AI Ethics: Data Usage & Privacy

[00:00:00] Doug Berger: Welcome to the latest installment of Brand of Brothers. I'm Doug.

[00:00:03] Johnny Diggz: And I'm Johnny. Today we're continuing our conversation about the Higgins-Berger scale of AI ethics. Today we're talking about data usage and privacy.

[00:00:11] Doug Berger: All right, let's get to it.

​

[00:00:22] Johnny Diggz: And we're back. So Doug, last time we talked about transparency Yes, we've spoken about transparency. Um, prior to that, uh, we talked about, uh- Potential for harm Potential for harm. Right, right, right. So the Higgins-Burger Scale of Generative AI Ethics, we've been talking about that, uh, for, for a few epis- Three s- Yeah

[00:00:44] Doug Berger: yeah, three shows so far, so this is our fourth. Yes, and today we're talking about data usage and privacy. Yes. This is a little bit, little bit ahead of where things, like, before you get to transparency, before you even get to use a, a, a LLM, it has to be trained on some data- Mm-hmm ... and that it has to get that data from somewhere.

[00:01:09] Johnny Diggz: Mm-hmm. So my first question for you is, as a creator, when you share something on the internet or wherever, um, w- what are you sharing? So that is a fascinating question because it depends on context, right? So there are times where I am sharing what I think is an image. There are times where I think what I'm sharing is a status update, but what I'm actually sharing are basically ones and zeros, right?

[00:01:52] Doug Berger: I am sharing data. So the real question becomes who owns that data at any given particular point in time? And so that's part of privacy, but that's also where provenance comes into play. And so those two pieces together kind of are, are, are the crux of what we're discussing today. So when it comes to LLMs and, and training models, um, the complication that we're facing is the prevalence of how much people are utilizing platforms like OpenAI, uh, specifically ChatGPT, to generate content, to generate, whether it's verbal or visual, right?

[00:02:42] Doug Berger: Um, and so that's really our key focus is on content generation. Um, when you're generating content through OpenAI, again, ChatGPT, uh, a- and really any of these others, right, Claude, Perplexity, you name it, these are not closed environments. These are very much open environments, so not only is there a privacy concern about what content you're reeling in, but also what content you're proliferating.

[00:03:18] Johnny Diggz: So, but I'm talking about from a creator's perspective, um, are you, when you, when you publish something- And, uh, let, let's say you publish it on, uh, X or Facebook, something public. X, is that like Twitter? Yes. You mean Threads. Threads, yes. Okay, gotcha. Um, uh, Blue Sky. Oh, okay. Perfect. Mastodon. Mastodon. Um, so are you...

[00:03:46] Johnny Diggz: Obviously, you're doing that if y- assuming you're doing it, uh, in a, with a, your settings to public, there's an assumption that you're giving permission for other people to read it. Mm-hmm. Um, and but are you giving permission for AI, uh, LLMs to train on that? Probably. So, uh, we know that Meta has their own AI platform, and probably by utilizing the platform, I am giving them permission to do whatever they wish with what I upload.

[00:04:20] Doug Berger: Um, does it mean that I'm relinquishing my copyright? No. But it does mean that I'm relinquishing- The license to utilize it however it, they wish to utilize it. Um, we, we know that the same is true for, for X, right? They have Grok. Mm-hmm. Um, so generally all of these platforms have some sort of AI baked into it that there's no doubt in my mind, I haven't researched this, but there's no doubt in my mind that somewhere in their terms you have to opt out of having your data being mined for their AI.

[00:05:00] Johnny Diggz: As a matter of fact, I think, uh, most recently there was some news, um, because I think it was Instagram, um, had, uh, enabled a new feature, um, where they, they could turn- they were s- they were, uh, uh, you had to opt out in order it, to stop them from training on anything that you upload to Instagram. Um, and there was a lot of pushback and, um, I think I just read that this week, within the last week or so, they, they reversed that.

[00:05:30] Johnny Diggz: Yeah. And so you're not automatically opted in. But I guess this is, this is where the crux is, is your, um, when you upload something, your content, um, is it... You know, h- how can that be used? And is there assumption that, um, if you publish a book, um, you know, i- t- does that mean that an AI can train on your book or on your artwork or on your music?

[00:06:04] Doug Berger: So it really depends on how you put out that information. So if you decide that you're going to voluntarily upload the content to a platform, that basically you're agreeing by utilizing the platform to their terms, and you're agreeing to hand over the keys to your car- Everybody reads those terms, right?

[00:06:31] Doug Berger: Oh, completely. Right? I... Not, not only do I read them word for word- ... I also have my team of attorneys read them. Yeah. And, uh, they're s- they're still supposed to get back to me on, on some- It, it- ... of these details. You know, it seems like every, every few months, um- That's not true. Somebody, somebody actually does read the terms, and then there's a little bit of a viral post about, uh, that everybody gets worried that now they're stealing all of your data.

[00:06:56] Johnny Diggz: And there are, the, there's some language that, that is in all of these platforms, especially the social media ones anyway, that- protects them so they can actually, you know, if you post something, they want the permission to be able to repost it, 'cause that's how social media works, right? Right. Yeah. They wanna put it in other people's feeds and stuff like that.

[00:07:16] Doug Berger: The key, the key here in all of it is indemnification. So when you as a creator are leveraging an LLM to create something, you are opening yourself up to risk, right? So when it comes to our ethics scale, it, it's super complicated, and we try to make it a little less murky, even when you're using Firefly, right?

[00:07:46] Doug Berger: So f- Adobe Firefly is for all intents and purposes the safest visual creation LLM that designers have access to because... I see you making a face. Did you want to jump in? Well, no, I just... Well, what makes it, what makes it... How, how, how do you define that it is, uh- Yes. The ethically sourced. Yes Right? Yeah. So that's really what it comes down to, is ethically sourced material.

[00:08:13] Johnny Diggz: That's what we're getting to, right? So we've been talking about, uh, creators uploading their content, and there's risk on that side. But on the other side, using the tools- Yeah ... now you're, you're, you're using the con- po- potentially the c- the, the intellectual power, prog- you know, everything that creators have u- Yeah

[00:08:35] Johnny Diggz: have created, the history of mankind. And building something new. So it d- to a limited extent, that, that could be true if you're using a more open space like ChatGPT. However, when you're using Adobe Firefly, the training data is more confined. So- How so? They are using images and graphics and designs that they have contracted.

[00:09:09] Doug Berger: So all of the material that is being utilized is licensed material, or they have also... Not or, and they also have, uh, open source material that has been input for training data as well. So technically it's ethically sourced, not ethically dubious. So they're, they have, um, a- any of the content that's, that's b- was being u- used by Adobe to build Firefly's LLM, their, their generative AI, they, um, e- they either, you know, th- the authors co- either actively- Consented

[00:09:53] Johnny Diggz: a- actively consented or, uh, were compensated. Mm-hmm. Um, they, they know that, that, that this is being done. Unless it was put out into, uh, into open source. Oh, exactly. Yes, yes. So the, I guess that makes, uh, that particular model, the Fire- Firefly model, m- m- uh, it, it would score a better or lower- Oh ... in the HB- than HBS.

[00:10:22] Doug Berger: Oh, for sure. I mean, this is definitely what I would say lands in the fully compliant category, so that would give you your lowest score. However, there's always a but, um, and this one is a big but. Um, and I cannot lie. So when it comes to having the, uh, the, um, the, the generative AI create for you, that you're- you may end up with something that resembles a famous individual, which opens you up to problems, right?

[00:11:00] Doug Berger: Because that's a likeness problem. And then there's also the fact that maybe you're trying to make it create a design, and it inadvertently creates an already existing trademarked design. Or it gets even crazier because in the world of video editing, you can not only trademark, but you can also patent aspects of effects.

[00:11:24] Doug Berger: And so if you're making AI-generated videos and it utilizes an effect that simulates a patented effect, you are opening yourself up to risk, to li- that legal liability. I think- And did I mention that we're not attorneys? Yeah. As much as, you know, we kind of sound like a law firm. Well, yeah. Well, I mean, we're talking about ethics.

[00:11:48] Johnny Diggz: The, the le- what is, what is legal versus ethical sometimes overlap, many times overlap. Mm-hmm. Um, that's why the laws get created because of ethics usually. Um, but, um, but we're, we're definitely focused on the ethical side rather than, uh, recommending anything legal. Um, but- But can we, can we talk a little bit more about these LLMs and how it is possible to have your own local LLM and to train it- Sure

[00:12:21] Doug Berger: and how there are actually organizations out there- Sure ... who are doing this? Yeah. Because it, it, before we get into the, the dark- ... a- a- and, and, and nebulous side of, uh, of non-consensual or restricted use I'd like to kind of- Mm-hmm ... talk about how, um, there are organizations that are trying to make this as ethically viable as possible, and that includes energy consumption, right?

[00:12:50] Doug Berger: So all of a sudden, the ethical component of environment, which we, we've obviously moved away from, uh, environmental impact, but these, these, uh, e- LLMs that you can run locally, uh, uh, are, are definitely shifting the conversation in a more favorable direction. There, there are. There's, there, um, there are organizations that are both, uh, advocating and they're companies that are building, um, ethically sourced LLMs.

[00:13:26] Johnny Diggz: Um, the, none of these are the mainstream ones, uh, that you've heard of, but, uh... And, and I, I wish I could think of the name of any of them right now. I guess I could pull them up if I wanted to look it up. That's fine. But, but anybody can, anybody can look up, um, it, it just, you know, search for ethically s- ethically sourced, uh, LLMs.

[00:13:46] Johnny Diggz: Oh, we can put it in our notes. Yeah, we can add it to the notes. Um, but- Uh, so and then you can, you can, you know, in- install these models locally. Mm-hmm. Um, and there are different, uh, degrees of, of whether or not the... You know, I don't think there's any, any, uh, of the major LLMs that have, you know, clean records of where they, they not only have, uh, consent, they have compensated, they have, uh, they have a clean, um, uh, trail back to- Right.

[00:14:27] Johnny Diggz: Provenance ... pro- yes. Yes. Um, and, and so, um, none of them have that. So the, any of the, you know, whether you're talking Claude, Perplexity, uh, OpenAI, Deepsea, any, any Grok, they all have ethical issues with how their data was sourced, and there are literally hundreds of lawsuits going on right now about this very topic- Right

[00:14:52] Johnny Diggz: from, uh, uh, creators of- of- of we're talking writers, we're talking, uh, record companies, we're talking anybody who ha- you know, uh, uh, uh, artists, a- a- a- every- Mm-hmm ... anyone who has been a creator. It's li- it's limitless. Yeah, exactly. Yeah. Yeah. And so, and, and, and these lawsuits, um, are not gonna get s- settled quickly.

[00:15:13] Johnny Diggz: Some of them are. Some, some, like, uh, Suno has been working out, uh, agreements with some of the record companies and changing h- how their, their, their, you know, moving from the old data- Right ... to a new data, and- A- and now there's a, a new thing where, uh, they, uh... we're gonna be seeing an AI label, and, uh, there are two different- On the record

[00:15:35] Doug Berger: AI la- labels. Yeah, yeah. Um, so there's going to be an AI label, uh- From the RIAA. Yeah. It, it... Well, you know, obviously we're gonna... All artists w- will end up being compliant because the distributors are going to apply these labels. Sure. So, uh, you know, it, like DistroKid, for example, is going to very likely be applying these AI labels based on what their artists are contributing, for example.

[00:16:05] Johnny Diggz: Yeah, and how, and, and, yeah, whether it was completely AI generated or partially- Mm-hmm ... you know, there's human in the loop sort of thing. Mm-hmm. Um, and, you know, that, that we're seeing some of those aspects of transparency, then that would, that would be much more of a, a- That was two episodes ago ... that was two episodes ago, so go back and, and watch that one.

[00:16:25] Johnny Diggz: Um- And like and subscribe. Um, so but, uh, to get back to where w- w- the, the- You had told me... Sorry to interrupt- Yeah, yeah ... but you had told, you had, you had, uh, made this analogy about a cake and how there are basically- Mm ... three different types of cakes, right? There's the store-bought cake- And then there- The restaurant There's the restaurant cake Experience.

[00:16:51] Johnny Diggz: It- And then, and then there's, there's make your own cake. Right. So, so talk to me about what that means and how that's relevant to the conversation. So the idea goes like this. If you go into a restaurant and you have a piece of cake, um, you can't take that cake home. You don't know what ingredients are in it, and but you can still enjoy it.

[00:17:11] Johnny Diggz: And so this is like using, uh, Chat, ChatGPT today. You don't know w- what's, how, how it got there. It's still delicious. Um, and, uh, so any of your major, uh, platforms, OpenAI, Claude, Anthropic's, uh, uh, any of those are- l- existing like you're, you're eating your cake at the store. You have no idea w- uh, what the ingredients are, how it was made, um, uh, how good it is for you.

[00:17:42] Johnny Diggz: But it's still gonna be delicious. Now, the second option is, uh, a cake that you buy at the store, you bring home, you can still enjoy it. Um, but you still don't know what's in it, right? You still don't know- It's just cake mix ... it's just, yeah, it's just cake mix. But you get to decorate it, so you know- Yeah, mm-hmm

[00:17:57] Doug Berger: what frosting was used. Yes. Yes? Yeah. Sure. So, so t- t- what, take me through this a little bit more, a little deeper. It... Go ahead. Go ahead. No. Well, and then the th- the third one is, you know, you make cake from scratch, and you know exactly everything that's, that's in it, and you build your cake. And that, that, that would be the, the, the highest amount of provenance.

[00:18:22] Doug Berger: I see. I see. So, so basically you've got your OpenAI's, uh, ChatGPT type of thing, and that's your restaurant cake. Yep. Then you have a, let's say that you grab your own LLM that you're operating locally- For example, like Llama- But- We're talking about, like, Llama or something you can install on a l- a local machine.

[00:18:43] Doug Berger: But you don't know how it was trained. You just know that you can operate it locally. Correct. Correct. Got it. Okay. Which, yeah- I see ... which is, which is, you know, we haven't really gotten into the privacy side of things, but that is really where you wanna be to help maintain a le- a higher level of privacy with the data that you are, uh, exposing, your own data.

[00:19:07] Johnny Diggz: So this is, you know... Uh, do we wanna get into privacy yet? But- So I, I, I don't know how, how deep we wanna get into privacy- Right ... I, because it's so incredibly nuanced. Um, so a- as, as I see this, uh, so i- in the world of LLMs, we have open source- Mm-hmm ... and we have open weights. Yes. Right? Yes, yes. So when it comes to the LLM that you are downloading and installing, but it already has its training infused- Right

[00:19:39] Doug Berger: you're dealing with something that is open source, but it's not open weights, right? Exactly. It- So it, it, basically that, that comes down to, um, it gives you, um, s- it, it gives you th- sort of the finished data model. Yeah. But, um, but you're, it's still processing locally. It's still generating the text and doing all of the thinking locally.

[00:20:09] Johnny Diggz: But it, but we're- ... we're still kind of in an ethically dubious space, right? Specifically- Because we don't- ... because of the data, that you don't know where the- Yeah, we don't know where the data- Yeah. Y- Okay ... it could be- Again, provenance. Yeah, yeah. That could be, you know, um, th- it could've been trained on Copyrighted material Right And so, um, versus building it from scratch your own data or ethically sourced data from sources that you know, um, that, you know, there was consent and, and compensation and all of, all of the things that, that you need, that would be the only way.

[00:20:47] Johnny Diggz: And they're, they're really- And that's where you're making the cake entirely yourself. Yeah. Yeah, yeah, yeah. And, and can we, can we take the analogy- And you can... I mean, there are a couple companies out there that are, that are, that are offering this- Yeah ... um, but they, they're very, um, they're very few and far between and, um, and expensive.

[00:21:08] Doug Berger: But they can also take you to that next level. So we've talked about there's a restaurant-bought cake, there's a store-bought cake, and then there's the homemade cake, but then you can take that homemade cake to the next level, which is you can now get certification. Oh, yeah, yeah, yeah. It- There's organizations that will certify your cake.

[00:21:30] Doug Berger: Yes. So well, much like you can have a certified organic cake, for example- Sure. Yeah ... or certified gluten-free cake or- Yeah. Well- ... or vegan or whatever, um, i- in this instance, it can be certified, uh, ethical based on data use and privacy. Now, privacy. Let's go back to pri- Yeah ... we, we had briefly touched on privacy.

[00:21:52] Doug Berger: Sure. I, again, don't wanna get too deep into it, but obviously the ethical nature of this homemade LLM Becomes thrown into question when you begin importing your content into it. So obviously we're staying local, so you're not really proliferating, but if you're training it with private or, or sensitive data, then that is going to impact anything else you're doing locally.

[00:22:32] Johnny Diggz: Yes. Um, privacy is, is not so much about secrecy. It's more about the expectation of what is going to... Where, where your data is being used. Right. Um, and especially when you're in, uh, in a creative environment like an agency and you might have access to client information, client missions, and what their goals are, and they, you know, like all of these, these can be, you know, competitive secrets- Mm-hmm

[00:23:06] Johnny Diggz: that if they're launching a new brand or whatever, um, that, uh, if you're using a LLM to assist you with, um, let's say even something as simple as, uh, you know, a, a br- a brand evaluation or some- something like that, um- You're presumably gonna have to upload some of this client data to get that, that analysis back.

[00:23:33] Johnny Diggz: Mm-hmm. And so if you're using a public chat, like- Oh, well that opens you up even more. Yeah. Yeah, yeah. So, uh, so talk about that. Where, where, where do you draw... Where, where, where are those ethical boundaries that, that, that you, that you see as potential pitfalls as an agency owner? So when it comes to the training data, um, i- specifically when you're using, uh, something like ChatGPT, ChatGPT says that on the commercial side of things, that the chats are basically closed chats, and that you have control over it to a, a limited extent.

[00:24:15] Doug Berger: Let's be realistic. We don't know how true that statement is. The only way that you can begin to control the, the sensitive data is by having local LLM instances that are closed per client. And we know that that is cost-prohibitive, um, a- Mm-hmm ... a- if, if not completely untenable for most businesses. Don't, don...

[00:24:45] Johnny Diggz: I know I haven't really played with it much, but I know that, like, for example, OpenAI has, like, a business, uh, version of, of OpenAI. Like, is... Do- Yeah. Do use that. Okay. So, and th- that has certain restrictions on- Allegedly ... allegedly, right. So they, they claim that they're not sharing that data or using- Yeah

[00:25:07] Johnny Diggz: training on that data to build future LLMs, right? Listen, I, I don't mean to change the subject. But at the end of the day, it's still gonna be that 80/20 rule, um, where m- minimally there has to be 20% human-in-the-loop involvement. Otherwise, you are still opening up yourself to risk, because you've gotta make sure that you're not violating someone's copyright.

[00:25:34] Johnny Diggz: Right. And you've got to make sure especially that you're not putting your clients at risk, regardless of whatever indemnification E&O insurance you may have, liability insurance. Right. As far as I'm aware, I don't believe any of the liability insurance that Remixed carries covers presenting AI-generated work as our own, which is why we don't do it.

[00:25:59] Johnny Diggz: Right. And so when it comes to utilizing AI-generated work, there's a, a degree of transparency and a degree of inclusion, right? Um, a- and we were talking about this earlier, that- I could dump 20 years' worth of data into a local LLM, but w- we didn't do every single element that you see in these brochures and in these flyers Some, some of them were licensed by third parties and, and- Oh

[00:26:31] Johnny Diggz: have limited scope of their license. And it's not just limited to photography, right? Right. It's, it, it's also typography. Sure. So- Fonts and stuff. Yeah ... yeah. I- if, if- She did a great episode about fonts a- about a year ago. Nice pitch. Um, a- and so with regard... Or plug. W- with regard to, to typography, you know, you can't just...

[00:26:53] Doug Berger: You can actually have ChatGPT create something using the font Gotham. But if, for example, you don't have a seat license for that font It's possible that Jonathan Heffler is gonna come knocking on your door. It's not probable. But if you're big enough and you're using their IP and it's not licensed, then you are moving into not just ethical gray areas, but questionable or unclear provenance.

[00:27:27] Johnny Diggz: Right? The, the, um, you know, it, it really comes back down to you have to look at it not only, you know, can you legally do something, but should you ethically- Yeah ... do something. So even though you might get away with it, um, are you just getting away with something? Uh, listen, I, I think we have this conversation time and again- Yeah

[00:27:50] Doug Berger: which is you have to have ethical boundaries to care about ethics. So if you don't care- ... this doesn't matter. Right. So this is really, uh, uh, uh, the, the HBS, the Higgins Burger Scale, uh, the utility especially, is meant for people who care. And so the, the real goal here is that they're not, uh, trying to be deceptive.

[00:28:21] Doug Berger: They're not trying to use restricted use material. They're not trying to, uh, circumvent the, uh, consent of other creators, um, because that would be the, the ultimate negative when it comes to, uh, to data usage is to just basically say F you to, uh, to other content creators when you yourself are a content creator.

[00:28:46] Johnny Diggz: Right. And s- and, you know, which is also why this is, this is a framework and a scale. It's not a y- it's not a black and white thing, right? It's, it's, it's, you know, we, we have these number values that we try to assign with the HBS that address the fact that we know that sometimes it is going to be a little gray because you can't, you can't know that it was ethically sourced or- A- a- and, and that, and that's when it's important for creators to go, "Okay, whatever I'm doing with AI, it's, it's strictly assistive."

[00:29:23] Doug Berger: You know? Right. It's not going to be the penultimate- Or even the ultimate, uh, concept. It'll be maybe it's something that's used for ideation, uh, maybe you're synthesizing focus group data, right? So you can do a lot of things that are ethically, uh, bound that are not violating people's IP, and that's really what this is all about is not violating other people's IP.

[00:29:57] Johnny Diggz: Yeah, and but on the flip side of that, if, if we require, you know, let's say, let's say, you know, five years from now there are laws in place that require every creator to actively opt in to allow their p- public creation to be used for training data. Mm. Does that, does that stifle innovation? Does that, does that somehow- I, I don't think-

[00:30:23] Doug Berger: cripple us? So I don't think it does because I think everything is built on something before it. And, and when it comes to innovation, it's still going to be humans that are going to be innovating. We know currently that AI is not really capable of innovating. Again, it's that 80% of the way concept. You're...

[00:30:44] Doug Berger: It- an AI is only as good as its operator. I've yet to see AI smarter than the operator when the operator has above a 100 IQ, let's say. Obviously, it's possible- Yeah ... for you to ask AI something that is the, a- akin to a search query, but, you know, the, the response that you get back, if you don't know what you're reading, if you're not intelligent enough about what you're reading, all the sudden the AI becomes the expert, and that is a slippery slope.

[00:31:18] Johnny Diggz: Well, we know that it's wrong at least 20% of the time. At least. You're, you're being so kind. Um, w- speaking of the operator, uh, so our, our next episode we're gonna be talking about- Displacement impact ... human displacement. Yes. So, um, let's tease that a little bit about w- what does that mean? The, the, the ability, the, the, the...

[00:31:43] Johnny Diggz: We're seeing, uh, a lot of companies announcing layoffs. Yeah. And blaming AI. Yeah. I mean- Attributing AI ... I mean, that's what this is all about, right? Right. So it, it is about how does AI, uh, impact human labor. And I think that's a good place to, uh, to stop and continue next time. Um, and so make sure that you guys like and subscribe.

[00:32:09] Johnny Diggz: I love some of the comments we've been getting. I've been seeing a few. Uh, keep 'em up and, uh, especially you, uh, Calvin. Um- And, and we'll see you next time

 

[00:32:20] Johnny Diggz: Thank you for tuning in to Brand of Brothers. Big thank you to our presenting sponsor, Remixed, the branding agency, along with production assistance from Johnny Diggz, Simon Jacobsohn, and me, Doug Berger. We can't forget music by PRO. Speaking of not forgetting, remember to do that like and subscribe thing and find us at BrandShowLive.

[00:32:34] Johnny Diggz: com and follow us on the socials at BrandShowLive.