The speaker outlines a practical checklist used by their organization to ensure ethical and responsible development of AI products in drug discovery and healthcare. The process begins with evaluating whether the training data are sufficiently representative and free of bias, including proactive steps to acquire additional data when gaps are identified. Next, the team defines the model’s intended use and systematically explores possible unintended applications, considering regulatory and societal impacts. Finally, they emphasize the necessity of continuous monitoring after deployment, requiring dedicated oversight to log performance, detect drift, and intervene as needed. The overall framework is presented as a set of anchoring principles—data integrity, use‑case clarity, and ongoing governance—to guide large‑scale, high‑impact AI initiatives.
Chapters
One of the things we have is sort of a checklist when we're starting to consider to build a product. So us as an organization, we only try and do large impactful things, right? We're not solving someone. I want to transfer this Word document to an Excel spreadsheet and building a thing for that. We're building big things, impact patients or help discover targets and things like that. And they have large term, long term commitments in the company. We discover a target, we do something, we're committing money to it, there's an opportunity cost, there's all these things that happen to that. So these are big problems we're working on. So the first thing we look at is, well, first of all, let's look at the data you're using, right? And you can't sit there and say, well this is the data I've got, this is that I can find. The first question is, is your data adequately representative? Right? It's got biases and if not, and you know, have other sources where you could find more data, right? And that's a very interesting question. And then you ask like, well, it's sometimes it's not okay to say I can't, I just couldn't find it. It's not out there. You say, well, should we gather it? Should we just go and actively gather the data? And that's actually pretty new to most people that come into machine learning, particularly at school. They're like, what do you mean? Like, well, we'll just buy more, we'll generate more, we can maybe run a study, how do we do this? And we can look at the cost of doing that right now if it's. Obviously there are a lot, there are time bounds and large amounts of money to do this sort of stuff. And it may be invisible to things depending on what you want to do. But you have to make an attempt and we have to have understood that and we pre specify what biases we know about that and we made reasonable attempts to address all of them. I think that's before you even write a single line of code, just to think through what that is. Then you think the other thing you think about is the what if scenario. I have the model, it's in production. So what, what do people do with it? What's my intended use? Right? And you can write that, well, that's easy. That's what I want to build a model for. And then what are my unintended uses? If someone else had access to this or what would use this, how else could they use this? In different settings. And that's where I worked in systems in the past that could predict people's probability of getting disease at time t given their past medical history. Really good for planning care, really good for doing pre implementations, things like that. All great stuff. Also really good factorial methods for insurance companies and pricing healthcare and things like that. Possibly good stuff, maybe less good stuff, depending on the regulatory environment. You have to also think of the intended unintended uses that you can when you put something in production and how it changed the world around them. So those are two anchoring principles. What goes into the machine milling thing and what comes out of it from the other side. And we have checklists and things that help people think through that. Right. And it's about thinking through that and to make sure, okay, this is, tell us what your original intent with your model is, what you. How you want to make the world a better place. And then let's work out the best way that we can practically do to that aim. And also how do we monitor. That's the third thing. How do we monitor this? Doing that in practice. Right. So you can't just build these things and like off it goes and I'll go and work another problem. Someone has to watch it and log its results and review it.
