A Fascinating Conversation on A.I. Consciousness (And Why It Matters)

1. 0:00 – 0:01

   Intro

   1. KRKevin Roose0:00

      Well, Casey, it's time to check

2. 0:01 – 23:30

   Social Media bans are coming

   1. KRKevin Roose0:01

      in on a story we've covered periodically on this show, which is the state of the social media backlash and these social media bans that have been going into effect in countries around the world.

   2. CNCasey Newton0:11

      Yes, Kevin, you may remember that at the end of last year, I made a prediction that by the end of 2026, 16 plus would become the new norm for getting a social media account worldwide. But I think I have one thing that I'm gonna feel confident about, with high confidence, in 2026, and it's this: I think that 16 plus becomes the new norm for social media accounts worldwide.

   3. KRKevin Roose0:35

      Mm.

   4. CNCasey Newton0:36

      And as we enter July, Australia, Brazil, Indonesia, Malaysia, France, the United Kingdom, Denmark, and Slovenia have either enacted, uh, laws or are preparing measures that would limit children's use of social platforms, so typically barring kids under 15 or 16 from TikTok, Instagram, Facebook, YouTube, and X. That's a big change.

   5. KRKevin Roose0:56

      That is a big change, and I remember you making that prediction, but I also remember that you didn't include Slovenia-

   6. CNCasey Newton1:01

      [laughs]

   7. KRKevin Roose1:01

      ... in your list of countries that would, uh, apply one of these bans, so I'm only gonna gr- grant you partial credit on that one.

   8. CNCasey Newton1:06

      Here's what I've always said about Slovenia. They could do anything.

   9. KRKevin Roose1:08

      [laughs]

   10. CNCasey Newton1:09

       They're... It's a very d- dynamic place, is, is what I'll say about Slovenia. So tho- that was the prediction, but over the last few weeks, we have gotten several big updates that make me feel like, uh, n- you know, number one, this, this prediction is, uh, absolutely gonna happen, but more importantly, I just think the internet is going to start feeling like a very different place sometime soon.

   11. KRKevin Roose1:30

       Okay, make your case. Maybe start close to home. Uh, what is happening in the US with social media bans?

   12. CNCasey Newton1:36

       So on Monday, the U.S. Supreme Court declined to block a Texas law that requires app stores and developers to enforce age verification and make kids get parental consent to download apps. So this is one of the stricter forms of, uh, age assurance, as it's called, that you will find out there. Texas passed a law in 2025 called the App Store Accountability Act, and if you live in Texas and you're a kid, you have to link your account to a parent or guardian, and then your parent or guardian has to approve any app before your minor can download it.

   13. KRKevin Roose2:12

       So the Supreme Court did not take up this case, which basically means that, what, this is now legal to do? You can, if you are a state, you can pass a law saying that you have to age verify to get onto social media?

   14. CNCasey Newton2:23

       Yeah, there was a legal fight over it. Industry groups tried to fight it in court, and a judge at one point blocked the law on the grounds that it likely violated the First Amendment. But in June, the 5th US Circuit Court of Appeals put that judge's order on hold. And so when the Supreme Court says, "Hey, we're not even going to look at this," that means that, yes, it goes into effect in Texas, and it will likely give other states who want to do the same thing, uh, carte blanche to do so. And in fact, there are four other states who have passed laws very similar to the one that Texas has passed.

   15. KRKevin Roose2:57

       Which one? Name them from memory. Don't look at your laptop.

   16. CNCasey Newton2:59

       I would love to name them from memory. They are, of course, Utah, Louisiana, and Alabama, so I guess it was really three states, plus Texas.

   17. KRKevin Roose3:06

       [laughs] Okay, very good.

   18. CNCasey Newton3:07

       [laughs] But California, interestingly, has passed a, uh, similar law that's less restrictive. It takes effect, uh, on January 1st of next year, and it requires operating system providers to collect age or birthdate information, uh, when you're setting up a new device. So, you know, what is interesting about that group of states is obviously now you have red and blue. This is becoming a bipartisan thing across the country.

   19. KRKevin Roose3:30

       And are all these states sort of defining social media similarly? Like, are th- are they all including the same websites? I know YouTube has sort of been on the cusp for some advocates of these bills, like, 'cause obviously, there's educational stuff on YouTube. Kids are using it in school. But it's also s- sort of social media, at least the way that some of these attempts to, to sort of age-gate it have, uh, have portrayed it.

   20. CNCasey Newton3:50

       Yeah, for, for a minute it did look like YouTube was gonna be able to wriggle out of these restrictions, but increasingly, it is being lumped in with these other apps. And I think it's important to say that that doesn't mean that a kid will be denied from using YouTube, because of course you can still open up a website and type in youtube.com and play videos. But it will prevent you from using an account. And advocates would say this does have safety benefits, so for example, nobody can send a message to your kid if they don't have an account. Uh, maybe the recommendations won't be perfectly tuned to, you know, their particular, uh, brain rot if, uh, you know, they aren't logging in every time they watch YouTube.

   21. KRKevin Roose4:26

       Right. So those are the state-level bills. Is there any movement on a federal ban on social media? Like, that was something that you and I had talked about earlier, that the US government actually step in and do something like this nationwide, but are we seeing any movement on that?

   22. CNCasey Newton4:39

       Not really. Uh, the Congress has been debating a raft of other child safety measures, and as usual, it seems like it's getting very close to the finish line and then falling apart at the last second. We've seen that, uh, several times over the past few weeks. I think what is interesting in, in the context of will we ever see a federal ban, is that Pew published new research last week which found that nearly six in 10 US adults do support banning social media for under 16. So this is now the majority position in the United States, and I would guess that if eventually you see, mm, let's say, like, half to three-quarters of states implement a ban like this, then that's probably when Congress will feel, uh, you know, confident enough to pass it at the federal level.

   23. KRKevin Roose5:18

       So the first kind of ban like this that we ever talked about was the one that went into effect in Australia-

   24. CNCasey Newton5:24

       Yeah

   25. KRKevin Roose5:24

       ... which I think has been the template or the model for a lot of these other bans going on around the world. And my understanding is that we actually now are starting to just get some data back about how the Australian social media ban is going. Um, there was a very good story by Jon Herman in New York Magazine recently about, uh, how hard it's been to ban teens from social media in Australia. But, like, what are some of the details there? What do we know about how that experiment is playing out?

   26. CNCasey Newton5:49

       Yeah, so there was a new study examining the early effects of the ban that came out two weeks ago, and basically these researchers at the, uh, University of Newcastle in Australia did a 90-day check-in, and the big headline was it found that more than 85% of kids reported using social media, uh, roughly three months after the ban took effect. And so that led to a lot of people sort of saying, "Aha, I told you so. These sorts of bans are impossible," or they said, "Aha, look at these terrible social media companies. They're doing nothing to kick kids off." Um, Australia's government in response is now introducing legislation that would double the fines against social media companies that do not take this more seriously. But I have to say, Kevin, as I dug into it, I became a lot less convinced that this is a failure.

   27. KRKevin Roose6:35

       What do you mean?

   28. CNCasey Newton6:35

       Well, when you look at what Australia's law actually requires, the requirement is that platforms take reasonable steps to try to keep kids off of these platforms. It does not dictate a particular method. And so what we're seeing the platforms do is adopt a range of strategies, uh, one of which is, um, what's called age inference, which is looking for signals that you're underage. So for example, if you, you know, create a new Snapchat account and everyone else that you message, you know, seems to be, like, a young teenager, this may give Snapchat a signal that you're underage. Or if you're-

   29. KRKevin Roose7:11

       Yeah, if you're using the word low-key a lot-

   30. CNCasey Newton7:13

       [laughs]
3. 23:30 – 55:58

   J-space; A.I. welfare and A.I. consciousness

   1. KRKevin Roose23:30

      Well, Casey, there was something that caught my attention this week, and it was something called JSpace, and at first I thought, "Another dating app for Jewish singles?"

   2. CNCasey Newton23:38

      [laughs] Now, that's interesting, Kevin, because I've been pronouncing it JSpace in the French style.

   3. KRKevin Roose23:43

      [laughs] But as it turns out, JSpace is a new discovery from interpretability researchers at Anthropic about what they describe as an internal workspace in Claude that is analogous to the way that humans internally think and process unconsciously.

   4. CNCasey Newton24:00

      And it was named after a mathematical concept called the Jacobian, which you probably already know, but just in case we have any kindergartners listening-

   5. KRKevin Roose24:08

      [laughs]

   6. CNCasey Newton24:09

      ... is a way to describe how a multivariable function changes locally. You can think of it as the multivariable analog of the derivative.

   7. KRKevin Roose24:15

      Oh, now it makes sense.

   8. CNCasey Newton24:17

      Mm-hmm.

   9. KRKevin Roose24:17

      I got it.

   10. CNCasey Newton24:18

       Yeah.

   11. KRKevin Roose24:18

       So this was fascinating. I, I stared at this paper kind of blankly for a while, just sort of trying to wrap my head around what it meant. Um, there's still a lot I want to, uh, understand about this research and what it means, but I thought that this would be a great week to have a discussion about A.I. consciousness-

   12. CNCasey Newton24:37

       Yes

   13. KRKevin Roose24:38

       ... because this is a topic that I think just a few years ago was very fringe, very taboo. You kind of couldn't find many credible people talking about it. Um, the people who were talking about it were largely ostracized or, or sort of seen as, as kind of out there. Um, but now this is becoming a real topic that labs including Anthropic are starting to study. They made very clear in their JSpace research publications that they were not saying that this proves that Claude is conscious or sentient or anything else, but it is kind of similar in s- in certain ways to some of the research that consciousness researchers have been doing on humans, uh, for many years now.

   14. CNCasey Newton25:17

       Yeah, and i- if nothing else, I think that this paper speaks to the increasing sophistication of these models, right? Like, we are a long way from pure next-token prediction. These models now have these, you know, internal elements that are, uh, appear to play a huge part in the way that they reason and communicate, and we are just now beginning to understand them and to have discussions about the potential implications of that sophistication continuing to grow.

   15. KRKevin Roose25:49

       Yes. So if you wanna learn a lot more about the Anthropic JSpace, uh, research, you can go read their, uh, papers and blog posts about it on their website. But today we're gonna have a conversation with someone who's been thinking about the topic of A.I. consciousness and welfare for longer than almost anyone. Jeff Sebo is an associate professor of environmental studies at NYU. He is also the director of the Center for Mind, Ethics, and Policy, and he is one of the foremost researchers in the world on this topic of A.I. consciousness. He's been writing about it for years, and he and a group of colleagues have a new report that just came out last week called Studying A.I. Welfare Empirically, where they dig into some of these thorny, hard-to-nail-down issues around A.I. sentience and consciousness, whether these models might become conscious someday, uh, whether they might be deserving of some kind of moral consideration, and how we would even start to answer those questions using a more scientific approach.

   16. CNCasey Newton26:45

       Yeah, so if you're in the mood for a sort of heady conversation about the big questions of life and the universe, I would say buckle up.

   17. KRKevin Roose26:52

       Yes, very heady stuff, and before we get into it with Jeff, we should make our A.I. disclosures. I work for The New York Times. We're just doing OpenAI, Microsoft, and Perplexity.

   18. CNCasey Newton27:01

       And my fiancée works for Anthropic.

   19. KRKevin Roose27:02

       Jeff Sebo, welcome to Hard Fork.

   20. JSJeff Sebo27:04

       Thanks so much for having me.

   21. KRKevin Roose27:05

       So I learned about your work first with a previous study that you and some of your colleagues did. This was a paper that came out in 2024 called Taking A.I. Welfare Seriously, and now you and your colleagues are back with a new report called Studying A.I. Welfare Empirically. We'll get to the report in a little bit, but just to kind of give, uh, some background and context for people who haven't been following these discussions about A.I. consciousness, when people in your field talk about A.I. consciousness, what exactly are you talking about?

   22. JSJeff Sebo27:37

       Yeah, that is a great first question because consciousness is a word that means many things to many people, even in science and philosophy. So people often use it to mean being awake instead of being asleep or being self-aware instead of being not self-aware, and even when you zoom in on the specific concepts that are most important in science and philosophy, there are still different ones. So two that might be especially important for our conversation today are what philosophers call access consciousness and phenomenal consciousness.

   23. KRKevin Roose28:06

       Now, I of course know what those things are, but, but define them for Casey.

   24. JSJeff Sebo28:09

       [laughs] Sure. Casey-

   25. KRKevin Roose28:11

       [laughs]

   26. JSJeff Sebo28:11

       ... please listen carefully.

   27. KRKevin Roose28:12

       All right.

   28. JSJeff Sebo28:13

       Access consciousness is a functional concept. It refers to states that can be accessed by the system and used for reporting, for reasoning, for control, whereas phenomenal consciousness is more about feeling. And so normally when people are talking about welfare and moral status and ethical responsibilities, they have phenomenal consciousness in mind. The question is: Does it feel like something to be this system? And in particular, might the system be capable of having feelings and emotions like pleasure, pain, happiness, suffering, satisfaction, frustration that feel good or bad?

   29. CNCasey Newton28:52

       And at a high level, I, I, you know, I, I think it's great that this is being studied, but when I've talked about it with some people, some people want to know, like, why even bother studying that? Like, it, it seems sort of so ridiculous to people on its face that anyone would, um, imagine that an LLM could have conscious. Tell us a little bit about why this is, um, such an area o- uh, of interest for you and your colleagues.

   30. JSJeff Sebo29:15

       Well, one reason is this is very important. Figuring out what entities in the world can have feelings and emotions like pleasure, pain, happiness, suffering, that makes a really big difference for how we ought to treat them, how we ought to interact with them, and it can be really bad to make mistakes in either direction. As many people have pointed out, it can be really bad to over-attribute consciousness to nonhumans- To see them as being conscious when they are in fact not because that could lead to inappropriate social and emotional bonds with them, misallocating concern to them, even interacting with them in ways that increase risks involving with misuse and loss of control in the case of AI. But then it can also be really bad to under-attribute consciousness to non-humans, to treat them as lacking consciousness when in fact they have it. That can lead to abuse and neglect of vulnerable populations, as has often been the case with non-human animals. We presumed they lacked consciousness. We scaled up industries like factory farming, and then later we realized they do in fact have sophisticated feelings and emotions, but now we are entrenched in these industries, and it will take a long time to transition away from them.
4. 55:58 – 1:14:57

   Casey and Kevin show us their new way of burning through tokens

   1. KRKevin Roose55:58

      Well, Casey, you came into our meeting this week very excited to tell me about some new tools that you are using, and so we thought what better time to break out our trusty old segment, Tool Time. [upbeat music]

   2. CNCasey Newton56:25

      Aw, yeah.

   3. KRKevin Roose56:26

      [laughs] All right, Casey, for our show and tell this week for Tool Time, what do you have to show me?

   4. CNCasey Newton56:33

      Yeah, so this conversation picks up on some of the vibe coding discussions that we had earlier in the year. Longtime listeners may remember that around the start of 2026, all of Silicon Valley was agog with the new possibilities that were being created by new coding agents and by Claude Code in particular. And while I made a bunch of things in those early days, Kevin, I had sort of fallen off. I had made the kind of stuff that I wanted to make, and then I heard about a new app called Glaze, and I wanted to try it out.

   5. KRKevin Roose57:07

      Uh, so tell me about Glaze.

   6. CNCasey Newton57:08

      So Glaze is made by the fine folks over at the Raycast company. Raycast makes another one of my favorite tools. It is a launcher. It is also my primary way for, uh, interacting with AI during the day, just like a very cool app. It's also free. I highly recommend you go check it out. But they decided to get into a new line of business, and that new line of business is vibe coding. They have an app called Glaze. They just, uh, started to allow everyone to use it this month, and once you install it, you can use it to make Mac desktop apps. And this winds up being really cool because while you could certainly make an app with a Claude Code or a Codex, this is a much more visual way of going about it, and they've already sort of set up all of the templates. So you're not gonna have to, you know, lose an afternoon to prompting and setting up your scaffolding and compiling all of your code. With Glaze, you make a Mac desktop app with just, you know, entering a few sentences into, um, a, a box, and after you get the result, Kevin, you can just edit it live. So you can even cir- If, if it gets one of the little user interface elements wrong, you can circle it and say, "Hey, I want this arrow bigger," something like that. So just winds up be- if you're a non-technical person like me, this winds up being a really fun way to make a Mac desktop app.

   7. KRKevin Roose58:28

      All right. So as, uh, any conversation in San Francisco these days must start, what are you building?

   8. CNCasey Newton58:33

      Well, I am so glad you asked. I'm gonna show you, uh, two things that I have built in, uh, the Glaze app so far. I will say that there is a sort of introductory free version of Glaze that they can use. They'll give you some credits. Uh, you may run out of those quickly, in which case you're gonna have to pay them, uh, 20 bucks a month if you wanna get some more credits and build. Um, just to sort of set your expectations there, I have gone ahead and spent the 20 bucks, and I spent a little bit more on top of that. But let me go ahead and show you-

   9. KRKevin Roose59:02

      That's the, the, they call that the Glaze Pays option.

   10. CNCasey Newton59:05

       That is the Glaze Pays. Very good. Thank you. So with that, I'm gonna go ahead and show you the first thing that I built. This is the one that is a little bit more kind of serious and useful. All right, so this, Kevin, is the Platformer app. Um, I know that for ever since I started Platformer, you've wondered, "How can I get this as a Mac desktop app?" Now there is finally a solution to that. But this is actually just a tool that I made that solves a common problem that I have, which is it's time to write my newsletter. I know that I've written about this subject before, but when was it exactly, and, like, how many times have I written about that over the years? And of course, the way that I have approached this problem until very recently was just by googling it, and that worked well enough. But it wasn't an app on my Mac, Kevin.

   11. KRKevin Roose59:50

       It didn't use billions of tokens-

   12. CNCasey Newton59:52

       Exactly

   13. KRKevin Roose59:52

       ... to accomplish the same thing.

   14. CNCasey Newton59:54

       It's-

   15. KRKevin Roose59:54

       And you thought, "I can solve this problem."

   16. CNCasey Newton59:56

       That's right. I wanted something that solved my problem while also exacerbating the urban heat island effect-

   17. KRKevin Roose1:00:01

       [laughs]

   18. CNCasey Newton1:00:01

       ... surrounding data centers. So this is what I built, and the way that I did it was I, uh, sort of exported every, uh, article that's ever been in Platformer. I ingested it into this app, and now I have a box that I can type into. Uh, you know, so go ahead. Ask me any question about Platformer, Kevin.

   19. KRKevin Roose1:00:19

       What are my biggest psychological insecurities?

   20. CNCasey Newton1:00:23

       What are my biggest psychological insecurities? Which is a question that a, a writer should always ask. And so now I have plugged in, um, a, an Anthropic API key here, not, not one of the expensive ones. I'm using one of the cheap models, uh, for this one. And, um, and so now it's gonna go ahead, and it's going to draw a selection of articles that it thinks might be relevant to that topic, and it's go a- it, it's gonna go ahead and, uh, answer your question. Um, and, uh- Let's see if it says anything interesting or funny here.

   21. KRKevin Roose1:00:54

       Ooh, the most consistent intellectual insecurity across the whole span is the fear of being wrong in public, and your unusually disciplined habit of admitting it.

   22. CNCasey Newton1:01:03

       Wow. Well, that's very nice of it to say-

   23. KRKevin Roose1:01:05

       [laughs]

   24. CNCasey Newton1:01:05

       ... I suppose. Um, but, you know, as you look through, it actually has lots of links to those columns, so I can click out and immediately dive into any of those columns. So this is honestly, like, much faster than a Google search at trying to, like, get an answer about a range of things that I have written.

   25. KRKevin Roose1:01:19

       This is the very high-tech version of, like, muttering to yourself while walking down the street. No, I do actually understand why this is, uh, why this is valuable to you. 'Cause often I am like, I forget everything I write-

   26. CNCasey Newton1:01:30

       Right

   27. KRKevin Roose1:01:30

       ... about six hours after I write it.

   28. CNCasey Newton1:01:32

       Com- completely.

   29. KRKevin Roose1:01:32

       And so people are always trying to, like, you know, ask me about something I wrote two weeks ago, and it's like, well, I, I couldn't tell you, I have the memory of a hummingbird.

   30. CNCasey Newton1:01:38

       Absolutely. So this is why I built the homepage of this app, to show me an RSS feed of my 10 most recent columns. So those are just always open on the homepage, and then I also created this little browser. So I had it extract all of the topics that I've written about the most, and all the people that I've written about the most. And so now if I'm like, "Hey, what, you know, what was the last time, like, I wrote a column about Elon Musk?" Now I just click on Elon Musk, and oh, look, here's, like, everything, you know, that I've written about him for the past two years. So, you know-

Episode duration: 1:14:57