States rush to police AI deepfakes ahead of midterm elections

**SPEAKER_1** (0:00)
Running a business is hard enough, so why make it harder with a dozen different apps that don't talk to each other? Introducing Odoo, the only business software you'll ever need. It's an all-in-one, fully integrated platform that makes your work easier, from CRM, accounting, inventory, e-commerce and more. And the best part, Odoo replaces multiple expensive platforms for a fraction of the cost. This is why over thousands of businesses have made the switch. So why not you?
Try Odoo for free at odoo.com. That's odoo.com.

**Meghan McCarty Carino** (0:31)
The Claude Slop Reckoning is here. From American Public Media, this is Marketplace Tech. I'm Meghan McCarty Carino.
Anthropic announced it's adding invisible watermarks to all text generated by its chatbot, Claude. We'll get into it on today's Marketplace Tech Bytes Week in Review. Plus, there's a patchwork of state laws that regulate AI-generated political messaging, and they're being put to the test ahead of the midterm elections. And federal regulators ordered Kalshi to continue to operate in New York, despite a state lawsuit seeking to shut the prediction market down.
But back to Anthropic's new watermarking, the AI startup is rolling out the policy globally as part of efforts to comply with the European Union AI Act. Maria Curi is here to explain. She's a tech policy reporter at Axios, and she broke down what the EU law actually says.

**Maria Curi** (1:39)
It's a transparency measure that essentially requires all of the AI labs to tell users when content is generated by artificial intelligence. And so this applied to all of the major labs, but Anthropic as far as we know is the only one that is applying it globally, even outside of the EU. And it kind of follows its reputation for being the lab that's most concerned with safety and making sure that those kinds of safeguards are in place. We'll see if the other labs follow suit.
But it's an imperfect system, right? You could, for example, put your original content into a model and just put it in there for editing and it'll still have this watermark on it. And is that AI authorship? Probably not, but it would have that watermark. So we'll see how it evolves.

**Meghan McCarty Carino** (2:27)
Yeah, it's interesting. There are two sort of ways that the watermark works. Apparently it sort of, it embeds something in the metadata if you copy and paste from Claude. And then also it sort of outputs the words in kind of an encoded way that, you know, to a machine, to the decoder that Anthropic has created, makes it clear that this is AI generated. Sort of an extra weight in the model that, you know, that is arranging words in a way that is, that's the watermark. And so yeah, there's been a lot of debate about kind of what some of the implications are for this, aside from obviously, you know, potentially outing a lot of the people who are using AI on social media and in their work and all this kind of stuff. But that there could be potentially some downsides, as you noted, you know, the, if you run your copy through Claude to fact check to, you know, copy, edit, and then sort of copy and paste back, that would embed this watermark. Or, you know, there are concerns about, you know, this potentially degrading the outputs of the model. So a lot of debate going on.

**Maria Curi** (3:48)
Yeah, absolutely. And in terms of the way that it works, they might also retroactively have some of the older models adopt these watermarks. So as of now, there might be AI generated text out there in the world that does not have a watermark because those older models did not have this tool embedded within it. But Anthropic might retroactively do so. So it's very interesting. There are a lot of teachers out there that are happy right now because, you know, they're going to be able to tell whether or not their students used AI.
But I am very interested to watch in this space how much AI use continues to be sort of like under the cheating umbrella, or if it becomes just more normalized. And I think we've seen that with images a bit, you know, on Instagram, for example, you have a lot of content creators just having the AI generated tag on it and people still engage. Sometimes it's not a complete deep fake. So I think it's creating a lot of very interesting societal questions about what is original, what is not, etc.

**Meghan McCarty Carino** (4:47)
Well, speaking of AI generated images, we want to talk about some reporting that you've been doing on state deep fake laws going into the midterms. 29 states, you report, have laws that regulate political deep fakes in some way.