Securing the Human Data Layer of AI with Siobhan Hanna

Russ Fordyce 🏆 (00:01.506) Welcome back to the Winner's Circle, everybody. I'm here with Siobhan from WeLo Data, and they are actually helping us clean up the data in AI, which is going to be miraculous. Siobhan, welcome to the Winner's Circle, and congratulations on the AI Excellence Award. Siobhan Hanna (00:15.485) Thank you so much, glad to be here. Russ Fordyce 🏆 (00:18.478) Before we get started in getting into the nitty gritty of kind of the challenges with AI data, give us a little bit of background on WeLO and kind of what they're all about. Siobhan Hanna (00:28.625) Absolutely. So, WheeloData is the world's leading multilingual AI human data provider. As you know, high quality data specifically is essential to train the world's leading AI models. We work with foundational LLM model builders and large and enterprise tech to provide those high quality human data sets. And really our niche is that we can provide those data sets across any language, any kind of cultural requirements across the world. Russ Fordyce 🏆 (01:05.486) You know, I've been thinking a lot about this because you read all the stories about where the LLMs are getting data and a lot of it is from social media sites and websites that Reddit that maybe aren't the most accurate in the world. What drew you into this problem of kind of finding fraud and AI data? Siobhan Hanna (01:24.681) Absolutely. we do work with our customers to solve complex problems and this is a very large scale problem solving for high quality data, whether that be human or synthetic to train these models. And because this work is frequently executed in a contractor crowd model, it is vulnerable to fraud, right? It is a big business. It is complex. There are many systems that we and our customers are working across. These are global remote dispersed teams and you know many many touch points and just many areas where there is an opportunity for model or data quality degradation to occur. So we observed that of course a number of years ago and we identified the opportunity to just pull together our solutions in into a more robust scalable framework, which is what Nemo is. Russ Fordyce 🏆 (02:29.067) So the question I guess kind of first in my mind is, you know, why has this been so underserved? Siobhan Hanna (02:36.807) Well, I think the AI industry obviously clearly is moving very, very fast. I think that the speed of change and evolution and scaling in the AI industry sort of created a bit of a blind spot. Of course, not to say that fraud is not something that has received a lot of attention, but there are many other elements that our industry is focused on. Obviously, model performance, superior model performance and continuing to improve on model performance. It's a very competitive space. Infrastructure, safety, scaling into multilingual, which is an area that we're very interested in. We were focused on all of those elements as well, of course, but we did feel that there was opportunity, certainly not speaking for others, but for our team to really ask ourselves some tougher questions around the integrity of the human layer that's producing the pre and the post training data. And because clearly that quality, we see fraud mitigation as part of the data quality assurance layer. That if we don't get that right, that can of course result in degraded model performance just as quickly as ineffective data quality workflows and under qualified contributor networks and so on. So we started to build NEMO because we saw that gap and we saw particularly the opportunity with how globally distributed our work workforces that we load data. And we saw kind of the lack of an off the shelf enterprise grade security framework that was available. So we felt we needed to build that ourselves. Russ Fordyce 🏆 (04:28.279) So we're talking about this kind of human generated data that's either fraudulent or misrepresented. And it's almost like a prompt injection, right? It's these LLM injections, it seems like. Can you walk us through the downstream effect on a model when that happens? Siobhan Hanna (04:43.017) when there's ineffective fraud mitigation or... Russ Fordyce 🏆 (04:47.445) No, when somebody's trying to inject LLM, when you see that fraud going into the LLMs, how does that affect everything downstream with that model? Siobhan Hanna (04:57.201) yes. Absolutely. Well, mean, ultimately it means that the quality of the human judgments, whether that's model eval or the base data sets, the integrity just may not be there as designed. And that may be difficult to detect as our customers, as model builders and developers are evolving their model until it's too late. It can be difficult to pinpoint what the source of degraded model performance is. And that is one of the challenges, right? It's often difficult to figure out exactly where in the data set did that degradation, did that issue, that gap crop up? And then it will, of course, proliferate across the model. We would, of course, rather do what we can to ensure that we get the integrity right from the start. And a big part of that is controlling our contributor process. not only pre-production while we're qualifying and selecting and onboarding our experts, but also in production itself. Russ Fordyce 🏆 (06:08.525) So you're out there trying to detect this stuff. It sounds like it's pretty hard. What makes it so hard to detect? Siobhan Hanna (06:15.685) There's just, again, a very high pace of change, very significant differences across infrastructures for different types of projects, and then lots of different ways that risk can creep into the process. That may be in the form of identity fraud. There's lots of ways that that can permutate. Of course, it could be coordinated manipulation, as I mentioned. the model itself is kind of vulnerable to that type of initiative. And then just general other data integrity issues that can creep in across these complex multi-step human data pipeline processes. Russ Fordyce 🏆 (07:02.647) Now we have a lot of AI kind of developers and companies in our audience. I imagine some of them are more risk than others depending on your applications. What types of applications are you seeing that are the most exposed to this kind of trading data integrity? Siobhan Hanna (07:19.305) I think really any AI system that is built on human data, which is most of them, that human judgment that is essential needs to be protected with the same rigor that we would apply to any other critical infrastructure would really be our take. So there's not necessarily one that is sort of more vulnerable to the other, but of course some model performance is more dependent on human judgment, know, more sensitive to cultural nuance and so on. And in those instances, you know, we think that understanding who these contributors are and getting the kind of base data quality right is really, really critical. Russ Fordyce 🏆 (08:10.473) It seems like it's always been a problem on the internet, kind of validating identity. All the big social platforms have tried to really crack down on this. If you've tried to make a fake Facebook profile at Daria, it's really, really difficult now. What's the difference between kind of this identity validation that you guys seem to be doing? And why is it so tricky compared to financial fraud or some other type of fraud out there? Siobhan Hanna (08:15.741) Yeah. Siobhan Hanna (08:35.689) Absolutely. So a good mention actually on financial fraud. We did try to borrow somewhat from financial services when we were envisioning NEMO on our broader fraud mitigation and misrepresentation mitigation framework. So KYC, know your customer principles, real time transaction monitoring throughout the life cycle of a project and engaging with our experts, behavioral analytics, all of those industries, obviously in financial services, learned decades ago that kind of point in time checks aren't enough and that continuous monitoring is the standard. So we did seek to apply that logic to our AI data operations. I think why this is a bit different from kind of traditional is, you know, traditional QA, for example, or validation is designed to catch errors and output. And NEMO is really intended to secure the environment. environment in which the output is produced, which is kind of a fundamentally, we think, different approach. And, you know, again, as I mentioned, the human layer is not necessarily a point of failure. It's an extremely important part in AI model development. We think, of course, we believe that that is where the most value is created. And we think then that that means it's worth protecting that value. The same qualities that make human experts, our contributors irreplaceable in model development and that's judgment, nuance, domain expertise, cultural insight. That's also something and contributions that are kind of vulnerable to bad actors impersonating or exploiting. Russ Fordyce 🏆 (10:27.543) Yeah, and I know a lot of the LLM builders are actually pulling in expertise and really human experts. hear about, I heard about OpenAI trying to take down Wall Street with 400, know, kind of Wall Street types coming in and sharing all their knowledge. How do you actually validate that Nemo's working and it's not just catching more activity, but catching the right activity? Siobhan Hanna (10:33.128) Yes. Siobhan Hanna (10:48.955) No, absolutely. plus one to what you just said. mean, before there can be artificial intelligence, there must be human intelligence. And many of our customers, like ourselves, have recognized that human expertise is essential in creating differentiated sort of models that perform in an extraordinary way. So sorry, do you mind repeating the question? Russ Fordyce 🏆 (11:17.537) Yeah, how do you validate that Nemo's working and not just catching kind of more activity, but catching the right activity? Siobhan Hanna (11:24.093) Yeah, absolutely. So it's a work in progress. We have data that indicates that it is working and we benchmark ourselves pretty rigorously. In terms of the signals or the data that we do use, we have at current current, we have a little over 130 unique behavioral variables that we monitor per session. So they're based on things like timing patterns, interaction signatures, cross-session and consistency, other anomalies that might suggest behaviors like a current sharing or coordinated activities are occurring. We tried to, and we're pretty rigorous about looking at the full picture of contributor behavior, not just the final output. And then we deploy that framework across the entire life cycle. So talent sourcing, candidate selection, live production, and That is kind of a combination of rules based and AI driven proprietary of course framework. That's very intentional as well. We know that rules based logic handles known threat patterns really efficiently. The AI driven layer is what we have found allows us to detect novel and evolving patterns. And that is important because bad actors, threat actors, do adapt, they evolve of course quite quickly and we know that a static detection system will eventually be outpaced. And then I think the last piece which allows us to keep pace with and hopefully stay ahead of and mitigate risk in active pipelines is behavioral psychology. So we do have organizational psychology team within our fraud mitigation organization. I think that this is a somewhat underappreciated dimension of the problem. Understanding human motivations and behavioral markers behind fraud makes our ability to detect it significantly more accurate than just purely technical signal analysis. And then, you know, obviously the overall intent to answer your question is to kind of stop threats before they get to the model. That's the design intent that we have because Siobhan Hanna (13:53.595) Because we really focus consistently on the security perimeter upstream before the data is produced, not after, but then we continue that rigor through the end of the data set development process. And because we do self-audit and we provide that audit data to our clients, if that is something that they're interested in, we're able to benchmark our activities. our operational framework and identify opportunities to improve. Russ Fordyce 🏆 (14:28.311) So you've looked at all this stuff, you've caught a bazillion fraudsters and fraud content in there in all these LLM. What surprised you the most that you started seeing in that kind of real fraud landscape? Siobhan Hanna (14:41.227) well... I probably don't want to talk too much about individual situations and behaviors and patterns that we've seen. I I think the pace of evolution and kind of the pace of sophistication growing in terms of rings is very, very interesting. I think I'm a little surprised that this is not a larger conversation in our industry. to get folded into sort of data quality and it is a data quality problem because as we discussed a moment ago, sort of failing to proactively address this will reduce and result in degraded data quality which will impact model output performance. But I think what surprised us most is that it isn't a larger conversation and then because we We are very focused on and a leading provider in the multilingual human data field. This is a larger challenge as we expand to global markets. We find that that is consistently underestimated. When I talked about our framework and all of the different checks that we have, just because it works for one line, does not mean that it will work for a hundred plus languages or markets and you know we operate in well over a hundred markets. The cultural context really really matters there and the security model itself needs to be adapted for the needs of different markets and that is just really not trivial and something that we're very interested in. Russ Fordyce 🏆 (16:33.449) No, the context matters so much, especially with different languages. I mean, you can walk down the street and get three different definitions for a common word, right? And you multiply that by 103 or however many languages you support. It's really a tough challenge. Now, you guys have built this system, you mentioned earlier, continuous monitoring system. I would imagine you kind of had the design principles when you started designing. Siobhan Hanna (16:45.001) Absolutely. Russ Fordyce 🏆 (17:02.859) I imagine you looked at kind of doing that one time screening check, that kind of onboarding of your experts versus this continuous system. What led you to kind of move towards that continuous monitoring system? Siobhan Hanna (17:13.833) Really just starting to hold ourselves accountable and observe evolving behaviors that were, despite best intentions upstream prior to the start of production, that were undermining those efforts in production is when we really understood, we need to apply the same rigor throughout the life cycle. The stakes are just so high, right? They're very high for our customers, so last enterprises, compromised training data is huge risk, but it's very very significant for us and for the industry overall and there is so much focus on trust in AI that this is just not a challenge that we could ignore. Russ Fordyce 🏆 (18:04.587) And we talked about the financial services and kind of that know your customer framework. It seems like you've adopted that kind of from that financial services industry and kind of are applying it to this AI training context. What translated well and what had to be kind of completely reinvented from that financial services model. Siobhan Hanna (18:22.601) Yes, I think that as I mentioned, the global scale and the understanding of culture, I think was really essential. And I do think that there's less that we can control in our work than financial services can control in terms of complexity, different models, different systems, different platforms. So the pace of change, I would imagine, and just the openness to evolving our strategy was probably just a key difference. we were originally, we did spend quite a bit of time studying how those financial services models had evolved and they did form the foundation. And then as I say, we kind of went into that kind three pronged approach, rules based, AI driven and behavioral psychology led. Russ Fordyce 🏆 (19:20.589) Now, the thing that I'm going to struggle with a little bit is you've got 103 languages, and now you apply kind of all the evolving regulation around this too that's all over the world. have the EU's done an AI act, you have seemingly millions of US executive orders coming out every day. All these standards seem to be emerging. Where does this training data integrity fit into the governance conversation? And is anyone mandating it yet? Siobhan Hanna (19:48.873) It is not mandated yet. That's a great question. The threat landscape is sort of really what we are focusing on as we try to think about where is AI risk headed. We know that the threat landscape obviously will continue to evolve in direct proportion, I would say, to the value of what's being protected. So as foundational model development accelerates and training data becomes more and more strategically valuable. The incentives for sophisticated actors to infiltrate these pipelines really increases and that's just the reality. you mentioned the regulatory side of things. We expect that data provenance and contributor verification will become formal requirements eventually rather than best practices, which is approximately where we are today. Yes, the conversations happening in Brussels and Washington are seemingly moving in that direction. Russ Fordyce 🏆 (20:50.379) Yeah, you really think about it. It's a huge, huge problem. It's not unlike email where you get Nigerian scammers asking for money. And I don't mean to pick on Nigerian scammers, but this kind of prompt injection thing is really, really tough. The responsible AI infrastructure you guys are building, when you go into a customer, I imagine that some of the organization, the data team is probably thinking about this. I imagine almost nobody else is. Siobhan Hanna (21:23.143) Yeah, I I think it actually is an important question that many model builders and tech companies are wrestling with at the moment. Really, if we have the opportunity to influence those conversations, we really do encourage our partners to start taking the provenance, the data provenance question seriously. Who produced this data? and those experts verified, and most of our customers are very responsible and proactive in this regard. What does the audit trail look like? Those questions really should be a key part of every vendor evaluation. just really starting to move, and again, most of our partners are thinking about this really strategically, but thinking about the AI supply chain the way more mature industries think about their supply change in terms of compliance and risk implications. The human layer is again very essential. We are very passionate about the human layer, including the experience of our experts, but it's a critical part of the chain and it deserves the same scrutiny as any other critical input. Russ Fordyce 🏆 (22:43.391) And if you look at all the work that you guys have already done, is there kind of one piece of AI infrastructure that seems to be the most vulnerable? Siobhan Hanna (22:53.161) Really, not necessarily, really any model that relies on high quality human data has the same vulnerability. Russ Fordyce 🏆 (23:04.471) So if I'm a founder, operator, or AI leader, what experience should I take away from this conversation about Nemo and how should I apply it, even if I don't choose Nemo as a provider? How should I kind of start looking at my data and seeing if I have a problem? Siobhan Hanna (23:19.306) AI integrity does not manage itself. It requires sustained investment and strong governance and a genuine belief, I think, that the human layer of AI development is worth protecting. do focus on that data compliance framework, who produced the data, making sure that your partners have a robust strategy in this regard, and that they are safe. safeguarding the provenance of the data that they're providing. Russ Fordyce 🏆 (23:53.997) Well, such an interesting space, such an interesting problem, and one that I think is going to get bigger. It's not hallucinations, folks. This is fraud. And they're going to be injecting into all these LLMs and telling you things that you think are true that just simply aren't. So it's going to be a real problem. And Siobhan, I really thank you for coming on and explaining all this to us. And congratulations again on the award. Siobhan Hanna (24:16.029) Thank you, thank you so much. We're thrilled.