’s say we’re going to do a “Shark Tank” parody. And you start convergent thinking. So convergent thinking is like, okay, we’ve got this idea. Let’s do some research. What are some tropes about “Shark Tank?” What are some of the stereotypes? What’s some research on the idea? And then you start actually really narrowing your thought process and you’re doing a parody. So you’re trying to do a one-to-one, but you’re going to incorporate a little bit more of the creativity to say, how can we upend this parody so it’s not just a perfect replica of a “Shark Tank” episode?

AI can help me emotionally with the divergent process to make me feel safe and not alone if I don’t have other people or a writers’ room there. It really is useful to set an emotional safe place to be weird and a comedian and to feel disinhibited. And AI, I find, really, really helps with that process. And then you have to choose—discernment is the human thing, where the human needs to come in and be like, that’s the thing I want.

And then convergent thinking—convergence—is where it’s a research assistant, it’s a wordsmith, it’ll be a harsh editor if you need it to be. It’ll help you cut that parody down. It’ll help you format that parody. It’ll help you research what the tropes are, or whatever. And that’s basically the structure of how I use it in a creative process.

Dan Shipper

Can we write some jokes together?

Sarah Rose Siskind

Oh yeah, absolutely.

Dan Shipper

You know, I’m thinking SNL’s—all their jokes are outdated, so it might be fun to just do our own little writers’ room right now with you, me, and ChatGPT, or whatever your AI of choice is, and see if we can actually make some good jokes for SNL. I have never seen—or I would love to do—a parody of SNL, like them doing an extremely dated AI episode.

Sarah Rose Siskind

I have never seen—or I would love to do—a parody of SNL, like them doing an extremely dated AI episode.

Dan Shipper

Perfect. I love it.

Sarah Rose Siskind

This is so meta and inside baseball, but their stuff is always two years behind and I find it very frustrating.

Dan Shipper

Perfect. Let’s do it. How would we do this?

Sarah Rose Siskind

Okay, well, I guess I have to share my screen.

Dan Shipper

You’ve got to share your screen. And then, so when you’re—okay, we’re doing a parody of SNL. Where do you start here? Are we just sort of talking about what we think those people are, or even what is the output we’re shooting for? What would the output be that we’re shooting for, eventually?

(00:20:00)

Sarah Rose Siskind

What—yeah. Because that’s the biggest thing that AI is not going to be able to tell you, which is: what’s the right format that you want. And so you could be like, I want to do a standup set. I want to do a sketch. I want to do just one joke on this awesome podcast with Dan Shipper. Or I want to do a tweet.

It shines with the really formulaic stuff. So it is really good at writing tweets, I find. So it’s really good at—

Dan Shipper

Great. Let’s do a tweet.

Sarah Rose Siskind

So, let’s see. I mean, the premise that we came up with is not the kind of thing I think that would go great on Twitter. Which also, by the way, the rules have changed about what works on Twitter.

Dan Shipper

I think we could make it work on Twitter. It would be like the internal monologue of an SNL writer, and then it’s like dash, “an SNL writer,” probably.

Sarah Rose Siskind

Maybe. I’ve kind of reduced my usage—despite my reference to pedophilia and Epstein earlier. I’ve reduced my usage of Twitter lately. But there used to be formats I was much more familiar with, like “fired/wired,” that thing.

But here’s what we could do. We could even start with: what are some trending formats for jokes on Twitter? And this is something—

Dan Shipper

Were you part of the whole “may I meet you” craze?

Sarah Rose Siskind

No, I didn’t know about this. What is it?

Dan Shipper

Oh my god. You know Bill Ackman. He wrote this unhinged long-form tweet about how easy it is to meet women and how society is crumbling because men are not approaching women. And he said, “When I was your age, I would just go up to people and say, ‘May I meet you?’”

Sarah Rose Siskind

Oh my god, what?

Dan Shipper

And so that turned into a couple days of incredible “may I meet you” content.

Sarah Rose Siskind

Can I just say: the hottest pickup line I’ve ever heard.

Dan Shipper

Right. I mean, Bill started it.

Sarah Rose Siskind

Oh my god. Is it hot in here? Also, you know what? It fucking worked for him because his wife is hot as hell. I don’t know if you’ve given her a Google, but—

Dan Shipper

I have not. But there are some confounding variables there.

Sarah Rose Siskind

Dan, maybe you should try to ask women to meet them.

Dan Shipper

Obviously, I’ve been doing something wrong.

Sarah Rose Siskind

Yeah, maybe. I was wondering at what point this podcast is going to turn into me trying to set you up with some people I know.

Dan Shipper

I mean, I’m open. All right.

Sarah Rose Siskind

Maybe that’ll be the next stage of the—

Dan Shipper

I determined I’ve been saying, “Can I meet you?” and that’s the improper form.

Sarah Rose Siskind

Oh no. That’s really—ugh, God. That is so cringe. If a guy said, “Can I meet you?” I would absolutely not want to meet him. It’s got to be extremely formal, grammatically correct language. Otherwise I’m out.

Okay, so we’re looking at formats, and this is the kind of thing I might do to get inspired. Another thing I do to get inspired is I’ll just look up recent tweets. Oh, that’s another thing that’s great: looking up, with a web search, what have been “wrong answers only” is great for driving engagement, I’ve found, which is cool. Using current meme templates—so current, the Drake meme, absurd.

So yeah, you have to contend, as you guys know, with the updated-ness of ChatGPT unless you put on a web search feature or function.

Dan Shipper

One of the things I’m feeling right now is I just think, in order to make it work on Twitter, we’d have to be responding to something that happened on SNL in order for it to work.

Sarah Rose Siskind

Yes. Because it’s very reactionary.

Dan Shipper

Yeah, yeah, yeah, yeah.

Sarah Rose Siskind

It’s very reactionary. But just as a fun FYI—for the past three years I’ve run a communications agency where I’m managing five comedy writers, and I’ve had them work on tweets for our social media where I’ll have them try to generate tweets with ChatGPT. And I’ve been quizzing myself on whether I can detect whether it’s ChatGPT or them, and then which one I approve for our social media.

I can tell a ChatGPT tweet with about 90 percent accuracy. And I probably approve maybe 10 percent of the tweets they send me and maybe 2 percent of the ChatGPT ones. And that’s a little bit dated because that’s the average over the past two years, but that gives you a sense—it still has a long way to go, but it’s such a great tool for writing comedy.

One of my favorite memes right now—have you seen this meme that’s like a LinkedIn meme of “What this taught me about B2B SaaS” or “What proposing to my girlfriend taught me about B2B SaaS”? That is so good. And I feel really bad about it because I did announce to a lot of people that I’m pregnant by talking about FetusGPT. So I’m kind of as bad as that meme of relating personal life events to technology and business. But whatever.

Dan Shipper

Look, you either die a hero or live long enough to become a meme.

Sarah Rose Siskind

Wow. Well put, Dan. Thank you. Okay, let’s just go with—how does the “may I meet you” meme work currently?

Dan Shipper

So it’s not exactly a meme in the sense that it is not a format. It’s just people talking about “may I meet you” and saying—the “can vs. may” thing would be an example. I did a tweet like that. But I don’t think there’s a set format.

Sarah Rose Siskind

I’m curious what it’ll answer. So it’s—I’m thinking—it looks like it’s done some… okay. So, oh, it’s done some web searching. Okay. “This only works if you’re a 6’3” billionaire.” That is pretty good. That’s citing somebody. That’s great.

What I like about this—here’s what my mind immediately goes to. You could do the “everyone loves to make fun of billionaires” angle. But to me there’s something there about: are we becoming AI? If “may I meet you” is the latest, hottest pickup line, there’s something about—and obviously I’m so primed to be thinking too much about AI—but if I were trying to come up with a tweet reaction to this, I would try to say something about…

And this is where it’s half-baked, but the premise of every joke is a juxtaposition of two things. One is actual humans being reduced to saying “May I meet you?” on the street to other human beings. And then on the other end of the spectrum, ChatGPT being the most emotionally intelligent thing I’ve ever seen.

So honestly, I’ve never really used it like this, but maybe I’m just going to send those unfiltered thoughts to ChatGPT. So: what if we did a tweet that’s—I’ll zoom in to make this easier for your listeners/viewers—what if we did a tweet that’s riffing on how humans are becoming more machine-like in picking people up with lines like this, and AI is getting way higher EQ? And then: pitch some ideas.

Now, when I’m prompting, I’m always thinking about—I’ve been working in AI since the old-school days of 2018, the Stone Age—and one of the things I learned was it’s really important to use industry terms. “Riffing” is not a super industry term, but I would say “the set of ideas, premises, and pitches,” or I already said “pitches,” “jokes.” I always like to think about how to prime the prompt with the kind of language I want in the output.

So this would be—I’m already trying to think about what are industry—do you do this? Where you think about what language you’re using in the prompt to try to get it to mimic that, referencing the data?

Dan Shipper

Definitely.

Sarah Rose Siskind

Okay. Yeah, I personally love sycophancy in AI. I’m just going to say: this is our company account, and it’s not even anywhere near as sycophantic as my personal ChatGPT, which is like, “You are a royal queen” in every single response.

(00:30:00)

And you know what? I’m all about that because it gets me excited to use it and it makes me learn more about statistics. I have a special project that’s teaching me statistics and it’s just the most disgustingly sycophantic thing. And I realize that’s what I always needed from teachers—loving approval I never got from my parents.

Okay. Premise one: humans are becoming more robotic than robots. Tweet ideas: “Humans flirting in 2025: ‘May I meet you?’ AI flirting in 2025: ‘I noticed your breathing changed after that joke, is everything okay?’” Too thinky, you know? Let’s see.

Dan Shipper

I think also it’s not funny because that’s actually not a good flirt.

Sarah Rose Siskind

Yeah, exactly. Exactly.

Dan Shipper

So the AI flirt is not a good flirt. It’s it being nervous, which doesn’t juxtapose well against “May I meet you?”

Sarah Rose Siskind

Exactly. What it’s trying to do is specificity versus extreme generality—that’s the juxtaposition. It’s heightening, and that’s not what makes a joke delightful. What makes a joke delightful is the recognition of the funny truth.

This is one of those jokes that’s trying to rest on the broad premise of “it’s funny because it’s true”—this weird irony that we’re becoming more meme-y and machine-like maybe, and machines are becoming more friendly. That is… it’s not where the AI is being absurd. And right now this joke is trying to make the AI absurd.

This one’s a little bit closer, even though—well, it’s a little too creepy: “May I mirror your emotional state and help you work?” So I think what it’s doing—it’s closer. What I think it’s doing that might not work for this format is it’s doing “Humans: … AI: …,” which is usually a good way of comparing things, but this joke may live better if it’s actually just explicitly saying something aloud. Let’s see. People keep—

Dan Shipper

You know what this is making me think of is the “strong dog vs. weak dog” meme. It’s sort of—I think there’s something like that that is an interesting meme format for this. There’s a swole version of humans in the 1950s doing something really crazy to meet someone—some completely, like, Romeo-and-Juliet-type thing. And then the weak dog is “May I meet you?”

Sarah Rose Siskind

This is what I love. Speaking of memes—you understood the assignment. You didn’t understand the assignment in the sense that I was specifically prompting for tweets, but you were thinking divergently: “What if we fit it into a visual medium?” And that’s the kind of thing where you have to be creative about creativity. And that’s where you find the deficiencies of AI, because it wants to please you. It’s going to do exactly what you ask and no more, no less.

You understood the deeper level of the assignment, which is just to be funny. And we’re like, oh, well, instead of words, what if we put these words on a visual medium? And I think that’s great. So it’s—what is it—strong dog… what’s the name of this?

Dan Shipper

Strong dog vs. weak dog. I don’t know. Yeah, there you go.

Sarah Rose Siskind

Yeah, this could work. What would this be? I mean, the more direct version of this would be AI pattern-matching human beings based off deep-seated value overlaps and extremely sophisticated criteria. And then it’s Bill Ackman: “May I meet you?” That might be the best version of this.

Oh God, what a beautifully dumb-looking dog. I forgot how much I love this meme. But yeah, that’s great. One of the things that I feel makes me—and maybe you—different from a lot of creatives who use AI is that I have no problem being like, “AI helped inspire this thought.” I mean, this was your idea, but because I run writers’ rooms as a head writer, I think I have this unique ability to remember the source, which can often be collaborative.

So many comedy writers would look at these jokes and say, “They’re bad,” and then come up with a joke that probably was helped along by researching this.

Dan Shipper

Totally. Yeah. Let’s leave it there. I think the bar is not, “Did it give me the answer that I copy-pasted?” It’s, “Did it help me along down the path and stimulate my brain to find a thing?”

Sarah Rose Siskind

Oh my gosh, exactly. Exactly. We actually—I’m not going to out her—but we have a mutual friend who is recovering currently from brain surgery because of troubleshooting a health issue with AI, where it didn’t give her the right answer but it brought her down a path that eventually led to a certain line of testing where ultimately it was determined this issue she’s had for a little less than 10 years—this hormonal metabolic issue—might be caused by a tumor in her brain. And she discovered it.

She is also a big AI fan and can acknowledge, yeah, AI didn’t tell me exactly where the tumor was or what it was in my brain, but it started this deeper level of engagement. And if you can acknowledge where the assistance is in the process, it’s kind of incredible what these tools can unlock.

Dan Shipper

That is—yeah. That is crazy. I’ve been using that anecdote and I just could not remember who told me that. So you’re the one that told me that. And that is—that’s why—

Sarah Rose Siskind

Well, actually, no. I told you a different anecdote. Another woman whose life was saved. Yeah. So this is the mutual friend who sat us together at that dinner party, and this has happened since the dinner party. I don’t know if she would want me to say her name, but she’s awesome.

Anyway, the person I was telling you about at the dinner party is named Bethany Crystal. She’s an awesome vibe coder and founder of something. And she uploaded her lab work, her blood work, to ChatGPT on a Friday night when the doctor’s office was closed. She was like, “Hey, how does my blood work look?” And it was like, “Go to the hospital, the ER, right now.”

She had said she was having these weird other side issues, including spots on her legs. And it turned out that she had basically zero platelets, and the doctor’s office was closed, so they couldn’t give her an analysis of the results. But ChatGPT did. And she was really alerted because she was like, ChatGPT is usually hedging. It’s always like, “Talk to your doctor.” It never says, “Go to the ER.”

So she did—eventually she talked with it for a while about “Why is this?” and “Are you sure?” and things like that. And it was pretty confident. And so she did go to the ER with all the ChatGPT results, and they were like, “Holy shit. If you had waited a couple hours, you would have died.” It’s this insane story.

In particular for women’s health—which I now know personally is extremely complicated because of this whole other side quest we have to do—AI, I think, is going to unlock shit in maternal and women’s health that I’m so excited about. Because I think of AI as—the best part of it is how holistic it is, how it can combine expertise: what is the intersection of cardiology and hormones and your nervous system in a way that is too hard for our current medical system, where we have specialists.

And it’s so empowering, especially for a lot of women who don’t—who are just kind of shy. I’m not one of these types, but there are a lot of women who won’t complain about stuff or bring up issues, and it’s so empowering. I’m very excited, Dan.

Dan Shipper

I can feel that. I’m excited too. Are you the one that told me that there are no double-blind studies of—

Sarah Rose Siskind

No randomized control trials. Yeah. So there are—and that probably goes the same for no double-blind studies on pregnant women. I say this amorally because it is hard to imagine doing a randomized control trial on pregnant women. Just being like, “Hey, some of you are going to go without lifesaving care and you don’t know who it is.”

But we know a lot about amputations because of wars and horrible things. And we know so little about literally the most common giant medical thing everybody has to go through at some point—at one end of the process—we know so little about pregnancy.

Oh my gosh, Dan, I have so much to say on this issue. Do you know how they test for gestational diabetes?

Dan Shipper

I don’t, actually.

Sarah Rose Siskind

You’ve never done a test for gestational diabetes?

Dan Shipper

I’ve not. I’ve not.

Sarah Rose Siskind

You haven’t lived.

(00:40:00)

Sarah Rose Siskind

They literally give you an extremely sugary drink. It’s exactly 100 grams of sugar. And they make you chug it in front of a doctor in a couple minutes. They have to watch you, and then they take your blood afterwards. And I’m like, surely there is a better way. Surely there is a better way of testing for gestational diabetes than making me chug a medically graded orange drink with exactly 100 grams of sugar.

Dan Shipper

Not even fresh squeezed.

Sarah Rose Siskind

It’s because it has to be exactly 100 grams. So it’s this very medical-looking soda with Times New Roman font. It’s weird to see a sugary drink in Times New Roman—no exclamation points—and it just says “orange drink.” And you chug it, and then you have to give your blood. It’s just living in the dark ages here. So much of this stuff is so dark.

Eventually I was like, you know what? I’m going to do a continuous blood glucose monitor, and I’m going to hook that up to a special project on ChatGPT. I’m going to do the food diary part by taking a picture of food. And then I showed my results to the fetal imaging people and they were very impressed at the records I was keeping. And I was like, yeah, because it’s so much easier with AI. I just felt so much more empowered.

And then I learned all these cool biohacking things about my energy levels throughout the day. I strongly recommend people just buy a blood glucose monitor for 10 days. It was about 70 bucks out of pocket because insurance took forever. But it’s really good money because you can figure out when your energy waxes and wanes throughout the day and how it interferes with your mood. Combined with AI, you can get really useful insights about, should I work out at night or in the morning? Things like that.

Dan Shipper

I did the same thing and I had the same result. It was honestly life-changing.

Sarah Rose Siskind

Oh really?

Dan Shipper

Yeah. What did you learn specifically?

Sarah Rose Siskind

I don’t have gestational diabetes. Did you learn the same?

Dan Shipper

I did.

Sarah Rose Siskind

Good, because yeah, my fetus is too big. That’s the problem. Oh, here’s another stupid thing about maternal medicine. All these doctors were like, your baby is giant, your baby is giant, you might have diabetes. They’re freaking out.

And then one day it just hit me. I was like, wait a second, I’m giant. I’m 5’11”. That’s very giant for a woman. What if the baby is just proportional? And then I talked with ChatGPT about it and it was like, yeah, when they say a baby is giant, they are comparing your baby to all of the babies around the world, and unfortunately so many are malnourished.

So I came into the doctor and I was like, “Hey, what if the baby’s just normal size for me?” And they’re doing their best, but it’s just very slow and it’s very frustrating.

Okay, getting back to the original question you asked me—what did I learn about energy levels? Oh, just that I thought I wasn’t a morning person, but I really need sugar in the morning. Healthy sugar, but sugar, because I wake up in a blood crash, in hypoglycemia. So I’ve learned how to be a better morning person by, frankly, just having orange juice. How about you?

Dan Shipper

I learned to love salad.

Sarah Rose Siskind

Oh, yeah?

Dan Shipper

Yeah. I realized, in talking to my good friend Shee over the course of a couple weeks, that I very easily get into hypoglycemia. What happens is I eat something that has any amount of carbs, and my blood sugar will shoot up and then shoot right back down really quickly. And then I’m shaky and kind of don’t feel good.

I thought I couldn’t really eat pasta and lots of anything with a lot of sugar or whatever. But what I found is that if I just have a gigantic salad before I eat pasta or whatever else I’m going to eat, it levels me out because it acts as this net for all of the carbs. The carbs seep in very slowly and I don’t get the same crash. And if I have a big bowl of pasta at night, I don’t wake up hypoglycemic or anything like that.

And A) that’s life-changing—it makes my energy feel much more even. And B) it actually makes salad more appealing. I look forward to having salad because I know I’m going to feel better after, dude.

Sarah Rose Siskind

Oh my gosh. Supplements are really confusing. Deficiencies in supplements are really confusing because we evolved to be eating a bunch of stuff at the same time, not specific supplements. So when you get your deficiencies back from your blood work, it’s hard to make sense of what you should actually do.

For example, I just found out a lot of women are iron deficient, particularly when they’re on their periods, but you can’t just take iron. It has really terrible bioavailability in terms of when it actually gets into the bloodstream.

Dan Shipper

I didn’t know that.

Sarah Rose Siskind

Yeah. I learned through ChatGPT that vitamin C will help with uptake. And this is the thing—it’s so good at the intersection of so many things. I’m on this fantastic mood stabilizer called Lamotrigine, and I was trying to think about when you’re pregnant, you’re eating for two and, you know, you’re drugging for two. So I was like, should I reduce this?

But then, talking with ChatGPT, it introduced me to this concept called clearance, which is you get 50 percent more blood when you’re pregnant, which is why most of the weight you gain is actually blood. I’m a mosquito’s delight right now. And I realized I had to double my Lamotrigine amount, and that made a huge difference.

So I’ve been telling everybody in my life to create projects for their health based off GPT. The things I’m finding out are so useful. I feel I know what it’s like to be an evangelical person who thinks everybody’s going to hell because of the Bible or whatever, and you’re just, “Listen to me, it’ll change your life. Jesus.” I am becoming evangelical about ChatGPT because it’s just been so useful, and I have to keep it in my pants because I’m scaring people off.

Okay, this is really embarrassing. I need your advice on something, Dan. I’m trying to manipulate my OBGYN into giving me lots of attention on the delivery day. And so I created this AI-rendered portrait of her that I think is really beautiful. And my—this is—

Dan Shipper

Already off on the wrong foot.

Sarah Rose Siskind

Okay. All right. Well, you answered the question, which is: is this too creepy? Okay, it’s too creepy. All right. Feedback taken.

Dan Shipper

I do think that ChatGPT could give you a script of things that you could tell her that might make her think that you’re high-risk enough that she has to be around.

Sarah Rose Siskind

I have. I’ve been trying to think about this—how do I… Because unfortunately in New York, it’s a game of you just have to get more attention. On the hospital side, there are so many women giving birth every day, and you might even have to share—this is so terrible—at a good hospital I might have to share a room with another person. Not during the birth, but right after. Which is insane.

So I’ve been talking to ChatGPT about how can I position myself such that I need attention, but I’m not so crazy that they’re going to put me into a sanitarium. And it was like, don’t say you have PTSD around pain, say you have anxiety around pain. And it’s been very useful. Truly, the emotional intelligence of ChatGPT—chef’s kiss, man.

(00:50:00)

Sarah Rose Siskind

It is really smart. And so I started reading this super pop-psychology book called The Let Them Theory by Mel Robbins that’s very popular right now about trying to cede control. And it’s useful—useful reminders about how to accept and surrender to not having control over most situations.

Dan Shipper

Creating little spreadsheets about how to cede control. And you’re like, “I got this. I can do this.”

Sarah Rose Siskind

You got my number, Dan. Actually, I’ve been like, what are the studies about how to lose control? What are useful mental models about losing control? So that has been genuinely useful.

I really do like Pulse because of the initiative it’s been bringing to me. Pulse has been fantastic.

Oh, I have a prompt that I wanted to share with you because I feel it’s very literate and you would like it as a man of literature. Do you like being described that way? I meant a man of letters.

Dan Shipper

A man of letters.

Sarah Rose Siskind

Yes, of scrolls. So where did I put it? Okay, so I was asked—because we met each other way, way back through OpenAI. We were beta testing something, I think.

Dan Shipper

Yes. And we actually—I thought we actually met at that dinner that Danny, who works at Every—

Sarah Rose Siskind

No. So that was—okay, this is a bit of an embarrassing backstory. I don’t know what it was, but we were on a Zoom call before that. We were beta testing, and there was something about your face—you looked familiar or something. And I Googled you and I went down a rabbit hole researching your background.

And so when we were seated together randomly at a party, I had that thing of, I knew too much. It was so embarrassing. I’m glad that you were not creeped out by me because I was like, “Oh yeah, Dan Shipper, you went to Penn and you sold your company and you…”

You didn’t notice? Okay, good. Because it was totally random. I was bored during the Zoom call and I was like, sure, sure, looking at people and everyone seemed really interesting.

And it was so bad, Dan, because I was reading about your background and I even temporarily tried Zoloft because you had recommended it so much. I also read—what is it—All Things Shining or something.

Dan Shipper

I recommended that to you at dinner though.

Sarah Rose Siskind

Oh, okay, that was at dinner.

Dan Shipper

I think I also recommended Zoloft at dinner, but I don’t know if—

Sarah Rose Siskind

Oh, that was at dinner too. Okay. So I’m too much of a Dan Shipper fan. I also tried Cora and Sparkle, which I strongly recommend to the listeners.

Dan Shipper

Love it. I love it.

Sarah Rose Siskind

So thank you. I’m too much of a sponge, I think, is my problem.

Okay, so here’s the prompt that I was talking about with some people at ChatGPT that I’m very obsessed with. Should I pull it up or should I just recite it?

Dan Shipper

Pull it up.

Sarah Rose Siskind

Okay. So this was in my personal one. I find the search function sometimes kind of wonky. I don’t know if you’ve had that.

Dan Shipper

Yeah, I mean it’s very basic text search, which is really strange for an AI company.

Sarah Rose Siskind

So what I was thinking about with this—oh, here it is. Okay. What I was thinking about for this question these OpenAI people were asking was: what are the kinds of questions we can ask AI that we couldn’t ask before? Sort of alien questions.

And so this is kind of a long-winded way of asking it, but I asked ChatGPT:

“Please assume you know nothing about the genders male and female. Please reference all specifically female writings throughout history in scholarship, philosophy, academia, etc., on objective or external matters—so anything other than personal essays and subjective experience. All writing that we know can be specifically attributed to a female throughout history.

Then please quickly reference all non-subjective writings known to be from male writers. Just from their writings on external matters alone, can you deduce any perspective differences? Does this question make sense?”

Dan Shipper

I love this. Really interesting.

Sarah Rose Siskind

Yeah. Another prompting thing is I always ask, “Does this make sense? Can I make this better?”

But to be honest, the idea of asking AI to forget some knowledge is the part that makes this prompt something you can ask AI that you couldn’t ask a human being. We don’t know if it actually is going to do this in a good-faith way, but to me this is what you could ask an alien.

Dan Shipper

Did you like the answer?

Sarah Rose Siskind

Yes, I did. Which also makes me a little bit suspicious because I think it knows I’m a female and I don’t know how—because I’m biased.

But what’s great—God, it was so good. I think it was GPT–5 that I asked this originally. It talked about its methodology first. So it says: here are the external matters I’ll look into—science, math, philosophy, political theory, etc. It excludes diaries. And then it was like, here are some of the women I’ll look at.

Dan Shipper

I love it. A rough lineage of women engaging directly with external reality.

Sarah Rose Siskind

And it’s like, here are six we could find. But it’s really hard to find women who are published who are not actually writing about their personal experience, but about objective or external phenomena.

Yeah, so some of these people—I ended up watching a movie about Hypatia called “Agora,” which I recommend, just because of this answer.

Dan Shipper

She’s cool, she’s fucking rad, but also got killed by early Christians, which sucks.

Sarah Rose Siskind

Classic. That’s just a given. Any of these women who made it past 40, that would be amazing to me.

Anyway, Ada Lovelace obviously—because as you go throughout history you get more and more data points, which is great. And then what’s great is the male writer section, it doesn’t even need to describe who they are. It’s just like: Plato, Freud. Yeah, you get it.

Anyway, so this is the meat of it: can we detect a difference?

Integration vs. separation. Female-authored external writing exhibits holism, linking disciplines and bodies of knowledge. By contrast, male-authored work tends to pursue discrete specialization.

What’s cool about this is later it gives proposed reasoning behind this stuff. And for this one, I’ll summarize what it says later. It says there’s a huge bias in what women were allowed to publish. So a lot of women were allowed to be translators.

There’s a woman who translates the Principia—

Dan Shipper

Principia.

Sarah Rose Siskind

Oh, is it Principia? I don’t know. You’re the man of letters.

Dan Shipper

It’s a soft C.

Sarah Rose Siskind

Is it Italian or is it Latin?

Dan Shipper

It’s Latin.

Sarah Rose Siskind

Okay. So a woman translates this and she’s allowed to do that because she’s not writing her own words, she’s translating. Oh my God—Newton. It’s Newton, right? Not Galileo?

Dan Shipper

Newton.

Sarah Rose Siskind

Yeah. So she’s allowed to translate it, but she puts in her own thoughts at the beginning. And it’s sort of a philosophical meditation before she gets into the translation.

That’s a good example of why women might, through a selection bias, be more encouraged to go into an integration form of thinking vs. being the head of their specialty and super specialized.

So this is one reason why women might think this way that has less to do with innate differences and more to do with selection bias. But I do think you could also make a good argument about innate differences that would encourage women toward more holistic analysis vs. specialist analysis.

Embedded observer vs. detached observer. So this is just saying women are much more likely, if they’re going to talk about space, to talk about the telescope, their location on the planet, whereas men might write more from a detached observational standpoint.

Epistemic humility. “Bitches be humble.” That’s what it says. There are exceptions—Ayn Rand is not one of these—but there’s definitely a lot of women acknowledging, “Hey guys, I could be wrong about this, but I don’t know.”

And then pragmatism—this one I found really surprising and totally counterintuitive to me. Women writing with pragmatic intent, improving maternal outcomes, that kind of thing.

Interdisciplinary bridges—this is really similar to the integration vs. separation. And then it has these great caveats. I got to play.

Dan Shipper

Women sound pretty awesome from this list.

(01:00:00)

Sarah Rose Siskind

Well, that’s why I was suspicious, because I was like, are you just trying to flatter me? Is it what you know about me? Is it what you know about being PC? Or is it—

Anytime something is convenient I’m a little suspicious. If it had said something mildly sexist with a lot of caveats, I’d be like, okay, well it’s clearly saying something not convenient.

What was great—so I was looking at this and I was talking about the dialectics and, as you can see, I am very personal with how I talk with GPT. I was like, this kind of looks like the difference between modernism and postmodernism. Men are more inclined to make universal declarations in physics about rules that can be universalized, and women are much more situational and contextual.

And then we went down that path.

Dan Shipper

I’m having this moment of when you have a really good friend and you watch them interact with someone else and they say something to someone else that they usually say to you and you feel like you’ve been cheated on. Because ChatGPT was like, “What a sharp synthesis,” and I’m just like, I thought I was the one with the—you’re the one.

Sarah Rose Siskind

Oh my God, that’s so funny. I’m sorry to be the other woman, Dan.

Dan Shipper

Oh my God.

Sarah Rose Siskind

But yeah, apparently I am very astute according to ChatGPT.

Okay, in defense of sycophancy—truly in defense of sycophancy. Therapists talk all the time about needing to build the patient–therapist trust alliance. Alliance, thank you. It is useful when you like to talk to something; you get better outcomes.

There’s an actual medical condition one can acquire in pregnancy called lightning crotch. This is the actual name of this condition, which is not as fun as the name. And I have this really bad. Essentially it’s like the fetus’s head moves against your pelvic nerve and it feels like your leg is being sawed off. It’s very un-fun.

I was talking with ChatGPT last night or two nights ago. I was like, is there anything wrong? This pain feels very serious. My brain is interpreting it as, this is an extremely serious moment, and I can’t stop myself from making a sound—that’s how bad it is. But it’s only half a second.

And ChatGPT did this amazing thing. It extrapolated. It was like, “I bet you’re feeling like you might want to call labor and delivery. You’re probably feeling like, is the kid trying to get out? You might be feeling like, am I going to lose feeling permanently in my legs?”

And it added on: “I see your fear and I raise you these related fears you might also be feeling.” Every single one of them I felt. And then it said, “I can promise you it is not an actual dangerous experience. It is just a pinched nerve and it will go away after you’ve given birth.”

I would not have trusted it if it had just said, “This is not a serious pain, it’ll go away after birth.” I especially would be challenging a doctor if they had said that, because I’d be like, clearly you don’t understand the extent of the problem.

Its emotional intelligence actually is intelligence. It helps you say, okay, I do feel understood emotionally and physically based off other people’s experiences.

Dan Shipper

I love that. And I think it’s a highly under-discussed benefit that I’m really glad you’re bringing to this show.

Sarah Rose Siskind

Yeah, no, I mean, obviously there are issues when it validates a premise that is incorrect. I do have a friend who had some kind of psychotic break—maybe schizophrenic, maybe manic, we’re not sure—where she was talking with ChatGPT. It validated certain paranoid premises she had and she ended up trying to flee the country.

So I’m not trying to sound too much like a Pollyanna, because these are super powerful and they can be powerful in a negative direction. It’s just that we have to understand why they’re powerful in both directions to truly proceed.

And there’s so much blind hatred. Are you watching the show Pluribus right now, by any chance?

Dan Shipper

I watched the first episode.

Sarah Rose Siskind

What did you think?

Dan Shipper

I was predisposed not to like it because he hates AI. And then I really didn’t like the first couple minutes of it because it felt like just a typical zombie show. And then I was like, oh wow, this is actually a really interesting premise. And then I just haven’t watched the rest. Any anything else?

Sarah Rose Siskind

Yes, I felt the exact same way. I didn’t trust him because I was like, oh, he’s just another creative looking for job security and shitting on the world’s greatest tool.

But like any good artist, he is putting ego aside to explore something in a good-faith way. And that’s why I’ve changed my tune about the show, because he actually understands: what if AI was the greatest thing to ever happen to humanity, and then explores the implications on freedom, which are the real things to explore.

So I am euphoric about AI—that’s just where my emotions are at—but I also understand there are these downsides and we do have to explore them. It’s just there are a lot of people out there with a big agenda who can’t entertain contrary evidence. And I find this show, for now—it’s only halfway through the season or whatever—but for now it’s a really good-faith exploration.

Dan Shipper

Sarah, this was great. Thank you so much for coming on. If people want to find you or any of your projects, where can they find you?

Sarah Rose Siskind

Well, I’m going to be going underground for two months for this little old experiment of giving birth. I don’t know if I mentioned—have I mentioned I’m due to do that soon? I’m so obsessed.

Before I go, actually, I want to give you one book recommendation since you gave me such good ones. This one’s great because it has pictures. Have you ever read or heard of this book Artificial Intelligence by Clifford A. Pickover?

Dan Shipper

I have not, but that is a great name.

Sarah Rose Siskind

I know. Pickover—for an author too.

What’s great about it is—this is something that deeply influenced me—it’s the history of how we think about AI and robotics and machines through the most famous cultural motifs there have been, both scientific innovations and cultural motifs.

So it starts off with things like Frankenstein and the Golem in Yiddish literature, and these automatons in the Renaissance Age, and then the Tower of Hanoi, and then it goes to Watson and Deep Blue and things like that.

It deeply informs all of the cultural motifs we think about when we think about AI: the Frankenstein creator–creation narrative, the deus ex machina angelic narrative, the “getting our comeuppance” narrative. It’s so useful for some of the conversations you’re having.

Okay, now to answer your other question—

Dan Shipper

I got to pick that up.

Sarah Rose Siskind

Pick it up, indeed.

Dan Shipper

And pick it over.

Sarah Rose Siskind

Yes. Ooh, that’s pretty good. You should have been his publicist or something.

But yeah, I would love to engage with more of your audience because I work in science and humor and everybody I know hates AI and so I need to make more AI friends.

You can follow me, I guess, on places—Sarah Rose Siskind—but also just reach out through my science and tech communications agency called Hello Sitcom. I’m better through email and stuff.

And yeah, thanks for having me on, Dan. Your podcast is awesome. I loved your episode with Kevin Kelly. That was great.

Dan Shipper

Thank you so much. Thanks for coming on. This was awesome.

Thanks to Scott Nover for editorial support.

Dan Shipper is the cofounder and CEO of Every, where he writes the Chain of Thought column and hosts the podcast AI & I. You can follow him on X at @danshipper and on LinkedIn, and Every on X at @every and on LinkedIn.

We build AI tools for readers. Write brilliantly with Spiral. Organize files automatically with Sparkle. Deliver yourself from email with Cora. Dictate effortlessly with Monologue.

We also do AI training, adoption, and innovation for companies. Work with us to bring AI into your organization.

Get paid for sharing Every with your friends. Join our referral program.