Watermark My Words

But, you know, there is this kind of growing question of, of what can you trust online, what is real, and we're gonna talk a bit about fact-checking later, which I think was kind of the original attempt to help validate what people saw on the web.

But we have been waiting for the AI labs to really come up with their own play on this, and Anthropic this week announced something that I think fits neatly into that bucket.

Yeah.

So I...

You know, this sort of all traces back to the EU's AI Act, which, you know, I think has been seen as sort of a really early jump into the space of regulating AI in, in all sorts of ways.

And one element of that is a requirement for AI-produced content to be watermarked in some way.

You can understand the sort of thinking behind that, in that there was a lot of concern.

You know, a lot of the initial concerns about generative AI tools was that it was generating content that would fool people into believing that it was real.

I mean, the, the deepfakes, right? We spent years talking about the threat and risk of deepfakes and how they were fooling people, except that there was maybe not that much evidence of people actually being fooled by deepfakes.

In fact, it seemed like the bigger problem was people doing what's been known as the liar's dividend, where they would claim when they were really caught doing something that it was a deepfake.

But, so built into the EU Act is this requirement that AI companies somehow watermark or indicate content that is created by AI.

Now, there have been efforts underway.

There's a, a standard which is now pretty widely being adopted for image and audio and video AI-generated content that, from what I've heard, is getting pretty good and is fairly hard to break, though I imagine people will come up with all sorts of ways to break it over time.

But what was trickier was the idea of what AI is most used for, which is generating text, and how do you indicate and mark that text is originally generated by AI.

So Anthropic announced this week that they're now very soon starting to watermark all AI text that they generate and/or process, which is a, an interesting distinction that is worth digging a little deeper into.

But basically, what they're doing is they have put some sort of, at this point still hidden, kind of bias into their text generation that will be recognizable to a machine-readable system to indicate that the text has been processed by Anthropic's tools.

And so at some point they will make an API available that will allow you to take any text and then ship it back to Anthropic, and Anthropic will give you some sort of score as to whether or not that text was processed by Claude in, in some form or another.

Mm-hmm.

Now, this...

There's the, the regulatory aspect of it, which is that they're sort of required to do this.

It...

As people were talking about this, it came out that apparently Gemini, Google's AI model, has been doing this for some time already.

At some point, I assume that OpenAI will release something similar.

The reactions to this have been all over the map.

A lot of people are really, really mad about it, some for perhaps reasonable reasons, some I think for less than [chuckles] reasonable reasons.