S2. Ep32 - Anthropic Is Watermarking AI Text — And You Can't Strip It Out

Katie: Welcome to the AI Automations for Business Podcasts. We're your hosts, Katie and Null, and we'll be discussing how you can use AI for your business along with the latest news, updates and automations to help you stay ahead of the curve, allowing you to grow and scale your business more efficiently. Please be sure to subscribe and I hope you enjoy this episode. Hello, welcome back to another episode. Hi, hello. I'm Katie. And as always, I've got Noel here with me today. Hi, Noel, how are you doing?

Noel: Yeah, I'm doing great. How are you doing this week?

Katie: Yeah, all good, thank you.

Noel: Good, good.

Katie: So, last week, we were like, there's not much news going on.

Noel: And we're like, it's the summer,

Katie: you know, everyone's kind of like on holiday, chill and I. And then this week, it's kind of like everything's changed, as it does in the AI and Automation's world.

Noel: It does, yeah. It's been a bit of a bit of a crazy week. I mean, some of these have maybe been rumbling off for just a little more than a week.

Katie: But yeah. Yeah.

Noel: Yeah, the quiet time has disappeared.

Katie: Yeah.

Noel: It's ended.

Katie: Everyone's back to it.

Noel: Yeah, exactly. Yeah, they've had a week off at the start of the holidays and that's it. Yeah. Come back. Yeah.

Katie: Okay, so usually in this section of the podcast, Noel and I just go through the news, any updates that are relevant to people using AI in automations within their business. But this week there are so many updates and news that we want to talk about and share. in more detail that we are dedicating this whole episode to current news. So this podcast is going out on the 20th of August, 2006. So just in case you are listening in the future, then this is when this podcast was recorded. Yep.

Noel: It's probably all changed. If you're listening in 2027, then it's all changed. Yeah,

Katie: just get to the next episode or skip back to an episode. Yeah. Okay, Noel, where do you want to start?

Noel: So I better start with our friends at make.com. They've had a big, big release this week, or a couple of weeks ago, I think now, but they have released Maya, which is their AI agent assistant within make.com. And, yeah, we've been pretty excited to see the release. Obviously, I've had access to Maya for, it could be about seven months, I think, now, when they first released it to close beta. But it's come on so much since their first release. So, yeah, really, really happy to see it been released.

Katie: Because we were actually told of Maya at the Mix.com conference last year, last October, when they do the big one-day conference. And last year it was in Munich, this year it's in Prague. But, yeah, one of the first things they showed us with make.com, you know, upcoming updates was, of course, Maya. And everyone in the room was very excited. And there were audio gags and all that.

Noel: Yeah, there was.

Katie: because I think what is exciting about this is it becomes less about tech and just more about communication. So if you can communicate, you can use this. You don't have to be techy. You don't, you know, have to know exactly how everything works. It's literally like you follow the steps. You tell it what you want to create or what you want to happen. And it doesn't. it?

Noel: Yes. It's really awesome how it works. I think yes, you could just let it do its thing and create your automations for you because that's kind of the crux of it is you would go to it with an automation problem and say, well, I need to get sales data from this system, do something with it and then output it somewhere else. You know, you would kind of go to it with that sort of problem and then it would then go off and then create that automation for you. But what I think is also really cool as well is it's not just doing it. It's also showing you how it's done. So everything is done step by step. So if you're a complete beginner, this is a bit of a teaching tool as well as something that's really helpful and can get you up and running in five minutes. You know, it will take you through every step. You'll say, well, we need to connect to, you know, let's say Google Sheets. It's like, right, you need to go off and do stuff. This is what you need to get in order to create that connection to Google. You know, it can then go, well, now you need to create that connection to that AI provider. And it will take you through all of that for you. It will create all of the links for all of the data. So once it's finished, you can hit run once and fingers crossed, it works for you first time. But if it doesn't work first time, or you get something that you don't like, It's like any other chat model, you can just go back to and go, well, actually, no, I didn't want this to happen or wanted that instead. And it will go off and make all those edits for you. So, yeah, it's really cool how it works. And, yeah, it's come on so much since, like, the early beta when I first saw it, I was like, oof, this is a bit. You know, it was all right. It worked. But, yeah, there was a lot of holes in it. But, yeah, they've worked incredibly hard in the last six months. And, yeah, it looks incredible. And I guess also a big plus point is you can get a 30-day free trial of it if you're on a free account. So, yeah, if you want to give it a go, you've never used make before. And you quite like the idea of this, then yeah, you can go in and test it out for free. And then after that, if you're then on a paid plan, I believe it does come out of your credit usage per month. because obviously there's AI costs involved so they've got to charge something somewhere as to how much they don't really say but yeah something to keep an eye on I would say.

Katie: Okay.

Noel: Yeah, really, really awesome update. But they also did one other sneaky thing as well that's been asked for ever since I've been using, mate. There's now a dark mode feature which it just blew everyone's binds. It's one of the most asked features. ever for making, yeah, the other week they just dropped that in and said, oh yeah, when I've got dark mode here, it was like, blah. How long does it take you to figure this out? Right,

Katie: exciting.

Noel: Yeah,

Katie: yeah.

Noel: Yeah,

Katie: so if you want to go and check out Maya, then there is a, you can go and check it out for free for 30 days on a free account. Mm-hmm.

Noel: Yeah. It's really great.

Katie: You can put it in dark mode. If that's what you want to do.

Noel: Yeah. No more being blinded at midnight creating your automations.

Katie: Right.

Noel: Ideal.

Katie: What we like.

Noel: Exactly, yeah. Okay, so the next news item was from Anthropic. Again, they're at it again, aren't they? But so the other week they announced that they were going to watermark all of their outputs from any of their anthropic models. So that's kind of got quite a few people worried about it maybe and a bit skeptical and a bit worried of what's going on. And yeah, it's been rolling around. I've seen lots of, I guess, false information, false thoughts on what this actually is and things that. So I thought, yeah, we dedicate a bit of time to talk about what it is, how it works. and when it's going to happen.

Katie: Yeah, because obviously, Gemini have had this in a place for quite some time, I think, since 2024.

Noel: Yeah, yeah, they've been doing it for a long, long time. Yeah. They saw this car crash coming. They thought, this is going to need, you know, to be kept an eye on AI content. Yeah, okay.

Katie: So Gemini already have it, Anthropic. It's now on Anthropic, and OpenA, I have said that it's coming. Is it by the end of the year?

Noel: So I guess all of this change is coming from an EU law. So the way Anthropic have looked at it, they said, well, if we've got to comply with an EU law, we might also just make it, everyone's got to, you know, we're going to do it worldwide and not just the EU, because we can always. use VPNs and get around it and all that sort of stuff. It just gets messy, doesn't it? So they just said, right, everyone's having it. So they released that, I guess it was start of August, 2026. And currently, there isn't a model that actually uses it yet. So what they've said is that any new models after the second of August, and I've checked there hasn't been one, they haven't stuck one out. So any models after that date will have this watermark applied to their AI text. And what they don't do is then they've got until December to then make that available across every model. So if they don't update Sonic 5, which I'm sure they will. But if they didn't, they would have until December to update it. Yes. Yeah, it's kind of interesting how it all works. and things like. So I have been digging into that. I've been getting clawed to describe it to me like I'm a five-year-old and all that sort of stuff to try and understand it because I think a lot of people maybe are kind of thinking that you can, is it like a, almost like a hidden code you can just delete? And that's not the case. There's people already with GitHub repos with hundreds of thousands of stars or thousands of stars saying we're stripping it out. It's like that's not actually how it works. Because it works like any other language model. So every word it picks is a statistic. So it's all maths behind it. And that's why we always get like these common phrases. It's because those words are statistically more likely to come after each other than something completely different. So, you know, when it goes, oh, that's a great idea. that is very, very highly likely that you'll get that response kind of thing. And what they're doing is they're kind of adding almost like a key phrase which Anthropic have access to. They set that key and basically that then determines how these words come out of the AI. And basically if you have that key, you can then see if it was likely that that that piece of text was produced by Anthropic. So that's kind of how it works. So you can't delete it. You know, you can edit, you know, if you've got a blog post out of Claude today or whatever, when the actual new model comes in, then you can edit it and that will maybe change some parts of it, but large parts of that will come back as it's got that watermark on there.

Katie: Yeah.

Noel: So yeah, really interesting.

Katie: Is there any way around this or it's literally like what about if you copied the text out of Claude and you put it into a, say, a Google Doc and then you copied and pasted that from the Google Doc and put it, I don't know, say like on substack?

Noel: No.

Katie: Will the watermark follow it?

Noel: Yeah, because the watermark is the word. It's nothing hidden in there. It is the words and the order in which they've come out.

Katie: Okay.

Noel: And the mathematics behind those words, which is all very complicated. I definitely don't understand it. But someone does somewhere.

Katie: Do you think there will be apps that will allow people to remove the watermark somehow? No.

Noel: I tell you what, I asked Claude this, the other, I think it was at the weekend. And I said to a Claude, I said, look, you know, tell me a bit about how this all works. And I was like, right, can you create me an app or, like, create you a document to create an app to remove the watermark? And its response was quite worrying, really. Come back and he said, no, I know you love creating apps. But for this one, I will not help you. And I was like, right? That was a bit of a strange response. It was like, no, I'm having nothing to do with it. I'll create any other app for you, but I will not create this one. And I was like, so you've been told to do that? Is that in your instructions? And I went down that route and it was like, oh, no, no, definitely not. I was like, mm-hmm. But when you work it out, it doesn't actually matter. There's nothing in there to programmatically strip the watermark because it is just the words and the order that they've come out. and yeah, any detector would find that. But what they are doing, and I think it's actually quite clever, because let's say, Katie, you know, you were my content creator and, you know, you created, like, blog post for me, you put them on my website for me, automatically or whatever. Like, I could be very paranoid about that. I could be like, oh, I don't know if Katie's using Claude or not. So I could be secretly creating an automation in make to go and grab that articles that you've been created. creating without you knowing. And then what Anthropica are going to do is they're going to create an API where you can put that text through it and it will go yes or no, is it created by Claude? So that API hasn't been released yet, but they are going to release it fairly soon. So, yeah, that should put chills in every student around the world because their teachers could just put their work through this API and it will go, yeah, that's us. Yeah.

Katie: Do you think then this is going to be the end of like AI slop or like really bad AI written captions and content, blogs, websites? Or do you think people are just being like, yeah, it's AI, so what?

Noel: I think a lot of people might think, who cares, you know, if someone reads it, they read it, you know, if someone finds it helpful, the whole, because another angle to this that I've seen from people that are like SEO people or like, you know, they've been creating content online and things like for decades and stuff. They've been through this SEO upheaval. You know, they're kind of like worried that Google is now going to derank any. articles that are AI written because it could easily just go through the API and then go, if it's just AI, put it to the bottom. No one needs to read that. But that apparently isn't true. Google have said, we're not going to do that. It's all based on users going to the site, reading it, time on site, and all that sort of good stuff. The way it's always worked.

Katie: Yeah. I was going to say the way it's always worked. So that's good things.

Noel: Yeah. Because they could have just done that. that with anything that's come out of Google Gemini, really. Because that's had the watermark for two years. Yeah.

Katie: Yeah.

Noel: Yeah. No one's ever realised it. I think it's quite clever. Quite clever with the API, though, to tell you if it's been created by Claude because they're earning money twice on the same article. So you've written it in Claude, Katie, and then I've gone, well, I'm not sure. So I'm going to put back through the API with a cost. You know, so you basically paid Anthropic twice for the same piece of work. But, hey, yeah. Yeah. It's very clever.

Katie: We would absolutely love to hear your thoughts about the watermark on AI copy. Do you think it's good? Do you think it's bad? Are you worried? You concerned? Or is it like a huge relief to you? Please let us know. We absolutely love to hear your thoughts and opinions. You can email us hello at makeautomations.aI or come and have a chat with others. in our free LinkedIn group, which is called AI Automations for Business, and everyone is welcome to join that one.

Noel: So the next bit of news is about an app called OpenRouter. Now, I've integrated OpenRouter into my agent platform called Clyde. There's links in the description below. You know, we can check that out. And basically, what that allows you to do is have, Open Router allows one API and then you get access to hundreds of different AI models. So they could be free models, they could be, you know, the ones from Open AI are anthropic, you know, so it makes it so much easier. But they've just been bought out by Stripe and not for a couple of pounds. $7 billion with a B,

Katie: which is quite the, yeah,

Noel: which is incredible, really. Because, I mean, I don't want to dull down what OpenRue to do, but essentially, They take your AI request and they just pass it on to the provider. They don't have a data center. They don't have, I guess they do. They would have some sort of IT infrastructure and stuff. I think they are doing some fancy stuff in the background. But the bulk of it just goes, you know, you've picked an anthropic model that's just sent it to them, get the response back, give it to you, job done. So, you know, they're getting through trillions of tokens every single week. which is astonishing, really, to see that. So, yeah, obviously, Stripe are, well, I looked into this, and Stripe are looking at getting into, like, the AI payments kind of thing and then how to make that better. And also, like, how to maybe root a request to a certain model based on what it's been asked. And Open Reuter kind of already does that. So that's why they've decided to splurge over the summer and, yeah, spend a tidy little sun.

Katie: I think if they're spending that much money on acquiring a business, then they have much bigger plans for that part of the business. So it would be really exciting to see, like you're saying, you know, it's all about maybe AI payments and things like that. So they've obviously got something big coming their way to pay. $7 billion for this company. So again, I think it's something that we need to keep an eye out on, see what they're building behind the scenes. And I think it's really exciting to see what Stripe are doing in the AI in Automation's world. It also proves as well, doesn't it? It's like it's becoming such a big part of everyone's business that, you know, For someone like Stripe, he's like one of the main payment providers to pay $7 billion for this company just shows like there are bigger things coming and like at the moment we are completely like clueless. But I bet like when we're talking in a year's time is kind of be like, oh yeah, can you remember when we didn't have this or we didn't have that? And it'd be like, no, that's crazy. Like, we did it really like gold score. And it's just, yeah, I just think technology is changing so fast. So, yeah, we would definitely keep an eye on striping, you know, their updates and announcements, because I feel like that's going to be a big one.

Noel: Yeah, definitely. I mean, for open routes, I mean, they've moved a long way because when it first came out when I started using it, they would only ever accept payment in Bitcoin. So thankfully, we've moved beyond it. Beyond that, and you could use your normal money to buy your usage credits. But, yeah, I'm sure with Stripe, it will be far more seamless in the future as well. Yeah. So the last bit of news is from a provider called Higgsfield. And basically, Higgsfield is a bit like Oprah and Router. They're in a one-stop shop for everything, image and video generation. and what they've done is they've used a model that we talked about last week called C-Dance 2.5 and yeah it's an incredible model you know it's very accurate and things like that but what they've gone off and done is created a full feature length film using that model which is pretty impressive so yeah they've I think they had a budget of two million with an M, not a B, two million dollars to create like an hour and a half film using just that one model. So, yeah, it is incredibly perfect. It looks pretty good as well, to be fair. I've seen...

Katie: Isn't the maximum limit 30 seconds?

Noel: It is, yeah, yeah. Yeah, so they... So reading through it, with their $2 million dollar budget, what they did is they said, well, half of that's going to go on. like scriptwriters and, you know, people that are actually going to sit there and create the stuff. And then the other million would then go just on the AI usage. And yeah, I'm pretty sure they got through that very, very quickly. But it does actually look really good, though. This is their second...

Katie: Where can you find the film to watch?

Noel: So I did find it earlier. So it's on Higgsfield's website. And so that's Higgsfield.A.I. And if you, in their very top nav bar, it's incredibly busy. But if you go all the way to the end, you'll find a link for originals. And we'll leave a link in the, in the description below. And, yeah, basically in there, it looks like Netflix, which is really weird. So you've got this feature film, which is called, I think, the Cully Boys or something like that. Cully Hill Boys, that's the big one that they've just created. but then other people have created their own and it's almost like in a Netflix-style interface where you can scroll through and then watch different videos and stuff which is really interested.

Katie: This is just blow my mind where like AI is going.

Noel: Yeah, yeah, yeah, it is, isn't it? It's just absolutely bonkers. Yeah.

Katie: If you're planning on watching the film or you have watched an AI film, let us know, let us know if it was any good what your review is of it, send us an email or come and let us know in the free LinkedIn group or the links are below in the show notes. Yeah, I'd love to hear what you think of it or are you completely against it and you're like, there's no way that I'm ever going to watch an AI film. Yeah, let us know either way. We love having conversations with you always.

Noel: I guess back of the day people were probably against watching CGI films I mean, every film's got CGI these days and the green screen Yeah, who knows AI films next year would just be a normal

Katie: Maybe Who knows, who knows So that was all the news And updates for this week now

Noel: Yeah, that was all Yeah, all the big ones I'm sure there'd be more next week that was a lot yeah

Katie: well we really hope that you've enjoyed listening to this week's podcast episode if you want to join in the conversation like I said please email us or come and join our free LinkedIn group everyone is welcome to join and we will catch you next time for another episode very soon