Kelly Wortham: Hello, everybody. For all of those of you who are joining us later, let's take a minute. Even though we've been chatting for a bit, let's take a minute to talk about who we are, why we're here, and what we hope to accomplish today. But first, like always, I want to thank our sponsors. So let's start by thanking Conductrix. They've been with us from the very beginning. Thank you, Matt, Nate, Ezekiel, Ashley, and the entire team at Conductrix. So, so much for sponsoring the tlc. We'd also like to thank our other TLC newsletter sponsors, including Conductrix, Growth Book and Ask why. Thank you all so much for providing your thought leadership. The community really benefits from that, from that conversation. And all three of you, thank you for sponsoring Experimentation island and thank you to Convert for sponsoring the Amazing Conversations Channel. If you haven't taken a chance to go read some of those questions and answer them, get to know your peers in the industry better. I highly encourage you to do so. So you might win a prize, but along the way, you get to learn more about others and they get to learn about you. So go do that. And last but not least, thank you, Eddie Agalar and blazing growth. So really quick. This is it. Open ticket sales are now live while. While they last. So as long as they last, you can have a chance to come to Experimentation island. Just go to experimentationisland.com and request your ticket. And, you know, if we like you, we think you'll make a good mix. We'll approve you, so hope you can make it because it's. It's going to be a lot of fun. All right, let's talk about why we are here. Oh, that still says agentic commerce. God, I have to fix it. Everybody wait for just a second.

Slobodan Manić: Meet your Camilla. Camilla is still Carlos.

Kelly Wortham: No, it's not.

Iqbal Ali: No, it's not. It's all right.

Slobodan Manić: Tlc. TLC username.

Kelly Wortham: Are you serious?

Camila Dutzig: I was wondering about that, but I'm okay with it. I think it's kind of like both. I'm not sure if they're both actually Latin names, but they sound Latin to me.

Kelly Wortham: That's amazing. Okay. It should say Camilla. Your new name is Camilla. It's just your name, isn't it? And I love how it's different color. You guys, real time here. Real time. All right, There we go. Feel great about this. I feel really good about this recording. So who are we and why are we here? And in all of that, I also put it full screen. So now I don't have My notes. Oh my gosh, people, y' all are being the best. Let's do this. There we go.

Camila Dutzig: All right.

Kelly Wortham: So as all of you know, we're living through this paradox and a digital shift, not just in experimentation, but also in experimentation. We have more power than ever to personalize, optimize experiences. And at the same time, same time, we are facing harder questions about responsibility, trust and restraint. So today is not about perfect answers. It's about like asking the right questions and the decisions that you all as brands and practitioners are being asked to make every day. So most of you, I hope, have heard about some of the amazing work that Sonny's been doing around AI, visibility and trust. And you know, he discovered this real world, real world AI prompt data exposure. And so I reached out to him actually after our last conversation with Namrata and said, hey, we need to do more on this. It was really, the discussion after the fact was, was so interesting, let's do more. And will you be my co host again?

Camila Dutzig: And he said yes.

Kelly Wortham: So he's going to talk a little bit about trust failure and maybe the public react into what he discovered. Iqbal, I think everybody knows Iqbal, but he talks all the time about responsible AI and algorithmic fairness. So that makes sense to invite Iqbal. He has really deep expertise in how bias and ethics show up inside systems and experiments. And then Camila, I had the honor of seeing Camila talk at Conversion Hotel and she is a experimentation leader in a will say sensitive type job with sensitive data. And so bringing that perspective of this like privacy first and highly sensitive industries, it really made sense to invite Camilla. So thank you all for joining. But before we fully dive in to the conversation, let's start and each of you can add a little bit more about why you think you're here as you answer these questions. But let's start about like when we hear the phrase ethics in experimentation, what does it mean to you in practice? And let me start with Sonny.

Slobodan Manić: So sorry, my daughter called me from the other room, she wants to get here and take some Nutella. Sorry, I had to go off screen for a second.

Kelly Wortham: So ethics, Nutella really matters.

Slobodan Manić: I know, I know, I'm sorry about that. But yes, there she is. Ethics and expertise. If you, if you have a line. And this is where you're testing to improve your product, and this is where you're testing to manipulate users into buying your shitty product. Ethics is somewhere here and it's as vague as it can be. I know, but ethics is something that is not in the room when you're experimenting unless you invite ethics into the room. So it's not going to just show up on its own. And, yeah, that's how complicated it is. That's how complicated I think it is. Also, maybe it's hot take, but I don't think you can be fully ethical when you're experimenting on people without them knowing on a web.

Kelly Wortham: Ooh, hot take. All right, Iqbal.

Iqbal Ali: So I'm gonna quote quite heavily from Olivia Gamblin's book Responsible AI throughout, so I'll make a start with the definition as well. So the way she defines ethics is kind of like a moral compass that's used to align decision making according to a set of values. And those values could be stuff like fairness, sustainability, privacy, transparency. So all of those are specific components of ethics, and each of those have a sliding scale, and every single one of us have got a different sort of position on that sliding scale. And. Yeah, so ethics is like that moral compass to help drive decision making.

Kelly Wortham: Perfect, Camilla.

Camila Dutzig: Yeah, very interesting definitions to begin with. I will maybe start by saying that although I fully can understand the Sunny's point of view, it cannot be fully ethical or it's hard to be fully ethical. I definitely think that it is possible, it is definitely possible for you to optimize for the right things, but it has to be invited to the conversation, like you so cleverly said. And I do work in experimentation in the experimentation team, but actually by training, by, you know, my area is actually UX design. And so I tend to see ethical ethics as a. As a design issue, as a system design issue. Right. It is hardly that people act in unethical ways on purpose, but it's rather like, what is the system around you? What is it built for? What are you optimizing for? Because then ethical considerations become an afterthought rather than embedded into your kind of like, default mode in which you operate. So I tend to see it as a. As a design experiment on its own.

Kelly Wortham: All right, thank you all. Our first topic is about why ethics matters right now. And let's frame the conversation, if you will. Sonny, can you walk us through what you discovered with the AI chat prompt exposure?

Slobodan Manić: And, like, preface it by saying, it wasn't just me, it was Jason Packer and another friend. Uh, that. And we discovered something strange happening in that friend's Google search console. We saw entire prompts from ChatGPT, entire, hey, my boyfriend doesn't like me. What should I do? Those kind of prompts. We dug around a little bit and we figured out how it's happening. And then Ars Technica got involved and their investigative reporter figured out that reached out actually to OpenAI and to Google confirmed this was really happening. OpenAI said it's not a big deal. Like it's just a few prompts, just a few accounts, which means every single account Google confirmed it was happening as well. OpenAI said they fixed it, but they haven't. So what that. Well, it didn't make me realize because it was obvious for OpenAI and those big tech companies, they don't care about privacy, they don't care about the users at all. They never will. By definition. Their job is not to care about the users. So ethics is for us to deal with and not for them.

Kelly Wortham: But aren't they just another brand and product and.

Slobodan Manić: I guess yes. But what is their business mission and what are they trying to do? Are they trying to make our lives better or something else? Wow.

Audience Member: It wasn't the only leak. There was a leak several weeks before that where everyone's private chats were getting leaked on Google search.

Slobodan Manić: Yes, but those were opt in because people. It was a dark pattern, but people checked the box and it was a.

Audience Member: Dark pattern, which is. Yeah, it was really bad. People didn't realize they were sharing them.

Slobodan Manić: Absolutely, I agree with that. It was opt in. This was someone's random prompt in some random Google search console which, like, I don't know how that can happen.

Audience Member: Which one is worse? Both of them were really bad.

Iqbal Ali: Absolutely.

Slobodan Manić: I 100% agree. But yeah, it's not the first one. I think it was the first one that someone prove that they're using Google to search the web. So we always have that.

Kelly Wortham: So Iqbal, if, you know, from what Sonny shared, like this is. And he, he, he believes it's our responsibility but like from a systems responsible AI perspective, like what gaps are exposed about how, how system AI systems are being built by companies like Open AI and Google and, and like how can we govern the data they generate? Or can we. Or should we have to? Or should they have to. Or like how does this. Are these like moving too fast?

Iqbal Ali: Yeah, I mean, I think there is an element of like the innovation and this, this whole thing that I personally hate, which is, oh, ask for forgiveness later than ask for permission. And I think we just say they.

Kelly Wortham: Can ask from our fault because we said quick ask and break things.

Iqbal Ali: Yeah. And I think the way it's supposed to work is that these companies are supposed to have a responsible AI department that are supposed to lay out some guidelines and rules to kind of determine what they. What they can and can't do. There should be some frameworks there, for instance, for a value like privacy, where they. And I look at it in terms of like a sliding scale. So on one side you've got completely safe, no issue, and then on the other side you have illegal. This is really, really illegal. Don't do it, you're going to go to prison. And then somewhere along between that, there's a line and there's a sliding scale in terms of where we, each one of us would put ourselves. And what these companies are supposed to have are some guidelines to say, here's our line. Here's the framework from which to determine whether or not we are partaking in unethical practices. This is what we should be doing. And it seems like what's happening is because of this whole, oh, yeah, yeah, AI Innovation, innovation. We need to move fast. Let's ask for forgiveness later. They're just kind of sacrificing a lot of those ethical principles. And then that causes what's. What's called ethical drift in terms of, like, you make a little concession that builds, that builds, that builds until you have a major gaping sort of failure like, like we saw here. I mean, that's one way of looking at it. But it's basically their responsible AI system, if they even have one, is. Has failed. That is essentially what I think has happened here.

Slobodan Manić: One thing we should be aware of, these are the same or same group of companies that was responsible for Cambridge Analytica and God knows how many other examples. We cannot expect them to suddenly become ethical because it's probably not going to happen.

Audience Member: They're also, they're also trying to bypass this by buying political influence. They recently open. AI, recently hired George Osborne.

Slobodan Manić: I almost didn't mention that. But my favorite example is no sex chatbots for kids law in California. That was supposed to be passed and they were just supposed to say, just make it not do this to children, not harm children. Instead, they lobbied with Gavin Newsom and he vetoed the bill. Yeah, it made me angry.

Audience Member: This lobbying is becoming a big, A big part of their playbook. They don't need to worry about ethics if they don't have, as someone just mentioned, the accountability factor through regulatory oversight.

Iqbal Ali: Yeah, so there, there is like, there's supposed to be an organization called the. Well, there is an organization called the oecd, which is an organization for economic cooperation and development. And what they are, is like an independent international forum for setting global standards. So what they, what they are supposed to do is set Some standards for AI governance, for all of these sort of values. What are the values that we're going to have and what are the rules that we're supposed to follow? And then what we're supposed to have is like a government level sort of set of rules. These are a government level, these are the rules you should follow. Then at an organizational level, these are the rules they should follow. Sorry, not on an organization, industry level. And then at an organizational level and then at perhaps a team level and then there's an independent, like an individual level. So there's different sort of set of ethical guidelines that exist in each one of those sectors and they are supposed to be well defined at the government level, at the industry level and at the organizational level. But sometimes they're not. And so yeah, I think there is an element of self regulation where at the moment, because, you know, they're wanting to go fast and because the whole AI boom and everything, that the, that there's no real, you know, proper regulation happening at the, at the government level because these things move too slowly. Whereas in an individual level we're able to shift very quickly with, with like what our moral position should be.

Slobodan Manić: But also what Camilla said.

Iqbal Ali: Yeah, sorry, yeah, go ahead.

Camila Dutzig: You said my name, so.

Slobodan Manić: No, what you said that this needs to be designed. System that controls, all of these things need to be designed. So that's what I was going to say.

Kelly Wortham: Yeah.

Camila Dutzig: There is also like the EU has AI acts that specify like there is a governmental attempt there that of course like, you know, was not established like early enough and now things are already rolling. But it does establish that there's a certain guidelines that it has to follow is very extensive and but it's starts with pretty much saying how that AI softwares, models, whatever you, whatever it is, they have to specify clearly how are they working to begin with, so how are they getting their data, how are they processing their data, what were they trained on? So then when you use them you can understand the type of, and how accurate they might be so that you in advance as a user can already kind of like assess the risk. Right. Because it is a responsibility, of course, of the businesses and the companies that are creating these things. It is the responsibility of the government and it is kind of like really, really sad that they bypass all these rules. So often, like many things, even a seemingly harmless piece of innovation can also be used for bad things. Right. Even a car can become something really dangerous for somebody that doesn't know how to drive it. Right. So by design they are not Safe. But we also have the responsibility of actually, like, what are the things that we need to be aware of? What are the things that we need to do? And do them consistently and systematically to avoid us being kind of like fooled and misusing AI as well. Because then we are the ones that taking the technology and actually deploy it in our websites, in our services and our products.

Kelly Wortham: Right. Lucas, you have a question?

Lucas: I want to respond to something Iqbal said and it's actually related to what Camilla is saying. I'm not sure the problem is that the law isn't following up fast enough. Iqbal, you seem to suggest that the world is changing too fast and the law can't keep up. But you said earlier, like, move fast. Move fast and break things for some companies includes breaking the law.

Iqbal Ali: Yeah. Not necessarily that.

Lucas: Even the laws that exist and have existed for a very long time are being broken. So it's not the speed of the law that's a problem. And David said something interesting in the chat, like companies are going to try to find ways to maximize profit. And I think it was you in the beginning who said that the ethics is fuzzy or. No, Sani said that ethics is fuzzy.

Slobodan Manić: Right.

Lucas: So the moment you draw a line and say this is the rule, the rule is guaranteed to be wrong because the ethics are fuzzy. And there's always exceptions. And so then the companies are going to try and find the exception of trying to find the edges of that and they will do things that are unethical or do we consider to be unethical?

Kelly Wortham: Bob and weave.

Lucas: I'm trying to wonder whether this is an unsolvable problem.

Audience Member: Right.

Slobodan Manić: Well, there is like Detroit, Communism and Soviet Union. I think they have a way of solving these problems. I'm not saying that's the way for everyone, but there's a way to solve every problem.

Audience Member: Gulag Siberia. A trip to Siberia.

Camila Dutzig: Have you guys tried drinking as well? I find the distance to help.

Slobodan Manić: I think they tried the ketamine.

Kelly Wortham: Dear.

Slobodan Manić: Sonny, but it doesn't work.

Audience Member: Some regulation has got to be better than no regulation. We definitely have better regulation of AI stuff or consequences of it in the EU than the USA has. It's a start. It's not perfect.

Iqbal Ali: I think like to the ethics is fuzzy aspect. I think there is an element of subjectivity in ethics, definitely. But there is some area of objectivity there in terms of we all agree that X is wrong or we all agree that this is illegal. Now, when it comes to, you know, like what I was saying before, like going fast and breaking rules and stuff, it's kind of like, and this comes to the concept of ethical drift, where you kind of, or the concept of slippery slope, you kind of give in on little small aspect, kind of go, yeah, but you know, we could do this. It's not a huge deal, but then the impact of that small change because of this technology that is so powerful and so sort of influential, that could be big. And then as you have more and more of these small sort of, sort of shortcuts that you make, you build up a sort of like a technical debt or sort of ethical debt, which then turns into something much more mountainous. So I guess like, and, and having some sort of rules or having laws to govern, specifically a very technical area where, you know, you make such and such change to a, a model, you need to follow these principles. That sort of thing I don't think is as firmed up as it can.

Kelly Wortham: Be at the better, like frameworks and guidelines around. If doing a, then need BCD before publishing, for example, or exposing to us, we humans who are going to be impacted by it.

Iqbal Ali: Yeah. And as an example, I'm sorry, this will be the last point on this, but as an example, the very recent LinkedIn algorithm thing where it's biasing, making gender biases. Now, the issue going into it is that they actually don't have gender in the model at all. So what the model is doing is it is biasing or it is favoring certain types of conversations or certain types of posts which just happen to align with a certain gender, a certain stereotypical sort of group which happens to be white men. So, and the problem is, is that they kind of LinkedIn kind of go, well, we weren't. Gender was never part of the algorithm, but gender didn't need to be part of that algorithm. And because Geldra isn't part of the equation at all, they don't have any metrics in, in place to make sure they're not analyzing. Exactly. And this is, this goes to Olivia Gamblin's point in, in her book, and she makes some really prescient points there, which have just come up in that LinkedIn bias scandal where she kind of says, well, you know, even if you may not have any biases going into it, you need to put some metrics in place to make sure that, you know, there are no biases when it actually goes out.

Kelly Wortham: Yeah, that's, that's brilliant, Camille. Oh, you're already going to say something, but I have a question for you when you're done. Go ahead, Camille.

Camila Dutzig: This is by far I think the, the biggest misunderstanding when it comes to data privacy right now because I think we all come from this understanding or like majority of us have reached a certain level of understanding of like, kind of like the GDPR related data privacy that we require, right? That it needs to be anonymized you all these things but in the current state status, the biggest threat is no longer only right? This is still a threat and this is still very serious. Only identifying an individual person because of the data, but because of AI systems is taking all this seemingly unrelated data that comes from perhaps many different parts of your product or your system or moments in which you share the data that you never thought that the AI would actually be able to take some unrelated metrics and actually turn them into patterns and use these patterns to actually then you know, exploit people's vulnerabilities without it actually being by design or by default because AI has the ability so if you give it the control over the end experience it might as well do that things that humans would not be able to actually concretely plan. Right. It would take us more time. We would have to like look at all the data, make sure that we're asking the right questions to actually be able to influence this kind of behavior. And in this process we will probably have the time to actually wonder about the consequences of what we're doing. But in the speed of it, if you give the AI the control that is the real issue with the data, right? It's no longer identifying people but forming patterns at a level that we couldn't before and actually then exploiting these patterns to the gain of the system.

Slobodan Manić: So which is why I said ethics are not in the room unless we bring them into the room. I mean every, even before AI. Sorry Kelly.

Kelly Wortham: No, 100% I, I, the question I wanted to ask you, you kind of touched on it a little bit but, and you're welcome to share as little or as not or as much as you want. But in privacy First Industries we'll say you're dealing with information that well frankly it can carry a stigma. Definitely cultural sensitivity like we were talking about before. Who's true north is, who's true's north and definitely regulatory scrutiny. So how do you recommend we approach experimentation, optimization, personalization, like all of those things differently when you know that the data you're dealing with itself, like in Iqbal's example, they were very careful. Well we're not going to collect that data because that data is sensitive. But in fact if you ignore that data, you're running the Risk of optimizing towards a predictor, a leading indicator of that data and then you're biasing the system.

Camila Dutzig: Yeah, there was a lot of layers to this and there is like a multitude of things that can be done and that should be done. But like maybe just to touch, kind of like the most high level ones that I think are often also the most impactful is that we kind of like in our experimentation program we took a step back first right in advance and we thought and we defined what are the principles, the themes, the tactics that we will use that we will explore and we will not experiment towards. And we tend to focus on things that are usability, accessibility, clear communication. We one of like our primary kind of like signals to say, hey, we should experiment on this is actually listening to people's feedback and trying to then solve their actual problems. So the things that people are saying, hey, I am suffering with this, I don't like this. So basically it's our kind of like attempt that by design. We're not simply using tactics of, you know, persuasion or urgency or things like this to like, to like lead towards growth, but we're trying to lead towards growth by solving their actual problems, helping them like not persuade them to make decisions, but like helping them understand the.

Kelly Wortham: Decision that they're making, easier for them to make decisions. So by focusing on accessibility and usability, the end result is growth.

Camila Dutzig: Exactly, exactly. So we have a bunch like for us to create an experiment in the team, there is a few metrics that are required and is to say like what is the user feedback? What are in the themes that we selected by Advanced that we know that our users like, you know, that is really important to them. Privacy being one of them is kind of like already a default kind of, you know, pre check mark. Let's say that you have to say, hey, how does this impact privacy? So we keep it in mind by design and we don't think about it later. So I think that this is a great way to get started. But it doesn't just. But it's not just that. Right.

Kelly Wortham: Like I think clarify when you say we don't think about it later, it's because you don't have to, because you've planned ahead, you're being proactive instead of reactive.

Camila Dutzig: We still, we still of course do it right, but it, but it becomes harder. I think that is the, the biggest misconception in my opinion about ethics is that we tend to think that the people that do unethical things often do it on purpose. Right. That they are not good people or they did it because they, whatever it might be, but I think you guys even said it before, is often this kind of like seemingly small risk that becomes bigger and bigger or that your morality changes or people don't understand what they're doing, you know, they're giving their data maybe to ChatGPT, they don't understand it's going to have this impact. So you know, ethical shifts happen very often for lack of like strong systems and misunderstandings and guidelines. So we try to, we try to prevent that. We of course think about it after as well, but we try to make sure that by default we are concentrating on the things that matter most and not causing harm. Essentially.

Kelly Wortham: This to all three of you. Do any of you, like when you're, when you're either experimenting with AI or experimenting in general because of AI, do you change the way like what Camilla just described to me doesn't sound like something you should only do in a we'll call sensitive industry. It sounds like if we all did that, if we change the way we design our experiments and plan, you know, good hearts law, let's make sure we've got the right metrics here. If we're optimizing to the right metrics, can that be a, a way to invite ethics into the room to, to Sonny's point, and have you changed the way you design experiments today compared to the way you used to say, you know, five years ago.

Iqbal Ali: Sonny, you look, you're making a face like you want to talk. So I'm a big fan of pre mortems. So before running a test, kind of like figure out what are the different outcomes that could happen and when it comes to sort of ethics and ethical impacts, especially when it comes to values like privacy or sustainability or whatever, thinking ahead to what is the worst case scenario that could happen because of what you're going to be going to be deploying and then from that worst case scenario work back to what other metrics that you could collect to ensure that that is or isn't happening. So yeah, just, but just basically, but even before that, just documenting what is your position when it comes to a set number of values? Because a lot of companies say, oh we're sustainable, but what is actually in your company regulation, company rule book when somebody is doing like when product owners are releasing some feature or whatever to kind of meet with sustainability guidelines, what are the practices that you're, you're taking and then just kind of implementing those practices into some of a pre mortem to make sure that Whatever it is that we're releasing meets with these guidelines and how we measuring that we are meeting those guidelines.

Slobodan Manić: It's almost like an ethical operating system you're describing that overrides anything that's related to any individual experiment. And I think that's probably the closest way to doing this. The right way. There's not going to be the right way. But how do we feel about, let's say a company hires an agency. We talked about this this morning, Iqbal. A company hires an agency and the agency uses some dark patterns, a CRO agency use patterns. And the tests win. Let's say they, on a subscribe button they say, no thanks, I don't want to save money or something stupid like that. And that's. That test is a winner. It doesn't harm the user. It's not unethical in that way, but it absolutely kills your brand's reputation in the long term. Is that also important, staying ethical as a brand? Or do you just say, okay, this is a 2% win, we want to say something like that on a subscribe button, even though we think it's horrible. Is there a line for that?

Kelly Wortham: They wouldn't even say that it doesn't.

Camila Dutzig: Necessarily harm the user yet. Perhaps it does, perhaps it does not. But I think that the question is first asking. I think that they have. There's also this interesting set of primary principles for ethical behavioral research experiments. And one of them is just three of them, which makes it actually quite easy and prioritized. And one of them is that people's agency, like, cannot be affected. So it's one thing to like, inform them so that they make a decision is another thing to like, make it hard for them or stir them.

Kelly Wortham: Understand? So you're informing them, but you're informing them in legalese. That does not count.

Camila Dutzig: Yeah. And informing and forcing are different things, right? Or coercing and, or manipulating or making them feel inadequate, making them feel idiotic for doing so. That's different than informing. So one thing is that. Then the second thing is like, how hard for them? How hard is it for them to recover from this action, Right? So if you are informing them on an action that they actually want, you know, to do that, another way to think about it is that would they want to do that anyway, right? You're just facilitating their process. Or they actually wouldn't have taken that action had you not like, like coerce them into it. And then how easy is it for them to actually like, take a step back and does it actually harm them? The action that they did. So long term or emotionally or whatever it is. If you can say no to all these questions, then your past, even if you use some kind of like intense persuasive copy and tactics, might not be that harmful and that unethical. Right. But if you cannot say an active note or this three main questions, then you're in very dangerous territory.

Slobodan Manić: Yeah. And that's from Belmont report that we talked about, right? Yeah, it's a 1978 report that was made by US government committee, whatever after the Tuskegee syphilis study. So this is why we have ethics and experimentation because some people did some bad things. It, it's horrible.

Kelly Wortham: Well, in that case, you know, the very unethical. But we could argue that they were optimizing for learning in a really gross way. But I think what happens with us is we're not even optimizing for learning. We're optimizing for revenue over or growth, you know, profitable outcomes. So how can we like I really liked Camilla's framing earlier of focus first on real user problems and solving those problems in accessibility. And you know, is, is it as simple as changing our metrics, our goals to make sure that we're being ethical, putting people first?

Slobodan Manić: Who's making that decision? That's the question. It's the person that wants more money.

Kelly Wortham: Well, I think as optimizers, as experimenters, you know, whether we're working as a practitioner with a brand or a consultant or vendor supporting them, everybody has that end goal. But it feels like to your all's point earlier, maybe it's our job to educate about how you get there and what the leading indicators can be like. I think it's really interesting that companies that are sit in those really privacy sensitive environments may have figured it out already. I mean, how many talks have we had about accessibility where everybody says like, here's the reasons you should do it and you know, the last reason is because people need it and you should be a good person. But the truth is it also works. It converts better. An accessible site converts better. So like maybe we need to do a better job of educating our brands, our clients about being ethical. Maybe ethics converts.

Camila Dutzig: It can. It definitely can. I think that that's a really good point. One thing does not exclude the other and you can have business metrics and you know, aim for increase of conversion. But I think that the risky thing is not that necessarily. That's great, let's do it. But the, you know, one way to approach it is to maybe have additional metrics that you measured to also have a look on the other impacts that bringing conversion up can have. You know, and, and this can translate in bad things for your business actually that you might not even like be looking at. And also the human body.

Kelly Wortham: Right.

Camila Dutzig: Because Herms conversion in this page might actually destroy brand trust in the long term. Satisfaction returns, nps. They can actually buy a lot more but then return all of it the next day and then your organization has an even bigger issue because logistically is a lot harder and expensive and you can be creating all kinds of issues. So yes, sure, you know, let's aim for conversion, let's aim for more sales, but like have the metrics in place and the guardrails as they say. So that this is like overall, this experience doesn't harm your own business even. And also, most likely when doing so, you're also, if you have, you know, customer feedback, retention, loyalty, nps, if you're looking at these things as metrics as well, you also know that the potential harm that you're doing to users is like you're keeping an eye on it systematically as well. And I think that this is also a good way to approach it.

Iqbal Ali: Well, also think like there needs to be some sort of business motivation to act in ethical ways. And if there's no business motivation, then nobody's going to act in that way. And there's a business motivation to increase conversion. But what motivation is there to, you know, meet with certain sort of values that a, that a business kind of says they aspire to. So I think without, without having something intrinsically part of the company structure that says we are a company that values sustainability and we're not just virtue signaling. These are the, these are the ways that we should be approaching things. And this, then there's like a business motivation almost say, like, okay, and when we meet these, when teams meet these regulations or meet these kind of standards, then we'll, I don't know, have a party or whatever, whatever the motivation is going to be. But with, without that in place, you know, nobody's going to be really going to be gunning for that. Especially when they've got this other motivation about increasing conversion rate and especially when, you know, they're put in charge of, of marking their own homework when it comes to. Were you being ethical? Yeah, absolutely I was being ethical. Yeah, look at my grade card. And by the way, we made you £10 million and it was all ethical. So yeah, so I think there is like, it needs to come into, from the fabric of the company and if it if it's not, then at a team level and individual level and experiment level, you got no chance.

Slobodan Manić: I think there's a positive way to look at AI in all of this. So basically, when a company does something bad, they can do 10 PR things, 10 big PR pushes, and it's going to be buried in Google search results. No one is going to be able to find it. With LLMs treating these entities, whenever you research a company, it will tell you. But they also had this scandal in 1997. If it's small and obscure, it will mention that one bad thing and this will become more and more important. The small shitty things the company does will be more visible no matter how small they are and no matter how many good things they try to cover it with. So I think that's the motivation. That's going to be the motivation moving forward. It will still stink. Yes, that's what I'm saying. With, with Google and with traditional pr, it was very easy. I mean there was a. I'm not going to say names, but there was a marketing person, a really. It didn't work. Lucas says someone got into some legal trouble and then they pushed a video. Let me tell you about that time. I got sued and they did get sued, but it was something like for traffic 15 years ago. But they pushed that to every single channel to bury the real news about them getting in some real, real serious legal trouble. And it kind of worked.

Kelly Wortham: Wow.

Slobodan Manić: I don't think that's going to work when the LLM is giving you the answers.

Camila Dutzig: No.

Kelly Wortham: And the LLM isn't just going to look for your press releases. It's also going to look exactly said on Reddit about your brand and their experience with your brand.

Slobodan Manić: And being loud is not going to be as important.

Camila Dutzig: Yes, that is assuming of course that they don't change the algorithms. Bypass this.

Slobodan Manić: I have, I have another ChatGPT leak story to share. So when what Craig mentioned when they, when they index those chats, I tried searching for that URL, that specific URL pattern and Sam Altman good. So chats that people had with threads they had with ChatGPT had thousands of results and under quotes. So it was like a direct search, exact search. Sam Altman bad. Zero results. There's no fucking way. Nobody ever searched for anything that says Sam Altman bad. There's no way. So they're not.

Kelly Wortham: Did you see Corey's thing? Until they change.

Slobodan Manić: This is. This is influenced already. Don't tell me those fragile ego baby men don't mess with those algorithms. Like, there's. There's no way. They don't. There's zero chance that Elon Musk. I mean, he does. Like, when you ask. I asked Grok once, I asked it something really, really disturbing. I asked it, would Elon Musk be better at organizing what happened in World War II? And he said, elon Musk has great organizational skills. He would probably do a better job. Yeah, that's how cool he is. So these algorithms don't work for us. They work for six men, basically, and we need to treat him that way.

Kelly Wortham: Oh, wow. All right. So we're in this, like, new frontier. I think we can all agree.

Slobodan Manić: Absolutely.

Kelly Wortham: And, like, people are generating experiments using gen AI or using AI in other ways to automate experimentation. So isn't the risk. And. And by the way, we got it on record. Sonny said large language models are good. Like, we all heard it. It's recorded.

Slobodan Manić: It lives on YouTube, if you know what they do.

Kelly Wortham: But my point is, like, if we're creating these experiments at scale, isn't the risk, like, the stakes, aren't they, like, exponentially multiplied? Especially, by the way, Camila, in sensitive industries. And, like, what do we do? Sunny, Actually, I want to start with, like, you're the AI chat prompts exposure. Like, in that story, in that example, is there anything that experimentation teams should learn from what you learned and what you all exposed that they might do differently to protect? When it comes to data and, yeah.

Slobodan Manić: That cloud that you cloud, you don't control, and sensitive data don't mix, period.

Kelly Wortham: Say that again.

Slobodan Manić: Anything. Cloud, cloud, cloud, cloud, anything. No, cloud is, as I said at Conversion Hotel, the middle child. Nobody cares about the middle child cloud. And, sorry, anthropic cloud and sensitive data should never mix. You should never, never share your sensitive data. That's how we got Cambridge Analytica. Do you want your sensitive business data to be a part of the next Cambridge Analytic or whatever it is? If you do, that's fine. You can go ahead and do it.

Kelly Wortham: Wow. What would you say, Camila?

Camila Dutzig: That's one aspect of it, right? Like, the data that you're feeding the system, where you are, like, teaching it. But then the second aspect of it is, like, what do you use this AI and how much agency do you give it? Like, how much can the AI actually make decisions on your behalf? Right. And what kind of decisions can you make? So what part of the experience does it control? If you give a lot of data and a lot of agency to make a lot of changes, you are possibly quite like, sorry, I Actually shouldn't say this and be recorded for quite.

Kelly Wortham: Okay.

Camila Dutzig: And especially, and this is what I think is also really tricky and on top of everything that we already know, quite complicated. The added layer is, can you reasonably explain yourself and to the end user, what is the AI doing, what data is collecting, how is it using, what exactly is going to be the outcome? And that outcome is actually reliable and consistent. If you cannot say yes to this, you should not be giving AI the ability to control an outcome in any way, shape or form.

Slobodan Manić: Succinct.

Kelly Wortham: When you're doing stuff like this at scale, do we have time, like to maintain, like you, you described frameworks that we need to be using, but like, how do we maintain that ethical oversight at that kind of speed and scale where people are like using experiments, using AI to automate experimentation, which really is in and of itself optimization.

Iqbal Ali: Yeah. So I think, you know, when you, when you're going at scale, it's an, it's an operational problem at that, at that point. And, and we're deploying small models, language models to solve those, a lot of operational problems. So when it comes to, when you do have some frameworks and you do have some definitive sort of, these are the values that we aspire to and these are, you know how important they are. These are primary values and these are the, the red lines that we cannot cross. And these are the kind of things that we need to look out for. There's nothing stopping somebody to, to write like a SLM or a workflow to kind of detect these at scale. And as well, you do need to have some people who are responsible AI champions within the company to be constantly checking these because it's either important to a company or it's not. And if it's not, then, you know, just let it fly by and don't pretend that it is. But if it is important, then you would be actually going through and installing these simple sort of frameworks and these simple workflows to catch at least these big problems at scale and avoid all of these ethical debts that you're going to be building up through these small minor infractions. Even.

Kelly Wortham: So, does it all come down to, you know, you, you described those like earlier, you called it a moral compass. But regardless, you have like these list of values that a company has, ethical values that they, that they want to make sure if we're experimenting, if we're building AI, if we're building systems or products or features, we follow these goals. Does it come down to knowing how to measure like we Always ask the question when somebody comes to us, like, were we successful? You have to start by what are you trying to accomplish? How would you measure it? How will you know if you get there? Do we have to ask those same magical questions of the goals that we have associated with ethics, figuring out how to measure?

Iqbal Ali: Yeah, I mean, I think like there's a lot of really good frameworks in Olivier Gamblin's book. So that's, you know, if, if people are interested, really pick up a responsible AI, you're going to put that link.

Kelly Wortham: In the TLC chat, right?

Iqbal Ali: Yeah, yeah, it's really good, really practical. It gives you these frameworks. What she says is basically you have set of values, so for instance, privacy. And then you have a government sort of set of guidelines for privacy like GDPR or whatever. You have an industry wide one, you have a organizational one. If you've got one for your company, if you don't, then you should write it. And then you, then you should know what your team or individual sort of guidelines are. And based around that, you can kind of decide what, what is important, what's not. And then after that you can, you can kind of use a sort of like the pre mortem approach. I was saying where before rolling out certain things, you'll know after you've rolled out that these are the problem areas in terms of, you know, like, hey, if we touch something around the algorithm, there's some warning signals there in terms of this is high risk and if we're doing something around the UI or clarity or whatever, that's low risk. But if we're doing something around algorithmic search, search engine changes or personalization or whatever, that's going to be high risk. So we need to put some extra sort of things in place to make sure that what's the worst thing that could happen? How can we make sure that the worst doesn't happen or if it does happen, we catch it quick enough so that we can, you know, turn it off or fix it, whatever. But yeah, the first thing needs to come with that documentation level and that kind of like going through and you know, really deciding on what those lines are for, for a company and for an individual.

Slobodan Manić: And also if AI is evolved, if algorithms are involved, guardrail metrics that you need to keep an eye on all the damn time. Like I'll just give an example. Does anyone know what I just read about it? NH predict is it's the AI algorithm used by UnitedHealthcare that denied or reversed 90% of claims that were approved by doctors for elderly people with cancer care. United Healthcare is a very well known company with a. Used to have a. CEO with a.

Kelly Wortham: What?

Slobodan Manić: Cannot let that happen. You cannot let stuff like that. I'm sure they have a new one, but they bragged about this algorithm that denied or reversed 90 of the claims that were approved by doctors, by human doctors. And that was a goal for them. I'm sure they trained that to do this. But. But if you're not careful, your company can do things that are equally bad, equally terrible. And if you're just not controlling what AI does and what the algorithm does, you will maybe not even know about it.

Kelly Wortham: One of my favorite stories of a medical company that is actually doing really good job with their experimentation. Do you know what the goal is of every single experiment Mayo Clinic runs? Health outcomes.

Slobodan Manić: Health outcomes, yeah.

Kelly Wortham: Try to measure that. It's fun, it's a great challenge. But improving health outcomes is the goal. And that's because doctors are in charge at Mayo Clinic. I love that. So looking at the clock, we, we have to start wrapping up and I just, I want to start by. And then I want to do like a rapid fire and ask each of you, so now you've got time to think, but ask each of you to share like, what is the one thing that either you've done or that you recommend a company do to be more ethical in their, in their practices related to AI experimentation, user research. But as you're thinking about that, I just want to say that like I said in the beginning, this shouldn't be about having the answers. It should be about asking the right questions. And I think that is our job. And I think all three of you have like, brought up a lot of ideas to inspire people to ask those questions. And those of us who do work with brands to build trust, I think that's, that's the end goal. And I think those are the ones that will succeed long term. So each of you, thank you. And if you want to close us out, each of you, what's the one piece of advice you would give?

Slobodan Manić: You said it. Ask the questions. Be skeptical. Because these are not companies you can trust blindly. When you're working with AI in whatever you're doing, you cannot trust them blindly. They need to earn the trust.

Iqbal Ali: Camila, for me, it's think, yeah, piggybacking off of that. Think critical think what is a worst case scenario of what it is that you're doing and then try and determine how do I know that that worst case scenario is happening as quickly as possible.

Camila Dutzig: I am going to be sneaky and give two answers in one. Because first I wanted to say, yeah, try to create an experimentation program that is as ethical as possible by design, so that it. It makes sure that you don't have to think so hard, that it's not so operational, that it doesn't like, require so much constant thinking and going back. But I think I already said that because before. So I'm actually going to be sneaky and give another one, which is at the end of the day, in the middle of all this chaos, we are all going to be using AI and that's. And, you know, do it responsibly, so on and so forth. But what you can do and what you should do is do not allow the AI to actually dictate the outcome of it. Right. Do not without human interference and careful one. So I think that this is the first step. We're just not there yet to be able to it do, do this ethically without harming people. And it's not about not using it or being against it, but do not let it dictate what our future will be like. Right.

Kelly Wortham: I love that. I love that. Thank you all for a wonderful conversation and I hope. Sonny. Cheers. I wish I had a something to do here. I'll toast you with my water. Enjoy your holiday break. And for the rest of you, the next time we chat will be a new year, hopefully a better one for everyone. So happy end of year, whatever holidays you're celebrating and Happy New Year. We'll see you in January. Thanks, everybody.

Slobodan Manić: Thank you.

Iqbal Ali: Thank you.

Kelly Wortham: Bye bye.