C2PA and Content Provenance in the AI Era – The Signal EP 02

Hey everyone. " This is a reoccurring series that we do here at Techstrong, along with our good friends at Digicert. And welcome to another episode.

" Well, not in today's world with all these deep fakes, maybe, but anyway, we're going to be talking about C2PA and content provenance in the AI era. I've got two great people to talk about this with. Let me introduce you.

First of all, from Digicert, he's one of my good friends there, we've been talking forever, my friend Mike Nelson. Mike, welcome. Hey, my friend, good to see you.

It's great to be back on. Absolutely. Mike, you, of course, are field CTO at Digicert, but give people a little bit of your background.

Yeah. I've been with the company for about 11 years. I spent my career, before coming to Digicert, largely in healthcare technology, but have been with the company for 11 years, and we've done a lot of great work in securing the internet, and then we've expanded in doing things like connected devices, software, and now in this age of digital content, we're doing tremendous work in that space, trying to ensure the integrity and safety of content that we're seeing online.

Excellent. Thank you. Joining Mike and I today is Nima Mohebbi.

Nima is a partner at Sidley Austin. Nima, welcome. It's great to have you here on this podcast.

It's great to be here. Thank you so much, Al. So Nima, give people a little bit of your background, and if you don't mind, a little Sidley Austin primer as well.

Absolutely. So I am a partner, as you said, in the Sidley Austin Century City office. I primarily practice litigation, but a lot of my work has to do with entertainment, sports, media, and a lot of AI tech-related work.

So this is right up the alley of all of my clients right now, and on the top of mind of every studio that we deal with. So very timely. Absolutely.

Look, generative AI, agentic AI, AI in general, physical AI, it's certainly changing how we create content, how we edit it, how we distribute it, how we consume it. How, and most importantly, how we trust it or not. Right?

Organizations are now facing two related questions, how to comply with these new transparency expectations. Not yet regulations, fortunately or unfortunately. At this point, it's still kind of expectations.

But more importantly, how to give audiences reliable information about what they're seeing and what they're hearing. Is it real or not? And so couldn't think of two better people to discuss this than an attorney whose practice is kind of sitting on top of that right now, and my friend Mike Nelson, and the folks at Digicert who bring maybe a little bit more of a technical perspective rather than legal.

So let's start with this, Mike. When we talk about C2PA content credentials, what are we talking about? Yeah, great question.

Well, I like starting with the problem, and I think you set it up well, which is in today's day and age, there's so much digital content, the documents that we're signing when we're closing a loan, the videos that we review online, the photos that we're sending around. " And it's just really, today I like to describe it as we don't know what we can trust. And that's the problem.

And it's not that editing is bad, but C2PA was formed to try to address the issue of content provenance. Where is content coming from? Where was it generated from, and what's the origin of changes that have been made, or what's the record of changes that were made?

And at this point today, that doesn't exist. And so C2PA, it stands for the Coalition for Content Provenance and Authenticity. But a group of organizations came together, largely video editing, video capturing camera manufacturers came together, and they said, "We have a problem.

" And that becomes really important as consumers and media outlets and judges are trying to determine the trustworthiness of content. And so that's what C2PA is trying to solve for. Yep.

And there's another element to this too, right? It's not just that this picture or this video, whatever this likeness is in fact truly that person, but there's also the intellectual, the copyright and IP rights of who gets paid for the use of that. You know, Nima, 100 years ago I went to law school.

I practiced law, right? For five, six years. I hated it.

Between you and me, and I went the other way into technology. But the big case when I was in law school was the Bela Lugosi case, which I'm sure you're aware of, right? And that set forth the principle that actors, public people are entitled to be compensated for the use of their likeness.

Yep. And so in this day and age where we can't tell if that's really supposed to be the likeness of this person or that person, or I used it here once, but can I re-use it? Do I still got to pay?

I mean, it's almost like an impossible task, and I know out in Hollywood and in the film industry and music industry, everywhere else, this is a real issue, Nima. Yeah, absolutely. So there's a lot in that question or setup, but I will say that there really are two separate questions.

We're talking about provenance, watermarking, things like that. Really transparency roles, if you will. And then the separate thing is who's liable for using the material?

Who gets compensated for using the material that ultimately is created? Who owns it, for that matter? And that's a much larger question.

We don't have answers to a lot of that stuff. There's court cases going all over the country, for the last few years that have been dealing with those sort of IP questions. But, as a practical matter, artists, celebrities, athletes, so forth, they've recognized that this is an issue.

And several have been proactive in essentially licensing their likeness, their voice, their face, pick your poison. The reality is that, could it still be fake? I mean, yes, the technology we have makes it easy to do that.

The question is whether or not the rules are sufficient enough to prevent this sort of stuff from happening. And I think that our current system of copyright rules and also our NIL rules, you mentioned the name, image, likeness, which is short for NIL. They apply nicely actually to content that is created by AI that is ultimately infringing on one's likeness or infringing on somebody's copyright And then going back to the earlier point that you said there hasn't been regulation on this.

I would just slightly correct that because really in the past few months, there have been two pretty significant laws that have been passed that deal directly with the issue of transparency in AI, and that's the EU Article 50- Yes ... and California SB 942. Mm-hmm.

And you can't leave us hanging there, Nima. Oh, yeah. How does this play into that?

Yeah. Of course. So let's talk about this.

I think Mike presented a very good summary of C2PA and how it fits within this realm, and I'm going to try my best not to trip over my own shoelaces. And Mike, you'll have to protect me here because while I do a lot of tech, I only know enough to be dangerous. I got your back.

Thanks, man. We're good. Appreciate it.

So, okay. So the EU act was pretty significant because it's an instance of a broader Western region that essentially requires that the providers of generative AI mark their output. So text, image, audio, video, in a machine-readable format that's essentially detectable as artificially generated or synthetic.

Now, separately, the law also requires deployers, so people who, not the creators, the LLMs of the world, OpenAI, it's like the generators of the artificial intelligence, but the deployers, folks who are actually deploying it. There's a separate disclosure obligation when they use those systems to essentially produce deepfakes, things that are used for advertising and are otherwise in the public interest. So that law just went into effect.

California has passed a suite of AI laws in the past year, and actually rounding out as of, I believe, yesterday, Gavin Newsom signed another round of AI protection laws. The big one, I guess the one that's probably most relevant here, is the AI transparency law, which was enhanced with this new suite of legislation. And that's SB 942, which facially applies only to large generative AI providers.

I guess it's defined as over a million monthly users, so you kind of know who they're aiming at there. And they are required to embed technical watermarks in AI-generative content. And separately, they're required to provide a free public detection tool so that the public can also use to verify whether or not something is AI-generated.

You might have seen the news recently about the Anthropic- Yeah, go ahead ... " But in reality, it's watermarked in a way to sort of comply with some of these transparency requirements. Yeah.

It's not an effort by Claude to own the content. No, not at all. But here's my problem.

We're at a phase in our use of AI where when we say AI was involved in creating content, there's sort of a scarlet letter stigma attached to it. Sure. And not all AI-generated content is equal.

I may write an article. I wrote it. I put it into AI to copy edit, and it made some good copy edits.

It's great at copy editing. Does that mean it's an AI-generated piece of content? Same thing with a picture.

I bring a picture in there and say, "You know what? I don't like that photo behind Nima on the wall there. Just take that out, but leave the rest of it alone.

" Well, see, now you're getting into the issue of IP and who owns what copyright, the rules that apply there. Well, but should it bear the scarlet letter? " So I want to hear Mike on this because I think that to me, this transcends legal and really what consumers and clients are willing to accept or what we're willing to essentially absorb as content.

Yeah, it's a tricky question. And what I will say is I think the world is trying to figure those questions out, and I don't think we've quite gotten there yet. Nope.

But I do see a rush to some of the things that we're talking about, like the watermarks. C2PA even has a little watermark for all images. So if you go to Digisert's LinkedIn profile, we're following protocol on this.

Every image we release has a C2PA watermark, and you can click on it, and it shows where the image was captured, what modifications, and you can go and you can see that. That level of transparency is what consumers need. There was a really interesting court case here in our backyard in Utah.

Charlie Kirk, Tyler Robinson, the man accused for the assassination there, was in court here. This was a couple of months ago. And really interesting, the judge in that trial, the state of Utah presented some videos, and in the videos, they had done some zooming, and they had made some green.

They had edited the video in some ways. Editing's not bad, but the judge, I found this really fascinating, said to the state, "How do I know that this video is authentic? " And the state kind of looked at each other like, "Oh boy.

" That's the kind of transparency that consumers need to be able to make assessments for what they're looking at. You think about it in the insurance world. Yeah, no.

In insurance. Yeah. Mike, with all due respect, you're talking about chain of custody and evidence.

Yeah. You go through the OJ trial, right? What happened to the gloves- Yeah ...

or all the other nonsense. Well, we're really good at doing it with physical evidence. Right.

In today's digital world, yeah, okay, maybe. But we're used to doing it with physical evidence. Uh-huh.

And maybe we're not that good at it, but it's standard protocol. But now that- But there's a different threshold, right, for something being admissible in court versus what a consumer is willing to accept on Facebook or LinkedIn. Yeah.

That's what I'm saying. Yeah, and I agree with that. I think that- I got a better example for you.

Okay. I don't know if either of you collect sports memorabilia. I do.

Look at my- I see it behind me, yep. Look, I do too, and I've conditioned that I look for my hologram stickers, right? I don't know if they're easily full of crap or not, but having that hologram sticker on there with the stamp that says it's from this and that, it's authentic.

That was good enough for me. Maybe not for a court of law, but as a consumer- Yep ... that was good enough for me.

Is C2PA a consumer- Yes ... kind of level, or does it rise to chain of custody legal evidence, which is a much higher burden? Yeah.

It can be both, Alan. It can be both. And the interesting thing about that is there is, when you have that documentation of origin, chain of custody, where it's been, who's owned it, it increases the value of it.

It increases the value of that card or that memorabilia because you can associate where it came from. Right? And we're early in this stage of digital content where we don't necessarily have that.

We're working on getting that. And it is applicable from consumers. " And that's what we're talking about here, is that it's not just CTPA, but other bodies of media creation are getting-- There's enough pain in consumers asking for it that they need more transparency.

They need to be able to provide that chain of custody, and where it was originated, and what modifications have been made. And that's what these approaches are doing, and the world needs it sooner than later. I totally agree with that.

And I would just add, this is a start, but I look at it as a good start. It's a good start, a good example of industry coming together and thinking about these issues, which in my view, to date, hasn't really happened until CTPA, in a meaningful way. And I think there's obviously downsides, too.

I think it's important to note that, let's take an example of a CTPA chain that can be broken, right? If you have CTPA on an image or some content, and somebody snaps a screenshot of it, then obviously that screenshot's not going to have the original metadata. It's not going to pass.

And so there's a lot of questions as to who owns liability if that happens. All of that's not really been sussed out yet. So, I think a lot of really good solutions to a lot of this is you do a kind of a two-prong approach, where you do the provenance.

Working with provenance doesn't necessarily equal truth, but you also do the watermarking, because watermarking is hard to strip out. Particularly, we used that Claude example earlier. The way that that works, and I found this fascinating, so I'll just say it again.

This might be obvious to everybody, but essentially the way it works, as I understand it, is that the watermarking is essentially a pattern in the text. So in other words, some pattern with letters, some combination of letters that only the generative AI can assess as being generated by AI. Well, any kind of tester that has the formula would be able to spot.

That's the idea behind it, is that it's very easy to say, okay, yes, this was put through Claude or what have you. But again, even there, did Claude write the whole thing? Did Claude copy edit the thing?

There's so much that needs to be fleshed out here. Mm-hmm. Right?

And it really does come down to trust. I think it really also comes down to we're going through this scarlet letter phase, because there's going to come a time where everything is touched by AI. Yeah.

The case that Nima talked about, the screen capture, I think it's important to address that, because CTPA has thought about that. So one of the approaches, and this is technical, but one of the things that they do is they do a digital signature as soon as an image is captured. So on your phone, as soon as you capture that image, it digitally signs it with a trusted certificate.

That's for cryptography, and that's why Digicert's involved. And it's a signature on that. Any time it's modified beyond that, whether it's through Adobe Photoshop, through Apple's photo editing tools, it re-signs it as an edit.

And so there's a chain of custody of trust. And so if you screenshot, you originate another image that has-- and it would generate that as a new photo. It wouldn't be the same.

And so, the standard has thought through that. But I think it's important, Nima, to highlight, and I think you were saying this, there's still a lot of complexities that need to be worked out, but- And there are nuances here. But let me ask you both a question.

Should every photo I take on my phone have to have that, or only ones that maybe I'm going to use commercially or post online or something? That's more of a policy question. And I think, again, it goes to consumer preference.

I think if you're using photos in the commercial sense, I think there's a good argument that you need- Absolutely ... that provenance. Personal photos, I can also see a good argument that you need it there as well, because I've seen some really horrible stories about people, generative AI fake photos of people in embarrassing situations used to- Well, no, there's a whole cottage industry.

I don't know, Mike, how old your girls are. But there's a whole cottage industry of high school, college-age kids, where there's a term, I don't remember the term, but basically they put your head on a new body or something and- Yep. Very common practice ...

put that around the school. Well, it's illegal in a lot of jurisdictions now. It is.

Yep. And again, I forgot, there's a name for it. Deepfake, right?

I mean, deepfake videos, but deepfake is fine. No, it's a particular thing. It's not deepfake.

Deepfake is all, just everything can be deepfake. None of us are young and hip enough to know the terminology, but No, but there's a thing like where you're putting your head on someone else's body and, not pornography, but in lurid kind of photos. And to a young girl, a young guy, or any young person, to any person- Yep ...

it's a terrible thing. Yeah. We do need to combat this, right?

And- Yep anyway. Hey- And that's where I would say- I'm the one who always asks the hard questions here. I appreciate you both.

And that's where I would say, yes, a lot of consumers will be naive and not understand what's going on. But the minute you need it, you want it. Exactly.

And that's where I would say it's a better practice to do it than not do it. That transparency and chain of custody, knowing where photos were originated is really important. Those use cases you're talking about, as a dad, a protective parent, I want that.

I want to know- Absolutely ... I want to know if there's photos of my kids floating around there, or of myself, or of anybody that I care about. Well, they're not even photos of...

It's purported photos of you, right? Yeah, exactly. Fake photo.

Yeah. I wish I could remember the name of it. Anyway, guys, we're about out of time.

We could probably talk about this for two more hours, but we're going to have to pull the plug here. Nima, I want to thank you for coming on- Of course ... and joining the conversation.

For people who want to follow along what you're doing over at Sidley Austin, is there a website or something they can follow or something on social they can follow? com, but I'm also on LinkedIn. If anyone wants, I typically try to post about relevant topics in AI whenever I can.

So if folks are interested- Very cool ... happy to talk. Always happy to do that.

Mike, where can we go find out more about this? com. We have a lot of great resources, and if you have questions about this, we're happy to help.

Digicert's right in the center, and cryptography is at the heart of how you do that authentic chain of trust, and so I'm happy to have conversations with anybody who's interested. I love it. Hey, guys, thank you so much for spending your time with us.

I hope you've enjoyed it out here. Look, this is a topic that hits home for all of us. So interested in your thoughts.

Of course, you can always email us, comment on social media, depending where you're watching this. And thanks for joining us today. Mike, Nima, have a great day.

This is Alan Shimel. We're out. Thank you