---
title: "Invisible Ink: AI Watermarking, SynthID & the Internet Meltdown"
podcast: "Modem Futura"
episode: "Episode 99"
kind: transcript
hosts: "Sean Leahy and Andrew Maynard"
date: 2026-08-25
audio: https://cdn.simplecast.com/media/audio/transcoded/fca32cee-ef4c-4b50-82e5-4c768318a8c1/018f87c3-4d42-4bc0-9d66-53873b51ec68/episodes/audio/group/68464dea-c54e-4225-9b83-50cd29a37693/group-item/2b47129d-b3d9-4839-ac66-eacd5d0f3194/128_default_tc.mp3?aid=rss_feed&feed=3frRI0HN
show_notes: https://text.futureofbeinghuman.com/modemfutura/99-invisible-ink-ai-watermarking-synthid-the-internet-meltdown.html
mirror: https://text.futureofbeinghuman.com/modemfutura/99-invisible-ink-ai-watermarking-synthid-the-internet-meltdown-transcript.html
---

# Invisible Ink: AI Watermarking, SynthID & the Internet Meltdown

*Full transcript*

Modem Futura · Episode 99 · August 25, 2026 · 49 min · Hosts: Sean Leahy and Andrew Maynard · about 10,200 words

Show notes: <https://text.futureofbeinghuman.com/modemfutura/99-invisible-ink-ai-watermarking-synthid-the-internet-meltdown.html> · Listen: [MP3 audio](https://cdn.simplecast.com/media/audio/transcoded/fca32cee-ef4c-4b50-82e5-4c768318a8c1/018f87c3-4d42-4bc0-9d66-53873b51ec68/episodes/audio/group/68464dea-c54e-4225-9b83-50cd29a37693/group-item/2b47129d-b3d9-4839-ac66-eacd5d0f3194/128_default_tc.mp3?aid=rss_feed&feed=3frRI0HN) · [Apple Podcasts](https://podcasts.apple.com/us/podcast/invisible-ink-ai-watermarking-synthid-the-internet/id1771688480?i=1000785703864&uo=4) · [Spotify (show)](https://open.spotify.com/show/3eFl4hY4t1qTCWE2Bxotrg) · [YouTube (video)](https://www.youtube.com/@ModemFutura)

---

**Andrew Maynard** (00:00): Okay, get in shot, scoot your head down so it fits the frame.

**Sean Leahy** (00:04): The most, again, always the humbling experience of getting this set up. I'm like, well, it doesn't fit half the time. So I was just, and of course, people couldn't hear this as I was setting up. I'm like, I'm the one who sets the camera up. And I'm like, why do I always make it so bad on myself?

**Andrew Maynard** (00:20): Yeah, yeah.

**Sean Leahy** (00:20): I don't know. Can't help it.

**Andrew Maynard** (00:22): But that's, you know, vanity.

**Sean Leahy** (00:23): It's very unpredictable. I don't know. An Easter egg for later.

(00:27) I don't know. Yeah. So, okay.

(00:31) I'm like, where do we pick up from all of the, I feel like we were on some burners last one.

**Andrew Maynard** (00:36): I know, but it's another week where part of me is thinking, can we talk about something other than AI? And the answer is no, because the train is going so fast and getting so close to getting off the rails. We can't not talk about it.

**Sean Leahy** (00:49): It's like a train that's like tipping from like one track to the other. It's like, it needs those, you know, like the amusement park rides that have those like, it needs the clamps. I don't know, because it is, it does feel like at any moment it's about to go off the rails. And so, yeah, so today I'm like, okay, how about we talk about, I even have, it's nice, I've been using Claude to help pull the show rundown together.

**Andrew Maynard** (01:13): And I love this meta aspect of this. We use Claude to create the show notes to then trash Claude.

**Sean Leahy** (01:20): Which I will talk about too, because it did not like that at all. But so I don't ask for this, but it titles the show. Which I'm like, hey, wait a second.

(01:30) That's my job. But I don't, we'll see if it sticks. But I like the way it just puts this together.

(01:34) It calls it, this episode is tentatively called Invisible Ink. I'm like, all right.

**Andrew Maynard** (01:40): Get that lemon juice out and start writing.

**Sean Leahy** (01:42): That's exactly it. Like lemon juice, milk. But yeah, so it's like, I love that Claude has entitled the show today.

(01:49) Invisible Ink. AI watermarking arise. Anthropic adopts Google's SynthID-Text.

(01:54) The EU AI Act forces the industry's hand and the internet melts down.

**Andrew Maynard** (02:00): Well said, I will give it that. I'm surprised that you didn't have a little bit at the end saying, do you want me to record the podcast for you? I wish. What tone would you like? Yes, exactly.

**Sean Leahy** (02:09): Well, that's something we can get into a little bit too because that is a ripe frustration. I came in this morning and I'm like, why is there no AI tool that can help me do the things that I do manually? Because that would genuinely be a massive time saver

**Andrew Maynard** (02:26): For me, of course, I think this is going to be part of the conversation. Sort of AI is fantastic for doing the things that it tells you it can do. It's lousy for doing the things that you actually wanted to do. A hundred percent.

**Sean Leahy** (02:37): And that's and again, we've talked about this before, but I think it's worth mentioning again, too. It's like, again, when you just are looking at those like the demo reels of like, oh, you want a website? No problem.

(02:46) Here's your website. You want a thing? Here's your thing.

(02:49) But when you actually are like, no, what I need is this exact thing because this is my real workflow and I need it to save me time. And it's like, nope.

**Andrew Maynard** (02:57): So, okay. So I've got to give you this anecdote, and then we must sort of get into the show. But the anecdote is, I think we talked about this. Did we talk about ASU's Atom AI generated?

**Sean Leahy** (03:10): We may have mentioned it briefly when it came out.

**Andrew Maynard** (03:13): So ASU came out with this platform. By surprise. Yes, but it slices and dices existing lectures.

(03:22) And if you subscribe to this, it's $5 a month, and I'm still paying $5 a month for this. But it has an AI sort of substrate or AI in the background that slices and dices these lectures to give you what you ask for. So you go in and say, I would like a training course on X, Y, and Z, and it asks you a load of questions, and then it gives you this.

(03:41) Apart from the fact that you can sort of say, say you go in and say, I want to learn about sort of responsible AI and the ethics. And it will ask you loads and loads of questions, and it'll say, here's a course on AI and business.

**Sean Leahy** (03:55): so no matter what you ask for you get what it's got oh my god i would like a course on fisheries and whatever like i've got a course on business and economics for you wait a minute that's not

**Andrew Maynard** (04:05): what i wanted so it's it's fantastic if you ask for exactly what it's got in the library that's

**Sean Leahy** (04:11): right it's like that thing it's like it's like the menu that has like four items and it's like yes or what was that what was the classic like henry ford thing you can get any color you want

**Andrew Maynard** (04:18): as long as it's black yes anyway with that we should get into yeah okay we'll stop there we'll

**Sean Leahy** (04:24): take a quick break and then we'll just we'll just jump in because i think this this is a this will be an interesting one because i think there myself included there are some strong feelings and reactions to this which aren't necessarily very warm and fuzzy yes um so we'll take a break and we'll jump right back in initializing waveform modulating signals for transmission

**Announcer** (04:49): exploring the possible probable and preferable futures

**Sean Leahy** (04:54): Welcome to Modem Futura. Oh my gosh, that giggle means... I know, we're back.

(05:17) This is Sean Leahy, joined by my co-host extraordinaire... Andrew Maynard. And that means you are listening to Modem Futura, the show that explores the intersection of technology, society, and the possible, probable, and preferable futures, asking deep questions like, what will it mean to be human in the future?

(05:32) And can you see your watermark? And can you tell? Have you been watermarked?

(05:36) Have you been inked? You know, so maybe that's a new thing. I have, but it's invisible.

(05:40) Yeah, you're like, you've been AI inked or something, you know? It's going to be this new fad. Or like, yeah, maybe you're not even going to have a choice.

(05:48) So, okay, so the big news around this is, of course, and I think one thing that's interesting is Anthropic, as of late, has been sort of top of the pile of news for quite a while with all these different things going on. And this is the latest sort of major announcement that as of August 2nd, 2026, due to the European Union, the AI Act, Article 50 transparency rules that was basically called for some sort of determinable, determinable watermark of sorts for text generated AI output. We've had watermarking or signatures, images and videos and things like that.

(06:31) But now this is a first mandated requirement that the text output carry with it some kind of signature. And we can talk about what that is because that alone is very weird. And there's a lot of confusion about how this is going to be done.

(06:49) How could you work around it? Right. But essentially, in compliance to this, Anthropic said, we will do this globally across the board.

(06:59) Other platforms like OpenAI and others are sort of, they haven't fully committed to what they're doing. But also I think there's probably a lot of internal hedging deciding if you have to do that regionally or not. And then of course that creates different, a lot of complexity.

(07:20) But in the end, what this means is that, and again, this is a fuzzy line too, because the language that Anthropic uses is that new models and retrofitted old models, but no timeline on is this current, is this happening right now?

**Andrew Maynard** (07:35): So there's a lot of uncertainty. A lot of uncertainty. Including how exactly they're doing this, because it's not just planting words. No. It's something more sophisticated than that.

**Sean Leahy** (07:42): Yes, and so, but like the rough sort of gist, if you're new to this sort of train of thought around, or this sort of news story that's unfolding around this, is that the output that you, if you use Anthropic's, OpenAI, or sorry. Claude. Claude. If you use Anthropic's Claude, that the text you use, someone would be able to easily basically identify that text as having been used with Claude.

**Andrew Maynard** (08:06): Yes.

**Sean Leahy** (08:07): And that's a, and there will, we can dig into that too, because there's a big distinction between that this is used by Claude versus Claude wrote this. Right. Which is, that's another thing.

(08:16) So basically there's this big, again, the internet did melt down, including, and Sean did too a little bit. Right. Just around this idea that all of a sudden, now any works that you do through Claude will contain essentially a watermark.

(08:31) And we can unpack that because it's not something so simple as like, and you know what a watermark, for people, if you're like, what's a watermark? That's that like little symbol or logo or otherwise stamp that's on material.

**Andrew Maynard** (08:42): Even there, so this is an irony. So when you think of a watermark, you think about something you can see. If you've got a sheet of paper with a watermark in, you hold it up and you can see the watermark and that says this is on official stationery or whatever, but you can see it.

(08:55) You can see it. Now, at the moment with Anthropic, you cannot see it. You get a bunch of text.

(09:01) You have no idea whether it's watermarked or not. So it's not quite the same. There is not that transparency.

**Sean Leahy** (09:05): No, and historically, like watermarking and putting digital identifiers and stuff, and there's a term, it's funny, I want to call it GUID, but it's not a graphical user interface. Right, okay. It's a global unified identifier, I think.

(09:21) So it's the same letters. They mean different things. And you know that if you've ever explored Microsoft Word files, if you've explored the metadata inside PDFs or images, there are lots of these types of metadata that is embedded to give provenance to where the file came from, who authored it.

(09:41) That is not new. The difference between that and a watermark is that is information that you can scrub out intentionally. You can remove metadata.

(09:49) Even some digital artifacts that you can't remove some of that stuff. So digital rights management elements, copyright information, the reason you can't share a movie you download on iTunes or whatever. Because that's embedded in.

(10:02) It's embedded. There's these restrictions. Those are also types of this idea, this concept of watermarking in a general sense of, again, labeling it, putting restrictions, putting some sort of provenance in it.

(10:14) So like if you take a digital file and move it around, you can say it came from Andrew. It came from this other person or whatever, whatever, whatever. But what's really interesting, of course, with artificial intelligence is you're able to crank out unfathomable amounts of text.

(10:31) And at present, a very difficult time discerning whether or not that was made by AI or a person or some combination. and how do you figure that out, which has led to this unending trail, I think, of false claims of AI detectors and, you know, all these kind of things.

**Andrew Maynard** (10:49): But it just, so this is where the landscape is so weird and murky. So it starts from asking, why do we need to tell in the first place? What is the rationale for being able to say this was AI generated or human generated?

(11:05) All the way through to how this is going to be used. And it's not just a case of sort of provenance in terms of accuracy, but there are some very, very soft social things here as well, with people saying, well, I don't like reading stuff written by AI. I want to have stuff written by a human, even if they can't tell the difference.

(11:22) And there have been studies that indicate that, in some cases, people prefer the AI written stuff if they don't know it was written by AI. So that question of, why are we doing this, is an important one. But then it gets all the way through to sort of authority and freedom of what you can do.

(11:42) And then what constitutes AI written versus not? So if you work with AI, you brainstorm it, and it throws you some ideas and use some, and you put your own words in, is that AI generated or not? Because that then keys into this.

**Sean Leahy** (11:53): Yeah, and so I think there's a lot of pieces. So I think when I was thinking about this, the initial, And we've had entire episodes devoted to things where we talk like the invisible upgrade, right? Right, yes. These things where the resonant frequency, if you will, between the human and the tool create this output.

**Andrew Maynard** (12:12): Right.

**Sean Leahy** (12:13): Who is anyone to say that is machine generated versus like slop content where you're like, make a book and I'm going to give you one sentence or one prompt. And you do it all. Yes, yep. And so you can see that, you know, depending, again, depending on people, and there are, again, the spectrum of feelings on AI-generated output go from everywhere from you should never even, even opening a browser that is connected to the internet is in some ways, you know, making you guilty versus people that are like everything and anything goes and whatever.

**Andrew Maynard** (12:47): And I've got to just throw in here because this is a thing that really irritates me. So we're having this conversation. How many apps now have the little buttons saying, do you want AI to generate this for you?

(12:57) If you're in email, if you're in a browser, if you're in LinkedIn, everywhere, they've got the AI button. So on one hand, you're being almost forced to use AI. On the other hand, there's this backlash against using AI.

(13:07) Well, absolutely.

**Sean Leahy** (13:08): And I think for me, one of the pieces that is almost an absurdity is like, it's like, it's too late, first of all. And because again, going back to, it's been a while since I've, so if you've got your bingo cards at home, players, I'm about to mention the cosmic background or the nuclear background. Not the concrete cement again, yes. And again, that just refers, go back and find that. I think it's a brilliant...

**Andrew Maynard** (13:31): If you have no idea what Sean is talking about, now is your time to go back into the archives.

**Sean Leahy** (13:35): It's the background radiation, which basically is a marker that, in this case as AI, anything after essentially late November of 2022... Is contaminated by AI. Is contaminated by AI. And so at this present time, and when I mean contaminated, I mean, you might not be using a tool directly, like do my homework.

**Andrew Maynard** (13:53): But AI has got his fingers on it.

**Sean Leahy** (13:54): Yeah, like you said, I go to email and it's like, you can't type one word without it trying to finish the sentence for you.

**Andrew Maynard** (13:59): Right, yes.

**Sean Leahy** (14:00): Which, you know, or you type in, you're searching something and you're getting the AI response, the Google, you know, there is almost nothing, especially in our digital world, that isn't already heavily influenced by AI, either directly or behind the scenes. Yep. So to say that the output now should be measured is whether it's clean from this or not.

(14:20) I'm like, how is that going to even happen? Right? And so that's one thing.

(14:26) And that kind of also comes back to this idea of like, why do you want to know? Why do we need to know? Because again, depending on your use case, you can see, well, right, if someone's generating like this idea, this false news or damaging material.

**Andrew Maynard** (14:43): But even, so if you're just looking at text, even false news. Humans are brilliant at doing that. We make up stuff all the time. We don't need AI to make up stories.

**Sean Leahy** (14:53): And now it's sort of like this thing where it's like, well, now because we have the technology, we should be able to use it to then to say definitively, I knew he was full of it. It's all fake. There's no way you could have, or like, I know you.

(15:05) There's no way you could have come up with something this clever. Like you must have been, like, I knew it. So there is, so I think there's, so there's that piece right off the top, which is more of like the social sort of psychology of it is of why.

(15:20) Why do we need to know? Should anyone be able to know? Because again, I think that is a legitimate question.

(15:28) As these things are changing how you think, how you act, how you do, again, depending on what level of output is mixed in there, because what this doesn't do is delegate any level of like threshold, right? And I can talk a little bit more about some of the tech behind this, because the way that it's using a Google tool to do this, which is interesting, it's using something that came out of DeepMind. It's called SynthID.

**Andrew Maynard** (15:53): I note, so if you go back a little over 12 months ago with the World Economic Forum, Top 10 Emerging Technologies,

**Sean Leahy** (15:59): we wrote about it last year.

**Andrew Maynard** (16:00): Yep, absolutely. It was one of our top 10.

**Sean Leahy** (16:01): It's one of the top 10. And yeah, I mean, again, ringer, ringer. Look at that.

**Andrew Maynard** (16:05): We were there.

**Sean Leahy** (16:07): Now, when do we get the microbreweries in our guts? That's what I want to know. That one can come true.

**Andrew Maynard** (16:12): I'd rather that one. That's another archive set. You know, or like, can I just, yeah,

**Sean Leahy** (16:16): I'd like to brew tea on my, I don't know. But so, so yeah, so, so you have this SynthID capability of watermarking the text. And in this particular case, the watermarking is not a visible stamp.

(16:28) It is not a digital metadata that's embedded. It is a, essentially a secret code that is used based on the probability of the way the LLMs work, right? So a really very simplistic example is if, and the idea here is that the hidden sort of decoder sheet is kept secret to, in this case, Anthropic.

(16:51) And the way it works is essentially if you're saying something simple, like write an email to my colleague saying, you know, so long, thanks for all the fish. It's kind of a thing or something, right? The AI comes back and it does this.

(17:03) And the way the AI works, if we roll back again to the simplicity, the probability engine on what's the next character, letter, whatever in the string. Well, it has options. So one easy example is like at the end of a sentence, instead of saying like, this has been exciting, it could be this has been fun.

(17:21) This has been amazing.

**Andrew Maynard** (17:22): So you have all these different word choices that different probabilities that it could choose.

**Sean Leahy** (17:26): And so the statistical probability of using either one of those could be equal. OK, fine. And so as the user, you would read that sentence and not think anything of where or why it chose that word over the other word.

(17:40) Right. But what you don't see is the secret sauce on the backside is actually telling it which one to use. Yes.

(17:47) Based on a probability, again, a secret.

**Andrew Maynard** (17:48): Because that then embeds a frequency of use.

**Sean Leahy** (17:52): Yep. In a secret layer so that through the text of your document, the certain appearance, not of the word itself, but of the probabilistic choices that it made. Yes. appear in a way that can say, this was our handiwork.

**Andrew Maynard** (18:05): So I've got to say here, this sort of bends my mind, boggles my mind, because the origins of large language models were predicting those next words and sentences from humans. So it trained on what humans do and said, well, this is the most probabilistic way that a human would write this. Now we're saying, well, actually, let's flip that and say, what's the most probabilistic way that an AI would write it?

(18:28) So now the AIs are in charge of how they write. Absolutely. And it's no longer sort of emulating humans, but it's emulating AIs pretending to be humans.

**Sean Leahy** (18:38): Yeah, right. It's emulating the humans with a little dash of like, but this is our secret sauce.

**Andrew Maynard** (18:41): Right, right. But also as a writer, I'm thinking about this. Yeah, that might make sense to somebody that doesn't write for a living.

(18:47) But if you're a writer, every word has meaning and weight. And you weigh each word depending on what you're trying to say. So you may think that that last word is trivial and you can replace it with something.

(18:59) that fits into this frequency chart. But as a writer, it doesn't work that way.

**Sean Leahy** (19:03): No. And in fact, what it is, okay, so then to go even deeper with it, right? So the other piece around it is, well, twofold. One is that unlike other watermarks, so again, going back in the, you know, 30 plus years of Photoshop, for example, right?

(19:20) When people are like, oh, don't worry, we've watermarked that image. And it's like, okay, that'll take 10 seconds to remove, right? And you go in and you're like, boom, watermark gone.

(19:27) Or it's even as simple as people would just crop and frame it out, right? I know, yes. So it's a cat and mouse game with watermarks.

(19:33) What's interesting about this, because the watermark itself is in the probabilistic nature of those words appearing, it survives things like copy and paste. It survives editing to an extent.

**Andrew Maynard** (19:46): It's more fractal-like. And it's a dangerous analogy, but I like fractals. Or actually, more hologram-like is even better.

(19:54) So if you know anything about holograms, the really cool thing about a hologram is you can have a sheet of glass with a hologram on and you shatter it. And every little individual bit of that shattered glass will give you a glimpse of the original. So it's a little bit more like that where even if you sort of break up the text, it still carries something of that original origin.

**Sean Leahy** (20:14): Yeah, yeah, absolutely. And so the challenge is normally, right? Like if you're like, okay, I got AI content.

(20:20) You're like, well, I'll just copy and paste that into a Word document or I'll run it through another AI tool or whatever, right? Like now write it like a fifth grader wrote it and then now write it, you know. And, you know, you can do all these things to kind of move them back and forth.

(20:35) And, but if the words, and so right now, as it is, significant editing might surely break the watermark. Yes. But that's going to entail you having to go through it.

(20:47) And again, going back, you might as well just written the dang thing. I know, yes. But also, and it was funny because I've heard some interesting, some funny anecdotes.

(20:54) So people would be like, well, just take it in English and convert it into a different language. Then back, and then it was like, you know, it was someone had brought this funny thing, you know, and I don't even know if this is something that had been done before, if it was just this one time anecdote, but it was like, you know, this idea of this phrase of like, you know, the soul was willing, but the flesh was weak. And like, turn that, like convert that to Russian.

(21:15) It comes back. Then it was like, the vodka's good. The meat is bad.

(21:20) And then you're like, okay, well, you know,

**Andrew Maynard** (21:22): Which is a whole other layer of AI detection.

**Sean Leahy** (21:25): A whole other layer. But it was funny too, because like the other piece about it is then kind of back to your point about the AI now taking its own sort of mechanicalistic or statistical probabilistic influence on the word choice, right? Can have dramatic impact in what the word says. Because one of the examples when I was looking at, you know, again, working with Claude to try to explain Claude's rules, which was, that was an interesting experiment in its own right. But one of the examples it gave, the word, I remember it very clearly, the word at the end of the sentence was one was significant

**Andrew Maynard** (22:00): and one was meaningful, right?

**Sean Leahy** (22:02): So in normal, I would say, normal everyday language, you might see those as synonyms. However, in academic language

**Andrew Maynard** (22:10): In technical language, absolutely not.

**Sean Leahy** (22:11): There is a big, when you, and if you know this, if you've ever passed a paper to a professor with the word significant

**Andrew Maynard** (22:18): Well, especially a stats professor. Yeah, no, no. Significant has a very specific weight. Substantial is actually an alternative that I would use when significant actually doesn't have the weight. That's right.

**Sean Leahy** (22:28): And it might have the generic everyday language synonym, but not what it means. Like when you say significant, you're like, okay, well then where is the proof? Where are the numbers?

(22:37) Yes. That is a claim you are making. So it was interesting.

(22:41) So even in there, you're like, okay, so that's really interesting where you would have potentially these really uncomfortable traps where the language that it might be trying to force you to use or suggesting as of like workflow edits. So again, I'm not even saying whole cloth. I'm saying you've written work and you said, revise my, help me make this, what was the metric for a better paper?

(23:06) It was the word scarcity or what was the metric it said? It was basically like the, basically you talk too much.

**Andrew Maynard** (23:15): Oh yes, yeah, yeah. No, I, oh goodness. Oh, what was that? Oh, I can't remember.

(23:18) Well, yeah. But yes, basically it's Word economy or something? It was a lack of word economy.

(23:23) I wish I could remember the exact phrase it used, yes.

**Sean Leahy** (23:26): But so now, so if you say, so you've got, let's say you have a five-page thing, and you're like, I want help reducing this to one and a half pages or something. So now what really comes into question is the word choice it might use might pull you away from what you're actually meaning to do because of the intent to water market. which then is like in this really awful scenario where it's like, it's no longer potentially doing what you want it to do.

**Andrew Maynard** (23:54): It's doing what the company wants it to do.

**Sean Leahy** (23:55): It's doing what it's been told to do, which now is, so then it raises a question like, you know, and this may have been a long, a barrier of the lead to get to this question is like, should we even be using these tools anymore? Yeah, yeah.

**Andrew Maynard** (24:06): Right? And this is where I really struggle and it comes back to sort of what is the point of this. And so I actually, I do not mind people using AI.

(24:15) I have visceral responses to bad AI use with writing. But I would always concede that if you're using it smartly to get an idea across, and AI helps you get that idea across better, that's great. But if we're beginning to A, nitpick here, and B, if the machine is beginning to sort of guide and steer exactly how you express yourself, so that that writing is no longer just about getting the idea across, but it's got other embedded things in it, That worries me a lot.

**Sean Leahy** (24:46): Absolutely. And I think, you know, and again, you have this weird, this weird slash troubling connection between the human evaluation. Because again, I mean, you can hear, you hear stories left, right, and center of these people having, you know, getting disciplinary warnings in schools about you cheated because it triggered some BS trigger.

**Andrew Maynard** (25:06): You know, it's a career killer. And we're already beginning to see this where somebody slips up and uses AI even benignly and somebody else publicly accuses them of that,

**Sean Leahy** (25:16): it gets messy. Well, and that's one of the pieces. So one of the main backlash elements here is it almost creates this scarlet letter where it's like, oh, well, if you used it here, then we must call into question everything that's, and it has this, you can see, and again, the question's why?

(25:34) Why would that matter? And even though it is inescapable, there is still a very strong feelings about whether how you should use, how you should use it professionally, especially if your profession is thinking and writing and doing these other things. There are very strong feelings about it that, again, runs that spectrum.

(25:54) But if you, again, it creates this, and I think there's been this since the beginning for people who have been sort of power users or really leveraging it to reach that new level of ways of creating work have had to do it sometimes in secret. Having to do, because there is a stigma There is a negative association. If you're like, well, I use AI for my whatever, then it's like, oh, well.

**Andrew Maynard** (26:21): And I think that we see this happening, but there is another aspect of this. So yes, Anthropic doing this, but then you also have the AI checkers. And I have to say, so Pangram is the leader of the pack at the moment.

(26:33) And with all of the caveats around, you really sort of can't trust these 100%. Pangram is incredibly good. I mean, I have been bowled over with how good it is, which basically means forget about the watermark.

(26:48) Already, you have employers and companies and others putting stuff through Pangram, and they can say with reasonable accuracy sort of the degree to which you are likely to have used AI. Now, I would, again, caution to say anybody who tries to use that either in a disciplinary way or a shaming way is on incredibly shaky grounds. But the reality is you cannot escape this.

(27:11) If you're using AI in your writing, somebody will be able to tell.

**Sean Leahy** (27:15): Yeah. Well, it's interesting too. I don't know enough about Pangram. Do you know how it detects, like what's different about it that

**Andrew Maynard** (27:22): Again, it uses a similar approach in terms of looking at statistical frequency.

**Sean Leahy** (27:25): So it basically You have 22.4 em dashes. Right.

**Andrew Maynard** (27:30): Well, no, so it doesn't look at the actual content. So forget about things like sort of patterns and em dashes and things. It looks at sort of basically, if it was an AI predicting the next sort of word it can tell whether what it's reading is more likely to sort of follow that predictive pattern versus not.

**Sean Leahy** (27:47): Interesting.

**Andrew Maynard** (27:48): And it is incredibly good.

**Sean Leahy** (27:50): Yeah, because I have not played with it. I have not played with it.

**Andrew Maynard** (27:53): It's worthwhile giving it a go.

**Sean Leahy** (27:55): Yeah, I have to play with it because initially, my initial gut reaction is to write them all off because they have historically been so, I mean, we're like, what was it?

**Andrew Maynard** (28:03): This is in a different category.

**Sean Leahy** (28:03): I forget which tool it was, but one of them, you know, the Declaration of Independence was like, 48% written by, and you're like, Well, it's because, okay, come on.

**Andrew Maynard** (28:11): And again, even with Pangram, you have these stories of people coming out and saying, well, you know, I write like AI and I have done for the last 40 years. So it thinks that I'm an AI. But I think that those are edge cases, to be honest.

**Sean Leahy** (28:22): And to some extent, we all do. Because if you come back to it, again, I go back to Jaron Lanier's thing, it's like, there is no AI, it's just our data. Right, right.

**Andrew Maynard** (28:30): Until now with the watermarking.

**Sean Leahy** (28:31): Well, it was interesting. So I was just looking to one of the, sort of a quote that was pulled up from John Gruber, kind of talking about this concept that the watermarking trades precise word choice with detectability, and basically calling that a corruption of writing, which is interesting, right?

**Andrew Maynard** (28:47): I would agree with that. Again, if you're a writer, if that is your craft.

**Sean Leahy** (28:51): Yeah, and it's really, and again, I come back to this issue of it's really interesting. And, you know, I think even before this tool has been put into place, we've already seen lots of behavior trying to like different, like again, I feel bad for people who were in like writing careers, journalistic careers who used the em dash. Like the em dash has existed for, I don't even know, for how many hundreds of years have the em dash existed?

**Andrew Maynard** (29:18): If somebody wants to accuse me of using AI, that's fine. I use em dashes.

**Sean Leahy** (29:23): That's how I write. I almost looked that up. How long has the em dash been part of the

**Andrew Maynard** (29:27): It's been a while, yes. But actually, but nobody talks about, do you use the American or the British version of it? One, you put spaces either side, the other you don't.

**Sean Leahy** (29:35): Oh, I put spaces. Which one is that? Oh, okay.

**Andrew Maynard** (29:38): I'm sorry, no, that's the British one. I think the American sort of formal approach is no spaces.

**Sean Leahy** (29:43): Oh, okay, yeah. Mine always has spaces because for me, I'm looking at my keyboard, it's option dash. Which for the longest time too, I mean, maybe I shouldn't admit this, I didn't realize there was a thing called an em dash. For the longest time, I'm just like, it's just a dash. I didn't even know that it had like, there was a difference between like an em dash and all this.

**Andrew Maynard** (30:00): So you were probably one of these people that mixed up em dashes and en dashes and just regular dashes.

**Sean Leahy** (30:06): I think I just use regular dashes. And I'm like, oh, that's embarrassing now. And it was funny because to that point, I saw, maybe it was on LinkedIn or something, somebody had posed, you know, how sometimes you see people will share these pictures of themselves at a talk or something, right? And this particular one on the slide was a citation of all of the em dashes, and he counted them. it was like all 48 em dashes were put in by me period and i'm just like see right but but it's interesting so even even you know without tools that are trying to like suss out whether or not there's been a usage of it you know we've already engaged in this behavior we've talked about this a little bit before of even just just consciously as you're writing a paragraph you are altering the word choice that you use to avoid yes to avoid that that scarlet letter of being like oh i don't

**Andrew Maynard** (30:55): even want someone to accuse this of writing. And actually, it gets even worse, because you have those people that will sort of use AI either to grammar check or do a bit of background research or something. So then they put a piece out, which is written by them, and somebody accuses them of using AI, and they're stuck between a rock and a hard place, because they can't claim absolutely no AI came anywhere near this, because they might have just checked something in sort of ChatGPT or something. And yet it didn't substantially affect the writing that came out. And what's really,

**Sean Leahy** (31:24): I mean, again, this is kind of like a cool feature that's also dangerous. But like, again, that word choice, that would impact too, is even if you just had it written and then you wrote it by hand, that wouldn't change. It would still, so you're like, whoa, now it's that like cryptography of it is like jumping mediums too, which is cool.

(31:43) But that's, so I want to bring this up too, because this is another piece. I get annoyed and frustrated by this because my thing is like, why do we need this? I mean, I guess, you know, if you think about it, if I put my, try to view this perspective from other disciplines, if you're in like news and you're trying to like fact check things and you're, you know, you're concerned that if I grab a quote from Andrew or that you used it and so then I'm passing it off.

(32:12) Now I'm guilty of using it even, you know, so like I understand in certain situations, and again, apparently too, I've, I've heard a lot of stories recently of how much AI generated content is making its way into like congressional staffers. And so it's starting to show up in places like bill legislation and stuff. So I understand from that perspective, but again, it's like whether it was

**Andrew Maynard** (32:33): written by an AI or a human, if it's checked and it's good. It's the content that should matter.

**Sean Leahy** (32:39): Does it matter who typed it in that sense? Right. So like, I have a lot of issues with it,

**Andrew Maynard** (32:44): but the other piece- Did you use a number three quill? I think this is actually the argument ad absurdum. Yeah. Sort of, if you didn't use the number three quill, it is not legitimate.

**Sean Leahy** (32:53): That's right. You got to use it. Did you, and I'm sorry, was that pencil HB or HB2?

(32:58) I don't understand. Yes. You know, but the other piece too, and one of the things that this reminds me of as well, because again, looking at the, what it can do right now is it can say with some degree of certainty that this content has come out of Claude.

**Andrew Maynard** (33:13): Yes.

**Sean Leahy** (33:13): Right. What it doesn't say is whether that was your content that Claude remixed, edited, touched.

**Andrew Maynard** (33:19): Whether it was original or I.

**Sean Leahy** (33:20): Or if it was slop, right? There's no way it can tell that. So the problem is, too, and this was happening, I'm trying to remember, I think it was, I'm trying to remember the platform.

(33:29) It must have been Instagram, I think, who was doing this. When you would upload an image that had, again, a metadata marker that had come out of Photoshop, it would say, this image has been manipulated by artificial intelligence because Photoshop contains a lot of artificial intelligence tools. And so it's like, well, wait a second.

(33:47) all that person did was crop the photo. Right. Or something, right?

(33:51) Like do a generic, socially acceptable, not even pixel pushing, which is where you change what, like literally just doing normal sort of image processing techniques onto an image was being flagged as AI generated. And there's an interesting story. Again, somebody had posted pictures of like their engagement and it was like, these photos are AI.

(34:13) And the person was like, no, how could this be? Right. But so it makes me think a lot of that as well, which is like, it's just going to then blanket everything.

(34:23) And I mean, I would be really curious because at this point, again, thinking about how you can't escape this technology in the digital world, I'm like, where do you draw the line?

**Andrew Maynard** (34:35): We've got to reframe the conversation. And this has resonances of the conversation that we've been having in education for the last four years. So if you think about that, ChatGPT comes out, the immediate sort of response from educators is, how do we tell when a student is using it to cheat?

(34:51) Homework is dead. Yeah. Existential.

**Sean Leahy** (34:54): Education's over.

**Andrew Maynard** (34:54): Sadly, that is still there. But many educators have moved on and said, no, what this means is we've got to go back to basics and ask, well, what are we trying to do with our students? What does learning actually mean here?

(35:06) And how do we actually get back to the core of what we're trying to sort of instill in our students? But now it feels that in public, we're back four years. where the question is, are you cheating or are you not cheating?

(35:19) Rather than what are you trying to achieve?

**Sean Leahy** (35:21): Exactly. And I think that's, you know, we start getting at these deeper issues, which all circle around, again, why? Why do we care?

(35:28) Like, again, if the output is good. Now, again, right? Of course, there are nuances and challenges.

**Andrew Maynard** (35:36): But I'll give you an example here. So say you have a grad student who has spent three years deeply researching something. They know their subject inside and out.

(35:46) And they have a really clear idea in their own head of the new knowledge they're generating. But they're really struggling to get that on paper in a way that other people will understand what they're doing. If they use AI to take the ideas out of their head, all originally generated from them, and into a format where somebody else can read that and it resonates with them and they understand, is that good or is that bad?

(36:10) Yeah. Or are we solely judging the scholarship of that person on their ability to communicate? Nothing to do with their actual scholarship.

**Sean Leahy** (36:17): Yeah, no, and again, it starts to look at, again, that's why I like the use of that, like, scarlet letter pieces, 'cause it associates this, like, dirtiness to the work that you're doing, regardless of what or how the work, like, you know, and again, drawing this line, I literally just was typing, how long has the em, and then Google finished the rest of it for me predictive type, right? Turns out it's been over 500 years. But again, that was AI. That's the AI overview. I just used AI. So now if I type, the em dash has been around for 500 years.

(36:51) You're tainted. By all accounts, that is not my information. That just came from these tools and systems.

**Andrew Maynard** (36:57): See, that's right. And that's where we've got to be really careful. Now, there is a counter pushback to that, which I think is really important. That's the question of care. And this is where the judgmental side of me comes in.

(37:12) And I would say this is something that I do not have resolution on. But when I read something that somebody has put together and it feels like it's AI generated, my first thought is, how much care did they take in writing this? Do they expect me to put more time and effort into reading it than they put into writing it?

(37:29) So then it's not about the content, but it's about, does this person care enough about me as a reader to actually put sweat equity into it? But that is very different from did they or didn't they? It's how did they actually construct this piece that I'm reading?

(37:45) And did they construct it with enough care for me to care about reading it?

**Sean Leahy** (37:49): Yeah, absolutely. I mean, you can, again, I even think about this, like, I can be childish about this. And also be like, it's okay, well, then why not disclose what type of keyboard you use?

(38:00) It's the number three quill. What keyboard did you use? What computer did you use?

(38:04) Did you use spellcheck on that or not? because if you spell check, you clearly don't know what you're doing, right? And so you're like, well, okay, where does that line stop?

**Andrew Maynard** (38:14): See, that's right. That's where we've got to completely reframe the conversation away from did they, didn't they use AI to sort of what is the process and what is the intent?

**Sean Leahy** (38:25): And what value do you get out of it as a reader? Absolutely. And it comes back to this concept of that, like we need to know the provenance of this.

(38:32) And I'm like, okay, well then where did it come from, right? Well, and we try to do that. We cite our sources.

(38:37) We do this kind of stuff. But like, where does it end?

**Andrew Maynard** (38:39): So again, you have, so this is actually an example I use from my risk research. Imagine you have a cliff, a 400 foot cliff, and you have a sign there saying, beware the cliff. And somebody walks along and says, you know, I'm not going to listen to that.

(38:53) That was written by AI. The point is not who wrote the sign. The point is the consequences of ignoring it.

**Sean Leahy** (38:58): Clearly this went through Claude. I want to see, prove it to me.

**Andrew Maynard** (39:02): The last thought that went through their head. i was went through now that is a cartoon i knew it

**Sean Leahy** (39:11): yeah but it's it's wild to me and then so like i and i i can't help this part too and again this is not a this is where some of the the futures improv part of my brain kicks in too is you know we've had these and again every time this conversation comes up we quickly will hear references back to like well we have to you know and we've talked about this at length and on our shows too about what is gained and what is lost. And, you know, there is a threat where, again, if you don't have the understanding of how a sentence is formed, eventually you just go, I guess it's whatever I get from the machine is what it is. And you're already seeing that.

**Andrew Maynard** (39:46): You're already seeing people beginning to write like an AI because they cannot discern what is different about AI speak and more than anything. And I, so we're drifting now into another episode, so I'm going to have to pull back from this. But the more advanced these models get, the more I find it very hard to stomach how they actually put sentences and paragraphs together.

**Sean Leahy** (40:08): Yeah, and I think, you know, maybe a tee up. I can tease us into another episode, too. Because, so the future, again, the future's improv of my brain is starting to tickle.

(40:16) And it goes, okay, why do we need to write? I mean, okay, right, like, now I realize that that's going to get people off their chair or whatever. But like, okay, so if we think about it, right, like is, and I think this is a fun place to explore.

(40:33) And so this, this could be homework for those of you who want to play along at home.

**Andrew Maynard** (40:37): Just give me a minute while I go away meditating.

**Sean Leahy** (40:40): Find my calm again. Yeah, that's right. You go like, so, so the question you can, this is a provocative question to put out there is, okay, so if these technological, if as a species, we have developed the technology to basically cognitively offload all of this. this thing called reading and writing because if we if it's ai written and it goes to something that it's like i don't have time to read this let me just get a let me just feed it to the thing so you end up essentially just taking a sidestep and letting it and the question of course right away you're like well this can't this this this can't this is a you know this is an abomination right um but then the question is why but then you're like yes is it yes because again holding reality putting a pause on that and getting to step outside of that, asking some of these really audacious questions like, well, maybe as a human species, we weren't meant to read and write this way forever. That maybe we get to, now you can call it a backward step or a forward step.

**Andrew Maynard** (41:37): But I think this is where it gets interesting. So I've got to say, so my first reaction is absolute nonsense. I can give you 10 reasons straight away why writing is so important.

(41:49) And then the second thought is, but all I'm doing is justifying the world as it is at the moment. And that is the difficulty. It's almost impossible to extract ourselves from our current situation and see things differently.

**Sean Leahy** (42:01): Yeah, and again, you could say, yes, maybe through evidence, you know, the human species has always found ways to communicate through multitude of ways, right? Written being one of them. But it's not something you're born with. It's not a natural thing. And even right now, we can't, as a global system,

**Andrew Maynard** (42:19): society. We can't agree on one language. No, no, no. And actually, you look at human history. So you could argue that certainly in Western society, and actually to a degree in Eastern society, the ability to write and communicate that way was a transformative point. It was a tipping point.

(42:35) And yet, we had societies, we still have societies that don't use the written word. We had societies that thrived before that. So yes, you can say with a lot of the world, it was a tipping point. But what is the next tipping point? Just because there was one doesn't mean there isn't going to be another.

(42:50) And maybe AI and getting rid of the writing is the next tipping point. It scares me. Right. Well,

**Sean Leahy** (42:56): Same. But again, this is where it gets fun as you're like, whoa, okay, let's, it's like you get to step off that cliff, but you don't fall, right?

**Andrew Maynard** (43:04): And the reality is, again, speaking as a writer, people don't read. People use AI to churn out words, but I know for a fact writing a Substack, people do not read.

**Sean Leahy** (43:11): Oh, I know. Well, and this is something, I mean, and this is not a new thing.

(43:15) that's been an issue growing for decades. A lot of controversy around the use of digital technologies and what it's doing to the ability to not only write, but also to read. But again, I come back to this wild thought of like, well, maybe there's something better out there because at some point, maybe we don't need that to do what it has done, which is to be this bedrock of our society.

(43:40) Maybe there is another thing on the horizon that we have yet to go across.

**Andrew Maynard** (43:44): I would say, so I've got to go back to my 10 things straight off the bat. But I think it's interesting. So you just glance at the bookshelf behind you.

(43:55) And let me challenge you in this way. So your ability to read means that you can scan that in half a second and get something out of it. What would that scanning look like in a world where you can't read with AI?

**Sean Leahy** (44:07): Well, maybe it's visual or it's happening in an audible fashion or something. I don't know.

**Andrew Maynard** (44:13): But the interesting thing is, this is parallel processing. You're not so linearly going through each book spine and going down. You're getting that snapshot and your brain is sort of capturing this 2D image and extrapolating from it.

**Sean Leahy** (44:25): How would that work? I don't know. Well, the funny is, well, the hard part too is I can't detangle my ability to read from that.

**Andrew Maynard** (44:31): See, that is. This is an AI question. You're an AI sort of, and you can't read.

(44:36) How would you actually do this? How would you look at it? Which, of course, but actually the ridiculous thing is an AI can't read.

(44:42) It gives the impression of reading. It's all statistics. But if you took a snapshot of that bookshelf and gave it to an AI, the way that it would assess it is not linearly reading it.

(44:52) No, right?

**Sean Leahy** (44:53): And it looks at everything simultaneously. So the beginning is the end and the end is the beginning. And what is it all for anyway? Why does it matter? Which means actually the future

**Andrew Maynard** (45:01): is a future of non-reading AIs. Humans won't get there. This is our succession.

**Sean Leahy** (45:05): Yeah, well, you know, I can't help but think of the movie Arrival. I'm like, this is how they talk in squid circles.

**Andrew Maynard** (45:11): There we are.

**Sean Leahy** (45:12): Right? That's what you need. It's no longer reading a book. It's just you magically do a little smoky ink circle and you're like, oh, that's the future. That's right.

**Andrew Maynard** (45:20): Next week, if anybody's watching the video, it's just going to be squid circles.

**Sean Leahy** (45:23): Just circles. You know what? Let's just do it this way because, and maybe, you know, hey, who's to say that we, that the written word is the end all be all way of communicating. Maybe there is some unknown symbology of shapes, colors, whatever that unlocks that message in your mind that allows you to see the future.

**Andrew Maynard** (45:44): What a nice futures improv way to wrap this up. But then we'll have to have the conversation about how do you put watermarks in squid circles?

**Sean Leahy** (45:50): Yeah, I'm sure there's a way. See, in that one, you also can convey the past, present, and future in their world. So therefore, not only can you, again, talk about a used future.

**Andrew Maynard** (46:02): It's like Is the AI deciding which future is going to go into this particular

**Sean Leahy** (46:06): You now have seen the future. Oh, okay.

**Andrew Maynard** (46:11): Well, so I'm not quite sure where that leaves us, but I'm sure the controversy is going to continue about watermarks in writing.

**Sean Leahy** (46:17): Yeah, watermarks. We can close the book on this one for now, but this is going to be an interesting thing because, again, from what's unclear is when these are actually enacted. Again, I had a conversation with Claude itself, and it didn't know about it.

(46:34) I had to give it the announcement from Claude itself, and I said, oh, yeah, you're right. Okay, yep, I got that. But still no clear pathway in terms of, like, when is this actually, when is it working?

(46:44) Is it now? And who actually gets to see?

**Andrew Maynard** (46:47): Do we all get access to the deconvoluting algorithm

**Sean Leahy** (46:50): that tells us whether something was from cool? That's a good question. That has been nowhere mentioned, is like other than Anthropic themselves, who gets access to that? Like what, yeah, who gets to tattle on you that says, well, actually, yes, you used 25%.

**Andrew Maynard** (47:03): Yeah, but I think we ought to wrap up and come back to sort of teaching and learning and education because this is where a lot of these conversations been happening over the last four years. And I think that people ought to pay more attention to that and ask, well, how do you move forward in a world where everyone at the press of a button can use AI? How do you reorient your thinking here rather than try and put restrictions in place to try and keep things in a pre-AI world? Yep, absolutely. And I think I'll go back and I'll

**Sean Leahy** (47:31): just end with this as a call. We've had some episodes in the past where we've talked about this, but I think the conversation continues to evolve. But also thinking back to conversations we've had with Punya Mishra on here, which is really what we should not be trying to do is change education for AI, but change education in a world that is completely dominated by artificial intelligence.

**Andrew Maynard** (47:53): With human flourishing at the center of it.

**Sean Leahy** (47:55): Slight change of language, but massive change in what you're actually trying to do. So, but yeah, so yeah. So, you know, again, you know, hey, give us, throw us a like, a follow, a comment, whatever.

(48:06) Let us know what, did we break a brain? Do we break your brain? Do we anger you?

(48:11) Do we inspire you? I don't know. Let us know.

(48:14) It's all fun and all fair game. So, and of course, if you don't like following through those traditional mechanisms, you can always go to our website.

**Andrew Maynard** (48:22): modemfutura.com slash sign up, and you can just sign up for a just quick email every week that a new episode drops.

**Sean Leahy** (48:29): That's right. And yeah, so we'll, like I said, we'll shut the, we'll shut the book. We'll put it back on the shelf.

**Andrew Maynard** (48:34): Maybe forget about it. I don't know. We'll find a few more used cliches and metaphors.

**Sean Leahy** (48:40): You can certainly look forward to those. But yeah, until then, thanks everyone. And we'll see you on the next episode of Modem Futura.

**Andrew Maynard** (48:46): We'll see you then.

**Sean Leahy** (48:48): Demodulating signal. Modem Futura is a production of the Future of Being Human Initiative at Arizona State University. Be sure to subscribe on Apple Podcasts, Spotify, or wherever you listen to your favorite shows. To learn more about the Future of Being Human Initiative in all of our other projects, please visit us on the web at futureofbeinghuman.asu.edu.

**Announcer** (49:21): End transmission.

---

*This is the full transcript of an episode of Modem Futura (<https://www.modemfutura.com>), a podcast hosted by Sean Leahy and Andrew Maynard and part of the ASU Future of Being Human initiative (<https://futureofbeinghuman.asu.edu/>). The episode’s show notes and listening links are at <https://text.futureofbeinghuman.com/modemfutura/99-invisible-ink-ai-watermarking-synthid-the-internet-meltdown.html>. Listen via the links above, or watch it on YouTube (<https://www.youtube.com/@ModemFutura>).*