Latest / The Best of the Money Show / Tech Thursday: Claude’s watermark plan and the future of trust
Transcript
- 0:00Speaker 1: No money show take Thursday.
- 0:03Speaker 2: Well the AI model claud Ai saying, now, now include
- 0:06Speaker 2: a watermark and everything it produces. That would be a
- 0:09Speaker 2: game changer. You can imagine being an author of publisher
- 0:12Speaker 2: in academia now you work as this watermark mark Cepumelela
- 0:16Speaker 2: Zondies art tech expert houses Pumle. I mean, people have
- 0:19Speaker 2: been talking about this idea for a while, and I
- 0:22Speaker 2: can understand how a watermark would work in an image,
- 0:25Speaker 2: but how would it work in this case?
- 0:28Speaker 1: Hello Sivien and and yes, So it's not a visible
- 0:30Speaker 1: watermark that's going that you're going to see when you
- 0:33Speaker 1: are using AI to generate let's say, speeches, content essays
- 0:40Speaker 1: for you. This watermark is going to be embedded in
- 0:43Speaker 1: the words and using metadata. And so it's only missions
- 0:49Speaker 1: machines that are essentially trained to detect the watermark that
- 0:53Speaker 1: are going to see that as someone actually or the
- 0:57Speaker 1: person or the individual or the student use AI to
- 1:01Speaker 1: produce the work for them, and they didn't do it themselves.
- 1:04Speaker 1: So it's going to be embedded in the text itself,
- 1:07Speaker 1: as opposed to it being something that's going to be
- 1:09Speaker 1: visible that can be seen with a naked eye.
- 1:12Speaker 2: So would you be able to copy and paste it
- 1:15Speaker 2: and get around it like that, or would a computers
- 1:17Speaker 2: still know that this is actually the result of AI
- 1:20Speaker 2: because it's been done in another way.
- 1:23Speaker 1: So listen, people always figure out ways of trying to
- 1:26Speaker 1: treat these things, and there are always as soon as
- 1:30Speaker 1: there's a way of detecting something, then someone creates an
- 1:34Speaker 1: AI that's going to try and bypass that. But teef,
- 1:36Speaker 1: initially you are going to you can copy and paste,
- 1:40Speaker 1: but even when it's copied and plasted, it's still going
- 1:42Speaker 1: to figure out that it was written by AI in
- 1:48Speaker 1: the way that it's going to be developed. But as
- 1:51Speaker 1: I say that, they're always there's always someone who's trying
- 1:54Speaker 1: to give people another method, another tool that's going to
- 1:58Speaker 1: try and cheat it. But initially, yes, even if you
- 2:01Speaker 1: copy and paste, it's going to be detected, but.
- 2:04Speaker 2: It would be still invisible, right, And I mean this
- 2:06Speaker 2: is the thing. No one wants to read something that
- 2:08Speaker 2: has that is sort of obscured by some kind of
- 2:11Speaker 2: watermark in any kind of way.
- 2:14Speaker 1: Well, definitely invisible. And it also with AI depends on
- 2:16Speaker 1: what you're using AI for, because what you find is
- 2:20Speaker 1: that there are people who use AI to pounds of ideas,
- 2:22Speaker 1: but they're not going to be producing AI in a
- 2:25Speaker 1: public space as the work that they did for themselves. Yes,
- 2:30Speaker 1: there have been situations like I came across a reader
- 2:33Speaker 1: of a Canadian politician who read AIG work and then
- 2:38Speaker 1: to the point that they even read when AI was
- 2:41Speaker 1: telling them to that's a better way, A different way
- 2:45Speaker 1: of doing something or saying something would be a particular method.
- 2:51Speaker 1: But in a case of where you're using AI to
- 2:55Speaker 1: pounds of ideas, then maybe it's okay. So no, it's
- 2:58Speaker 1: not going to be visible. Oh it's not going to
- 3:00Speaker 1: be detected. You're not going to see with a naked
- 3:04Speaker 1: eye that no, this is this is AI. And if
- 3:07Speaker 1: you're using it for yourself, then it shouldn't be a
- 3:09Speaker 1: problem because you're using it for your own personal use
- 3:13Speaker 1: and not to say that you produce these detect yourself.
- 3:17Speaker 2: Are we going to see the other AI models and
- 3:20Speaker 2: the other people who make it so Gemini or chech GPT,
- 3:22Speaker 2: are they going to do the same thing.
- 3:25Speaker 1: They haven't said whether they will right now, so in unlikeliness,
- 3:29Speaker 1: they're not starting right now. But this could be starting
- 3:31Speaker 1: a trend because there has been issues and very many
- 3:35Speaker 1: court cases, especially in Europe for example, where there's been
- 3:40Speaker 1: caught cases where there's worry that AI sort of generating
- 3:44Speaker 1: work that's created by humans or training AI with work
- 3:48Speaker 1: that's been created by humans without without the AI acknowledging
- 3:55Speaker 1: the original creators of the work. So because of those
- 3:58Speaker 1: cod cases, so AI companies need to figure out ways
- 4:02Speaker 1: of how they are trying to be transparent and acknowledging
- 4:06Speaker 1: because that this is something that came from elsewhere. So
- 4:09Speaker 1: in all likeliness, this is using a trend, and in
- 4:12Speaker 1: unlikeliness they are going to also try and adopt either
- 4:16Speaker 1: this method or other a similar methods.
- 4:19Speaker 2: Okay, I mean you know, to look at those court
- 4:21Speaker 2: cases in other way, so authors take the AI model
- 4:25Speaker 2: to court, they say, you've learnt from my work and
- 4:28Speaker 2: you haven't given me proper, proper credit for that. Would
- 4:32Speaker 2: this help them because over the longer term, if everything's watermarked,
- 4:35Speaker 2: you would know where it comes from. I don't know
- 4:37Speaker 2: if the watermark would then be able to tell you,
- 4:39Speaker 2: you know, like if you're going to Gemini and I,
- 4:41Speaker 2: so you do a Google search, it'll tell you where
- 4:43Speaker 2: the responses come from. That might work in the same
- 4:48Speaker 2: way to see the original author of content that's now
- 4:50Speaker 2: being presented to you.
- 4:52Speaker 1: Well, it might also lead people to go and research
- 4:56Speaker 1: or go to the original author. But the issue as
- 4:59Speaker 1: well has been on the money, right, because if you
- 5:02Speaker 1: training AI without an agreement, and yes, it might lead
- 5:06Speaker 1: to the people some people back to the original author
- 5:10Speaker 1: of the work. I'm not even sure if people click
- 5:13Speaker 1: on the link to the original also on on on Google,
- 5:16Speaker 1: Google AI when they're getting those AI responses, I would
- 5:19Speaker 1: think they don't. But also who then makes the money
- 5:23Speaker 1: because sometimes you find that things are hidden behind pay walls.
- 5:28Speaker 1: Do people then go back and do they pay? Because
- 5:31Speaker 1: the AI itself did not pay the origin the originators
- 5:35Speaker 1: of the concept before it trained its models on the AI.
- 5:39Speaker 1: So it might be a way that they are trying
- 5:42Speaker 1: to say, listen, we are safeguarding ourselves. But it also
- 5:46Speaker 1: doesn't solve the conundrum of the fact that the originators
- 5:49Speaker 1: of this content are losing money because a AI models
- 5:53Speaker 1: are being trained on the work that they would have created.
- 5:57Speaker 2: I mean, some publishing firms are looking for a EE
- 6:00Speaker 2: engineers as well. I mean, do we know what they're
- 6:02Speaker 2: going to use them for?
- 6:04Speaker 1: So and that's actually quite interesting because you worry when
- 6:08Speaker 1: you realize that publishing firms are looking for AI engineers
- 6:11Speaker 1: because they've always been dead set on saying that, listen,
- 6:15Speaker 1: we're looking for human authors, et cetera, et cetera. You
- 6:18Speaker 1: wants work that's created by humans, But for now they
- 6:21Speaker 1: are using AI engineers, not in in the authors of
- 6:27Speaker 1: the text. But they say it's to improve internal business operations,
- 6:32Speaker 1: manage data for optimization online for meta data tagging, so
- 6:38Speaker 1: their work and the work of the authors that they
- 6:41Speaker 1: publish is easily found online in order for them people
- 6:45Speaker 1: to go and purchase the work that would have been
- 6:48Speaker 1: authored by the people that they have deals with. So
- 6:53Speaker 1: so far it seems to be for business purposes and
- 6:56Speaker 1: marketing purposes, and it doesn't seem to be to author
- 7:01Speaker 1: the work for now at least.
- 7:04Speaker 2: I mean, I'm one of those people when I read something,
- 7:07Speaker 2: whether it's a letter or whatever, I go to the
- 7:10Speaker 2: bottom of it the last thing to see who wrote it, Right,
- 7:13Speaker 2: I want to know who that person is and what
- 7:14Speaker 2: their position is. And I'm sure not the only person
- 7:17Speaker 2: like that, because something is only as important as the
- 7:19Speaker 2: person who said it. It's a sad fact, but it's true.
- 7:21Speaker 2: The context of who wrote it is so important that
- 7:24Speaker 2: must be the case as to whether it was AI.
- 7:27Speaker 2: I mean, you can read something and be shocked and horrified,
- 7:29Speaker 2: and if it's written by a computer, you know it's
- 7:31Speaker 2: a hallucination.
- 7:32Speaker 1: And also with a computer that does not credits or
- 7:35Speaker 1: because sometimes with AI that it would not give you
- 7:39Speaker 1: credits or full wrote it. And also as you mentioned,
- 7:42Speaker 1: there've been hallucinations as well, and this is where this
- 7:45Speaker 1: can also help, so you know that it's a real human.
- 7:48Speaker 1: It didn't hallucinate and makeup stuff and creates that also
- 7:52Speaker 1: that don't even exist, because it's being found at times
- 7:55Speaker 1: to have created ausos that don't exist. So it has
- 7:58Speaker 1: citations you cannot find anywhere else. Think about AI policy
- 8:04Speaker 1: for example, that was I had to be pulled back
- 8:06Speaker 1: by the government where I found that it's hallucinated and
- 8:11Speaker 1: there were authors and there were citations that didn't exist
- 8:14Speaker 1: that were found there. So this is where it can
- 8:17Speaker 1: also help to know that it's AI that got to work.
- 8:20Speaker 1: And then because the is this watermark that founded the
- 8:25Speaker 1: wording and it also then shows where the work actually
- 8:29Speaker 1: came from.
- 8:30Speaker 2: See put mala Zoondie. Thank you to appreciate that tech eches,
- 8:33Speaker 2: but so much going on around AI as always, but
- 8:36Speaker 2: I mean the actual productive use of it. I have
- 8:39Speaker 2: to say, in my job, where you know, credibility is
- 8:43Speaker 2: everything and everything has to be true and it's all
- 8:45Speaker 2: about facts, I tend to sort of steer away from it.
- 8:48Speaker 2: I don't mean that I don't use it when I'm
- 8:51Speaker 2: searching for something, or what I really enjoy and find
- 8:54Speaker 2: incredibly powerful is instead of this Google search, I can say,
- 8:58Speaker 2: you know, what is that or how many or whatever,
- 9:01Speaker 2: and I'll get a coach and answer. And I really
- 9:02Speaker 2: appreciate that, and it does make life a lot quicker
- 9:06Speaker 2: and easier, and I think that's really important. But for
- 9:09Speaker 2: me otherwise, not as much as everybody says. But it
- 9:12Speaker 2: may just be me being old fashioned and a journalist
- 9:15Speaker 2: the same thing.