WEBVTT - Did AI Write This? 0:00:15.396 --> 0:00:33.356 Pushkin. I'm Jacob Goldstein, and this is What's Your Problem? 0:00:33.616 --> 0:00:36.976 My guest today is Max Spiro. He's the co-founder and 0:00:37.036 --> 0:00:40.936 CEO of a company called Pangram. And Max's problem is this. 0:00:41.656 --> 0:00:44.566 How do you build a tool that that reliably identifies 0:00:44.686 --> 0:00:47.826 whether a piece of writing or an image was created 0:00:47.926 --> 0:00:51.706 by AI. If you've heard about any of the recent 0:00:51.806 --> 0:00:56.376 controversies over whether something was written by AI, whether it 0:00:56.436 --> 0:01:01.696 was a prize-winning novel or an op-ed in a major newspaper, 0:01:02.296 --> 0:01:05.076 there's a good chance that you've actually heard about Pangram, 0:01:05.216 --> 0:01:07.996 whether you know it or not, because Pangram is the 0:01:08.036 --> 0:01:12.226 tool people almost always cite in those cases. It is 0:01:12.306 --> 0:01:18.306 generally viewed as the most reliable, widely available AI identification tool. 0:01:18.916 --> 0:01:21.316 Max and I talked about a lot of different things 0:01:21.436 --> 0:01:25.376 in our conversation. We talked about changing norms around when 0:01:25.416 --> 0:01:28.076 it's OK to use AI and when it's not OK. 0:01:28.816 --> 0:01:31.996 We talked about the risk of false positives, of PanGram 0:01:32.216 --> 0:01:35.976 wrongly accusing people of using AI to write things. And 0:01:36.036 --> 0:01:39.056 we talked about the new PanGram tool to detect whether 0:01:39.136 --> 0:01:42.656 an image was AI generated. But to start, we talked 0:01:42.676 --> 0:01:45.996 about two domains where PanGram is already in wide use, 0:01:46.676 --> 0:01:51.556 education and publishing. Tell me about PanGram and sort of 0:01:51.696 --> 0:01:54.266 where you are now. Who are your customers? Who's using it? 0:01:54.676 --> 0:01:58.536 Our customers, a bunch of customers in education, higher ed, 0:01:58.576 --> 0:02:01.916 and also like high schools, K-12, they want to know 0:02:02.096 --> 0:02:05.946 whether their students are submitting AI-generated essays. They want a 0:02:05.966 --> 0:02:12.596 way to teach writing or, like, have any writing-based assignments 0:02:13.136 --> 0:02:16.596 without having it sort of just devolve into students using AI. 0:02:16.676 --> 0:02:20.346 It seems like an extremely strong case. Like, you know, 0:02:20.466 --> 0:02:23.146 writing is a way of thinking hard about things. I 0:02:23.226 --> 0:02:26.726 buy that it is useful still for students to write. 0:02:27.906 --> 0:02:31.806 My sense is that students are using AI a lot and, 0:02:32.046 --> 0:02:34.306 you know, going to great lengths to sort of humanize 0:02:34.356 --> 0:02:36.536 what the AI says, to sort of write through AI 0:02:37.096 --> 0:02:38.616 or do what they have to do. Like, what is 0:02:38.656 --> 0:02:39.956 your sense of how it is going? 0:02:40.316 --> 0:02:42.956 Yeah, something like pangram here would be more of like, 0:02:43.366 --> 0:02:46.986 we're adding friction to the process of cheating. And, you know, 0:02:47.466 --> 0:02:49.546 it's going to make it less enticing if you have 0:02:49.606 --> 0:02:52.256 to go start with AI bones and rewrite it. 0:02:52.426 --> 0:02:54.366 Yes, you can't one-shot your assignment. 0:02:54.386 --> 0:02:54.886 Exactly. 0:02:55.086 --> 0:02:55.846 Yeah, yeah. 0:02:55.966 --> 0:02:59.426 And that's helpful in some way. And, you know, you 0:02:59.466 --> 0:03:02.466 can encourage students to, you know, start from scratch and 0:03:02.486 --> 0:03:03.346 not use AI at all. 0:03:03.386 --> 0:03:06.436 But I mean, certainly there are teachers who are requiring 0:03:06.496 --> 0:03:10.086 students to do their writing in the room. Right? Like, 0:03:10.206 --> 0:03:13.066 they're just going hard at, you have to sit here 0:03:13.106 --> 0:03:14.896 in front of me. So what did somebody call it? 0:03:14.986 --> 0:03:16.906 AI jail or something? You have to sit here in 0:03:16.956 --> 0:03:18.016 front of me and write an essay. 0:03:18.076 --> 0:03:20.796 It's kind of a bummer. Like, I feel like we 0:03:20.836 --> 0:03:24.336 lose a lot here. We lose the in-class instruction. We 0:03:24.456 --> 0:03:29.496 lose the ability to assign longer written pieces. It's true. 0:03:29.516 --> 0:03:31.136 When I was in college, I wrote a lot of 0:03:31.156 --> 0:03:33.526 like eight and 10, 12 page essays. 0:03:34.406 --> 0:03:34.706 Yeah. 0:03:34.746 --> 0:03:37.146 Like those, you can't do that in class. That takes 0:03:37.186 --> 0:03:37.706 a lot of hours. 0:03:38.266 --> 0:03:40.546 I mean, but that's just the nature of the world now, right? Like, 0:03:40.626 --> 0:03:41.766 it sort of is what it is. 0:03:42.406 --> 0:03:45.876 Yeah, and it's a little bit adversarial, as many of 0:03:46.116 --> 0:03:49.316 our use cases are, where, like, the way that Pangram 0:03:49.336 --> 0:03:51.976 is used is somebody is trying to get away with 0:03:52.016 --> 0:03:55.076 AI writing in a way. And so there is this 0:03:55.196 --> 0:03:56.256 adversarial component. 0:03:56.396 --> 0:03:58.936 Yes. I mean, it's got to be mostly adversarial, right? 0:03:58.956 --> 0:04:02.206 The weird non-adversarial one would be, I actually wrote this thing, 0:04:02.326 --> 0:04:04.266 and I don't want people to think it's AI, so 0:04:04.286 --> 0:04:06.006 I'm going to put it in Pangram, right? Like, that's 0:04:06.486 --> 0:04:10.686 an edge case. Like, I've started using fewer em dashes, actually. 0:04:10.826 --> 0:04:12.806 Oh, man. You got to use the em dashes. 0:04:13.306 --> 0:04:17.526 Journalists love em dashes. And actually, maybe because I love them, 0:04:17.606 --> 0:04:21.076 I didn't even notice it in AI pros. But once 0:04:21.116 --> 0:04:23.396 everybody started talking about the em dash as a giveaway, 0:04:23.496 --> 0:04:27.056 I just started doing, like, shorter independent clauses, basically. 0:04:27.176 --> 0:04:30.096 Em dashes are, like, they're no longer an AI tell. 0:04:30.416 --> 0:04:34.426 This was, like, a GPT-4 era thing. maybe like 4.0, 0:04:34.426 --> 0:04:37.286 really overuse the em dash. We're over it now. Em 0:04:37.306 --> 0:04:38.626 dashes are used a normal amount. 0:04:38.746 --> 0:04:40.616 I can bring the em dash back? I can bring 0:04:40.676 --> 0:04:40.946 it back? 0:04:40.976 --> 0:04:42.056 Yes, you should bring it back. 0:04:42.496 --> 0:04:46.936 Okay, great. So education is obviously a big use case. 0:04:47.556 --> 0:04:48.996 I'll tell you about publishing as well. 0:04:49.376 --> 0:04:51.156 Oh yeah, publishing. Well, so that one has been in 0:04:51.196 --> 0:04:53.616 the news, right? There have been a couple of people, 0:04:53.776 --> 0:04:56.916 one person lost a seven-figure book deal, right? Because, I 0:04:56.916 --> 0:04:59.536 don't want to say because of you, but because they 0:04:59.576 --> 0:05:01.996 were accused of using AI to write the book and 0:05:03.056 --> 0:05:04.176 you were definitely involved in that. 0:05:04.636 --> 0:05:06.816 Yeah, there have been a series of scandals this year 0:05:07.036 --> 0:05:10.696 on AI use, either in books that are already published 0:05:10.836 --> 0:05:14.146 or like big book deals. But kind of what it 0:05:14.436 --> 0:05:16.886 comes down to is that a lot of people in 0:05:16.906 --> 0:05:19.466 the industry are on the left side of the bell 0:05:19.506 --> 0:05:23.586 curve where they're not heavy AI users. They're like, they're 0:05:23.786 --> 0:05:30.696 book editors and publishers. They read like real books all day. 0:05:31.076 --> 0:05:34.636 So they're not like, you know, talking to chat. So 0:05:34.896 --> 0:05:36.796 I think like a lot of them, they like never 0:05:36.836 --> 0:05:39.476 really built up this intuition in a way that like 0:05:39.556 --> 0:05:42.816 all of a sudden, like AI getting good enough to 0:05:42.916 --> 0:05:46.616 like produce publishable prose was like very surprising and kind 0:05:46.636 --> 0:05:47.636 of caught them by surprise. 0:05:48.316 --> 0:05:53.566 So recently, famous investor Stanley Druckenmiller wrote an op-ed in 0:05:53.586 --> 0:05:57.126 the Wall Street Journal about the bond market. And everybody's like, wait, 0:05:57.306 --> 0:06:00.836 AI wrote that. And then Druckenmiller was like, you're damn 0:06:00.876 --> 0:06:02.776 right AI wrote that. I'm 80 years old and I 0:06:02.816 --> 0:06:05.696 know how to use AI. And then maybe most interestingly, 0:06:06.356 --> 0:06:09.336 the editor of the journal op-ed page was like, it's 0:06:09.376 --> 0:06:12.476 fine that AI wrote it. Like Druckenmiller is a famous investor. 0:06:12.556 --> 0:06:15.966 He really believes this. He stands behind it. Any famous 0:06:16.006 --> 0:06:19.046 person writing an op-ed in the paper usually has somebody 0:06:19.066 --> 0:06:21.186 else write it for him anyway. There's no problem here. 0:06:21.986 --> 0:06:23.846 Sort of surprisingly, I might think I would have a 0:06:23.846 --> 0:06:26.386 problem with it, but upon reflection, I don't. Do you 0:06:26.406 --> 0:06:27.026 think it's bad? 0:06:27.806 --> 0:06:30.486 I think the argument against is that it still does 0:06:30.926 --> 0:06:34.026 a disservice to the readers. Like, if you read it, it's, like, 0:06:34.146 --> 0:06:36.066 full of AI prose. 0:06:36.596 --> 0:06:39.056 So your critique is as a piece of writing. It's 0:06:39.136 --> 0:06:41.256 bad writing, and that is the criticism. 0:06:41.656 --> 0:06:44.976 Yeah, and, like, the Wall Street Journal has great editors. 0:06:45.076 --> 0:06:47.236 I'm sure, like, he could have just, like, written up 0:06:47.736 --> 0:06:51.816 a draft himself that was, like, not particularly flowery or 0:06:51.836 --> 0:06:53.916 whatever and then, you know, sent it to his editors and, like, 0:06:54.336 --> 0:06:58.476 they help him work it into something that's, like, very readable. 0:06:58.916 --> 0:07:01.996 You could also argue, like, it certainly would be less 0:07:02.036 --> 0:07:05.656 objectionable if it just said at the bottom, this was 0:07:05.716 --> 0:07:08.696 written with the assistance of AI, right? Like, that would 0:07:08.856 --> 0:07:11.516 seem to, like, if you have any criticism, it's like, well, 0:07:11.536 --> 0:07:14.576 you should tell me because it's weird. Yeah. Although you 0:07:14.616 --> 0:07:16.846 could make the same argument for ghostwriters and, like, pretty 0:07:16.876 --> 0:07:21.266 much anything you see written by a CEO, in most cases, 0:07:21.526 --> 0:07:23.866 of a big company is going to be ghostwritten, right? Like, 0:07:23.886 --> 0:07:26.366 there are people whose job is writing op-eds for CEOs. 0:07:26.666 --> 0:07:28.566 I'm going to make this other argument now, which I 0:07:28.606 --> 0:07:31.226 think is maybe more relevant. So I think there's like 0:07:31.286 --> 0:07:36.006 two types of AI use and AI writing. There's basically 0:07:36.066 --> 0:07:39.466 making something more concise or distilling. So for example, if 0:07:39.846 --> 0:07:42.726 Druckenmiller's like, hey, I wrote like, you know, five pages, 0:07:42.786 --> 0:07:44.926 here's all my research, here's all my thoughts. I need 0:07:45.026 --> 0:07:47.386 AI to help me like make this concise and get 0:07:47.426 --> 0:07:49.946 it into like op-ed size. And so I think typically 0:07:49.966 --> 0:07:53.786 AI writing in this case ends up being quite helpful 0:07:53.826 --> 0:07:56.656 because there's a lot to work with and it's distilling 0:07:56.696 --> 0:07:59.976 it down to its core components. The other side of 0:07:59.996 --> 0:08:03.096 AI writing that I see is expanding or padding, where 0:08:03.136 --> 0:08:07.156 somebody's like, here's a few bullet points. Here's a few 0:08:07.196 --> 0:08:10.426 things that I believe. Turn this into an email or 0:08:10.466 --> 0:08:14.006 an op-ed. I think that's where it's really doing a 0:08:14.046 --> 0:08:17.416 disservice to the readers because you're just putting more words 0:08:17.456 --> 0:08:19.396 that are basically padding and don't mean that much. 0:08:19.416 --> 0:08:21.336 Just publish the bullet points, bro. 0:08:21.376 --> 0:08:22.416 Exactly, exactly. 0:08:22.436 --> 0:08:25.676 Yeah, publish it in Axios. They love bullet points over there. 0:08:26.596 --> 0:08:29.416 There's a broader thing that's interesting to me about the 0:08:30.016 --> 0:08:33.876 Druckenmiller Wall Street Journal thing, and that is, like, it 0:08:33.956 --> 0:08:37.946 does make me wonder to what extent we're right now 0:08:38.106 --> 0:08:42.566 in a transitional moment where, you know, I could imagine 0:08:42.586 --> 0:08:44.006 in a year, a few years. 0:08:45.456 --> 0:08:48.336 In many settings, saying, did a person use AI to 0:08:48.356 --> 0:08:50.656 write this, would be like today saying, did they use 0:08:50.736 --> 0:08:54.616 the internet to do the research? Like, probably, and so 0:08:54.656 --> 0:08:59.216 what are the answers? Does that seem plausible to you? 0:08:59.996 --> 0:09:03.226 I think we're absolutely in this space or this time 0:09:03.236 --> 0:09:06.066 period where the norms are completely up in the air 0:09:06.226 --> 0:09:09.586 and people are so polarized on one side or the other, 0:09:09.606 --> 0:09:12.406 or some people are like, I'm going to let AI 0:09:12.666 --> 0:09:15.666 automate my entire online presence. And other people are swearing 0:09:15.986 --> 0:09:20.106 they'll never use AI ever. And so we have this 0:09:20.146 --> 0:09:25.216 really wide range of norms. Nobody's really settled on what 0:09:25.276 --> 0:09:29.896 is acceptable versus unacceptable. I think AI and LLMs, they're 0:09:29.916 --> 0:09:33.216 very powerful technology. They could do a lot of work 0:09:33.316 --> 0:09:36.096 for me. They can go do research for me. They 0:09:36.116 --> 0:09:39.696 can compile this into a spreadsheet for me to read 0:09:39.776 --> 0:09:40.746 and peruse. 0:09:40.886 --> 0:09:43.676 They can determine whether people are using AI to write. 0:09:44.696 --> 0:09:48.516 Yeah. But yeah, I think the answer is we really 0:09:48.596 --> 0:09:53.276 haven't settled on any norms societally. And the norms that 0:09:53.316 --> 0:09:56.196 we will settle to are probably that some degree of 0:09:56.216 --> 0:09:59.806 AI use is okay, but using AI to replace your 0:09:59.866 --> 0:10:03.966 voice completely, especially in settings where it's going to be 0:10:04.046 --> 0:10:08.846 read by thousands or millions, is probably less acceptable. Yeah. 0:10:08.986 --> 0:10:11.736 Well, and disclosure. I do feel like disclosure can solve 0:10:11.776 --> 0:10:14.836 a lot. It's also interesting to me. You see this 0:10:14.916 --> 0:10:17.986 certainly on LinkedIn and on Twitter. There are a lot 0:10:18.016 --> 0:10:21.826 of things that are so obviously AI. And initially, I 0:10:22.046 --> 0:10:24.526 was like, don't people know? But now I think maybe 0:10:24.566 --> 0:10:27.446 people know and don't care. Like, I don't know. It 0:10:27.486 --> 0:10:30.216 does feel... I want to be clear, I don't like 0:10:30.376 --> 0:10:32.696 AI prose. I'm not speaking of a moral objection, but 0:10:32.736 --> 0:10:36.756 just as a matter of taste. And then there is 0:10:36.796 --> 0:10:38.916 a little bit, like, it feels a little rude. It 0:10:38.956 --> 0:10:41.646 feels a little spammy. Like, I'm reading this thing, and 0:10:41.666 --> 0:10:44.566 this person didn't even take the time to lightly edit 0:10:44.626 --> 0:10:47.866 it to get rid of the obvious AI-isms. 0:10:47.886 --> 0:10:50.576 Yeah, or if you aren't going to take the time 0:10:50.616 --> 0:10:53.276 to write it yourself, then I'm so much less inclined 0:10:53.346 --> 0:10:53.786 to read it. 0:10:54.156 --> 0:10:56.646 So the question I have is, do the people who 0:10:56.706 --> 0:10:59.526 are posting that know that it is obviously AI and 0:10:59.546 --> 0:11:01.506 they just don't care? I mean, I guess it varies, 0:11:01.586 --> 0:11:03.686 but that's a weird one to me still. 0:11:03.766 --> 0:11:05.326 I think there's a little bit of a bell curve. 0:11:05.716 --> 0:11:07.616 There's a lot of people who are either new to 0:11:07.696 --> 0:11:11.076 using AI or they haven't really used that much AI before, 0:11:11.116 --> 0:11:13.956 and they can't really look at text and immediately grok 0:11:13.976 --> 0:11:17.676 that it's AI-generated. They can't immediately tell. And so, yeah, 0:11:17.816 --> 0:11:19.906 obviously you have the people who are just like, don't 0:11:19.986 --> 0:11:21.736 use AI much. And then you also have the people 0:11:21.746 --> 0:11:25.016 who like just discovered this technology and they're like, wow, 0:11:25.156 --> 0:11:27.236 it like writes so well, this is amazing. And I 0:11:27.236 --> 0:11:31.396 don't think they fully realized that like regular AI users 0:11:31.876 --> 0:11:35.156 can read this and immediately smell that it's AI generated. Yeah. Then, 0:11:35.236 --> 0:11:36.566 then you get to the middle of the bell curve 0:11:36.586 --> 0:11:39.846 where it's like, yeah, like, like people who use AI 0:11:40.566 --> 0:11:42.846 on a somewhat regular basis, they, they can kind of 0:11:42.866 --> 0:11:44.846 just tell. And I think you get to the other 0:11:44.926 --> 0:11:46.926 end where there's people who are like so deep in 0:11:47.026 --> 0:11:51.446 AI psychosis that, that they also can't tell because like 90% 0:11:51.446 --> 0:11:54.466 of what they're reading is like AI already. They talk 0:11:54.506 --> 0:11:56.686 to chatbots more than other people and they just like 0:11:57.076 --> 0:12:00.376 have lost this ability to differentiate. Uh-huh. 0:12:00.916 --> 0:12:02.936 So maybe we'll all be at that far end of 0:12:02.976 --> 0:12:03.956 the curve in a year. 0:12:04.196 --> 0:12:05.016 I really hope not. 0:12:05.036 --> 0:12:10.446 Yeah, I suppose I also hope not. So the companies 0:12:10.486 --> 0:12:17.146 that are making frontier models, Anthropic, OpenAI, arguably Alphabet, Google, 0:12:17.886 --> 0:12:23.426 are giant companies with wild amounts of money, full of geniuses. 0:12:23.946 --> 0:12:27.026 You'd think those companies could just do what you're doing. 0:12:27.186 --> 0:12:31.246 They could just build their own AI detectors. Why don't they? 0:12:31.726 --> 0:12:35.046 I agree. I don't disagree here. But I also think 0:12:35.116 --> 0:12:39.296 that this is just not in their top 10 priority list. 0:12:40.256 --> 0:12:43.736 They're building AGI. They're automating white-collar work. 0:12:44.226 --> 0:12:47.326 Probably the right answer. I mean, there's an answer, which 0:12:47.386 --> 0:12:52.006 is like, is it contrary to their incentives to do 0:12:52.046 --> 0:12:56.036 what you're doing, right? Is it in their interest for 0:12:56.096 --> 0:12:59.196 people to be able to distinguish AI pros from human pros, 0:12:59.236 --> 0:13:00.756 or is it contrary to their interest? 0:13:00.836 --> 0:13:02.956 I think it is contrary to their interest. There was 0:13:02.976 --> 0:13:07.136 an interesting article that I was exploring a couple of 0:13:07.156 --> 0:13:11.936 years ago, actually, when OpenAI had watermarking technology, they'd built it, 0:13:12.336 --> 0:13:14.376 but then they didn't want to release it. There was 0:13:14.396 --> 0:13:16.336 a lot of internal politics on this, and this was 0:13:16.436 --> 0:13:20.316 largely because they were afraid it would hurt their core 0:13:20.376 --> 0:13:23.436 user base, which was at least a plurality of students 0:13:23.476 --> 0:13:23.956 at this time. 0:13:24.416 --> 0:13:26.776 They would get caught cheating if there was a watermark. 0:13:26.836 --> 0:13:27.796 It's basically what you're saying. 0:13:27.836 --> 0:13:30.216 They would get caught cheating, or they would churn to 0:13:30.716 --> 0:13:32.686 a different AI service that didn't have a watermark. 0:13:33.036 --> 0:13:37.206 So you brought up watermarks, so let's talk about watermarks. 0:13:37.806 --> 0:13:41.526 The EU, as I understand, just passed a law requiring... 0:13:41.566 --> 0:13:45.066 watermarks on text from language models. Is that true? 0:13:45.146 --> 0:13:49.006 Yes, that's true. And OpenAI and Anthropic are both working 0:13:49.046 --> 0:13:51.526 on this. Google already watermarks text outputs. 0:13:52.186 --> 0:13:53.426 Is that bad for your business? 0:13:54.686 --> 0:13:56.826 I think it's really good that people are taking this 0:13:56.866 --> 0:14:00.006 problem seriously. It is competitive with us, but I think 0:14:00.106 --> 0:14:03.266 Pangram has a lot of advantages that watermarks don't have, 0:14:03.486 --> 0:14:05.386 which is that a watermark is just going to tell 0:14:05.406 --> 0:14:07.756 you either Claude was here or not, and Pangram will 0:14:07.796 --> 0:14:10.196 be able to tell you with higher granularity of like, 0:14:10.736 --> 0:14:12.716 What was the degree of AI assistance? Was it just 0:14:12.896 --> 0:14:14.586 helping or did it write the whole thing? 0:14:14.606 --> 0:14:20.946 I mean, presumably, there's some universe where Claude outputs the 0:14:20.966 --> 0:14:24.326 copy that's watermarked, and then you literally just retype it 0:14:24.406 --> 0:14:28.866 into a Google Doc. I mean, naively, I would think 0:14:28.906 --> 0:14:31.386 the watermark would not follow. Or is the watermark somehow 0:14:31.526 --> 0:14:34.096 encoded in the prose so that if you retype it, 0:14:34.126 --> 0:14:34.706 it's still there? 0:14:34.806 --> 0:14:40.026 Yes, the watermark is statistically encoded in the word choices themselves. 0:14:40.046 --> 0:14:43.706 Oh, that is rad. So even retyping it, the watermark 0:14:43.766 --> 0:14:44.326 is still there. 0:14:44.806 --> 0:14:50.326 Yes. With that said, you could run some automated paraphraser 0:14:50.406 --> 0:14:53.386 to just paraphrase all the words, and that would erase 0:14:53.426 --> 0:14:53.966 the watermark. 0:14:54.286 --> 0:14:57.786 Yes. Well, so there are humanizers now, right? There are 0:14:59.066 --> 0:15:03.306 things you can upload your language model pros into that 0:15:03.386 --> 0:15:07.086 claim to evade detection. Although I will say I tried one, 0:15:07.106 --> 0:15:09.716 I won't name it, but I tried one in preparing 0:15:09.736 --> 0:15:11.836 for this interview, and you'll be glad to know it 0:15:11.856 --> 0:15:14.396 didn't work. I had the language model output pros, I 0:15:14.436 --> 0:15:16.656 ran it through a humanizer, put it into PanGram, and 0:15:16.676 --> 0:15:18.076 PanGram said 100% AI. 0:15:18.236 --> 0:15:21.816 Amazing. Yeah, I mean, this is part of the adversarial 0:15:21.876 --> 0:15:24.766 nature of the business is We are also training on 0:15:24.866 --> 0:15:28.546 humanizer outputs because we need to make sure that pangram 0:15:28.906 --> 0:15:33.246 can detect anything that's been run through a humanizer. Otherwise, 0:15:33.306 --> 0:15:35.946 it would kind of be trivial to evade pangram. 0:15:36.586 --> 0:15:39.546 Why can't they beat you? I would think you could 0:15:39.586 --> 0:15:43.886 be at a degree of randomness at the level of 0:15:44.486 --> 0:15:50.216 word choice and syntax that would defeat you. Why can't they? 0:15:50.516 --> 0:15:52.836 I think these are kind of like two, there's two 0:15:52.936 --> 0:15:57.136 competing poles for text when you're humanizing. You want it 0:15:57.176 --> 0:16:02.776 to be coherent and, you know, retain its logical flow. 0:16:02.996 --> 0:16:05.356 And then on the other side, you want to just 0:16:05.436 --> 0:16:07.656 degrade it or perturb it a little bit enough that 0:16:07.696 --> 0:16:11.266 it is outside of the range of AI detectors. There 0:16:11.426 --> 0:16:15.226 are humanizers that can beat pangram and you could do 0:16:15.266 --> 0:16:19.066 this by just like perturbing the text way more just 0:16:19.186 --> 0:16:19.746 a lot more. 0:16:19.966 --> 0:16:24.566 Well, sure. I mean, there's a difference between degrading and perturbing, right? Like, 0:16:25.086 --> 0:16:28.266 you can make it really bad, but that's obviously not useful. 0:16:28.426 --> 0:16:31.726 Well, I think degrading is just a more extreme version 0:16:31.786 --> 0:16:32.386 of perturbing. 0:16:32.966 --> 0:16:35.446 Well, you could improve. You could perturb it a lot 0:16:35.486 --> 0:16:37.106 and improve it. Oh, yeah, yeah. 0:16:37.126 --> 0:16:39.406 But these humanizers, they don't have the resources to actually 0:16:39.476 --> 0:16:40.146 do that, right? 0:16:40.216 --> 0:16:40.696 Yeah. 0:16:41.016 --> 0:16:45.156 Most of them are, like, one- to ten-person companies. 0:16:45.716 --> 0:16:48.216 There's a more fundamental question here that is interesting to me, 0:16:48.236 --> 0:16:52.326 and I don't really understand the answer, which is, like, 0:16:52.566 --> 0:16:55.526 why does AI slop all sound the same? You know 0:16:55.546 --> 0:16:57.366 what I mean? Like, that's the thing that's so annoying. 0:16:57.386 --> 0:17:00.706 Why can't you just turn up the volume, not on, like, 0:17:00.926 --> 0:17:04.426 make it bad, but just, like, don't use that same rhythm. 0:17:04.466 --> 0:17:06.886 Don't do, like, the one sentence. Don't use these annoying 0:17:06.966 --> 0:17:10.306 words that are AI words. Like, why is that at 0:17:10.386 --> 0:17:11.666 a technical level? 0:17:12.046 --> 0:17:15.226 It's really hard to say. I think... So there's a 0:17:15.246 --> 0:17:18.526 concept in machine learning called mode collapse. And that means like, 0:17:18.906 --> 0:17:23.326 you know, once an AI model learns what is best 0:17:23.626 --> 0:17:26.846 or most common, which is the mode, then it will 0:17:27.126 --> 0:17:30.506 kind of like collapse the edges of its distribution and 0:17:30.546 --> 0:17:32.666 just output the mode much more often. 0:17:33.126 --> 0:17:36.526 Yes. The mode being the modal word, the most common word. 0:17:36.546 --> 0:17:39.386 Yeah, the most common word or phrase or sentence construction. 0:17:39.406 --> 0:17:42.266 I think that's part of it. And I think also... 0:17:42.536 --> 0:17:45.706 a lot of these labs, they do post-training. They'll create 0:17:45.806 --> 0:17:49.206 a reward model and they'll give the language model a reward. 0:17:49.406 --> 0:17:52.846 And so I can only suspect that there's somebody in 0:17:52.926 --> 0:17:57.096 here who's saying there's some reward for, quote, like good writing. 0:17:57.956 --> 0:18:01.976 And that might include, you know, writing with em dashes 0:18:02.136 --> 0:18:05.176 is probably correlated with good writing in the pre-AI world. 0:18:05.736 --> 0:18:08.206 Same thing with the like Like, it's not just X 0:18:08.246 --> 0:18:09.886 but Y sort of metaphors. 0:18:10.006 --> 0:18:12.026 Yes. That's another one that I have taken out of 0:18:12.086 --> 0:18:13.696 my own writing, actually. Yeah, yeah. 0:18:13.756 --> 0:18:16.956 Like, it's a very reasonable construction, but then you feed 0:18:16.976 --> 0:18:20.116 it into the reward model, and then the AI learns 0:18:20.136 --> 0:18:21.656 in its prose to turn that up to 11. 0:18:21.656 --> 0:18:23.176 It likes it too much, yeah. 0:18:23.416 --> 0:18:26.436 Yeah. So it just overuses what would have normally been 0:18:26.456 --> 0:18:27.876 considered good writing. 0:18:31.476 --> 0:18:43.016 We'll be back in just a minute. Okay, let's talk 0:18:43.056 --> 0:18:45.436 about false positives, right? This is a thing people care 0:18:45.476 --> 0:18:48.176 about and have written about. So obviously in any test, 0:18:48.196 --> 0:18:50.056 you can have a false positive or a false negative. 0:18:50.096 --> 0:18:52.416 A false negative, you're saying this is human generated and 0:18:52.456 --> 0:18:55.476 actually it's AI. That one people don't get so upset about. 0:18:55.496 --> 0:18:59.156 The one people are really afraid of is wrongly accused, right? 0:18:59.196 --> 0:19:02.776 Pangram says it was AI. Actually, it was written by 0:19:02.816 --> 0:19:06.896 a person. I've seen you say one in 10,000 is 0:19:07.196 --> 0:19:11.516 the false positive rate. And I've also seen outside studies that, 0:19:12.066 --> 0:19:14.806 find a very low false positive rate, a similar order 0:19:14.826 --> 0:19:16.966 of magnitude, one from the University of Chicago I found 0:19:17.006 --> 0:19:20.876 particularly compelling. But like, I don't know how many writing 0:19:20.916 --> 0:19:23.286 assignments students have in high school and college, but it's 0:19:23.496 --> 0:19:27.716 probably order of magnitude of hundreds, right? And so that 0:19:27.756 --> 0:19:29.876 means that some share of students are going to be 0:19:29.936 --> 0:19:35.476 wrongly accused, right? If it's 500 writing assignments, then it 0:19:35.496 --> 0:19:38.676 would be 5% of students will be wrongly accused by 0:19:38.716 --> 0:19:41.316 Pan Graham of using AI to write a thing that 0:19:41.336 --> 0:19:42.056 they actually wrote? 0:19:42.416 --> 0:19:45.476 Yeah, I do think like 1 in 10,000 is beyond 0:19:45.516 --> 0:19:48.996 the threshold where like you can use this as a 0:19:49.016 --> 0:19:52.686 tool and generally typically be confident that when it says 0:19:52.746 --> 0:19:53.866 it's AI, it's correct. 0:19:54.186 --> 0:19:56.726 It's true. That's definitely true. Like, I mean, I will 0:19:56.746 --> 0:19:59.656 say just I was actually accused by a teacher in 0:19:59.696 --> 0:20:02.236 high school of plagiarizing. That's the old school. And I 0:20:02.276 --> 0:20:05.676 had not plagiarized, for the record. And so that is 0:20:05.736 --> 0:20:08.666 the old school human version of this. And I'm sure 0:20:08.726 --> 0:20:12.846 his error rate was higher than 1 in 10,000. But 0:20:12.906 --> 0:20:14.886 I am sensitive to it. Like, that feeling of being 0:20:14.926 --> 0:20:17.046 wrongly accused is a terrible feeling. Mm-hmm. 0:20:17.586 --> 0:20:19.896 Yeah, so I think one thing that pangram helps with 0:20:19.996 --> 0:20:24.356 actually is it's much better than like the average human intuition. 0:20:24.736 --> 0:20:27.666 And it's also better than any other AI detector. So 0:20:27.746 --> 0:20:32.026 because of that, the existence of pangram is, I truly 0:20:32.046 --> 0:20:35.826 believe this, reducing the number of false accusations for AI 0:20:35.956 --> 0:20:38.156 use in a scenario like education. 0:20:38.956 --> 0:20:42.036 Yeah, that's fair. So like notionally, like, of course, teachers 0:20:42.076 --> 0:20:43.796 are on the lookout for AI one way or the 0:20:43.876 --> 0:20:45.476 other if they want students to write it. 0:20:45.496 --> 0:20:47.776 Exactly. And then the other side, I think, which is 0:20:47.816 --> 0:20:53.036 really important, is that it's not the only source of information. 0:20:53.556 --> 0:20:56.806 Once Pangram says that something's AI, we can always go deeper. 0:20:56.886 --> 0:21:01.146 We can either ask the student for drafts. If you 0:21:01.266 --> 0:21:04.666 use Google Docs, Pangram actually has this Chrome extension where 0:21:04.686 --> 0:21:07.586 you could see the full revision history, you could see 0:21:07.626 --> 0:21:09.006 that the document being typed out. 0:21:09.186 --> 0:21:11.946 And that history is available even without PanGram, right? 0:21:11.966 --> 0:21:12.366 Exactly. 0:21:12.466 --> 0:21:14.476 And if you're using Chrome, you can see the version 0:21:14.496 --> 0:21:15.316 history and go back. 0:21:15.676 --> 0:21:19.216 Exactly. Yes. Yeah. So that version history is there. But 0:21:19.596 --> 0:21:22.176