Ethics Practices in AI Development: An Empirical Study Across Roles and Regions

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Ethics Practices in AI Development: An Empirical Study Across Roles and Regions".

Jane: The paper was written by Wilder Baldwin, Sepideh Ghanavati and Manuel Woersdoerfer from University of Maine.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Welcome back to the show, everybody. Today we're digging into a paper that's been making the rounds, and it's called "Ethics Practices in AI Development: An Empirical Study Across Roles and Regions." Jane, I have to say, just the title alone tells you this isn't another opinion piece about AI ethics — this is actual data.

Jane: Exactly, Tom. And that's what got me excited when I first skimmed this. The authors, Baldwin, Ghanavati, and Woersdoerfer from the University of Maine, they didn't just ask a handful of developers what they think. They went out and surveyed four hundred fourteen people across forty-three countries. That's a big deal because most of the conversation about AI ethics happens in boardrooms or in philosophy papers, not on the ground with the people actually building these systems.

Tom: And that's the key thing, right? They talked to people in very different roles. You've got your AI managers, your developers, your quality assurance folks, even information security and privacy experts. I mean, that's the whole lifecycle of building an AI product, from the person writing the requirements to the person testing it for vulnerabilities.

Jane: Right, and that's what makes this study so valuable. Because when you only ask developers, you get one slice of the story. But when you ask the manager who sets the guidelines, the analyst who writes the requirements, the QA person who tests for bias, and the security expert who worries about data leaks, you start to see where the gaps are. And let me tell you, there are gaps.

Tom: Oh, big ones. I mean, one of the first things that jumped out at me was that people in quality assurance and information security roles were the least familiar with most AI ethics principles. These are the people whose whole job is to catch problems before they reach the user, and they're the ones who know the least about things like fairness or transparency.

Jane: It's almost backwards, isn't it? You'd think the gatekeepers would be the most informed. But the data shows that managers and researchers are the ones who are most familiar with these principles, while the people who actually need to apply them in testing and security are left in the dark.

Tom: So the paper isn't just saying "everyone should know more about ethics." It's saying different roles need different kinds of training, and right now, the training isn't reaching the people who need it most.

Jane: And that's the kind of insight you can only get from a study like this, with this many people, across this many regions. We're going to dig into the actual numbers and what they mean for how teams should be structured. Stick around.

Summary: Tom: So we're back with "Ethics Practices in AI Development: An Empirical Study Across Roles and Regions," and Jane, I want to get into the actual findings, because there's a lot here that surprised me.

Jane: Same here. Let's start with the big one. They asked people how familiar they were with nine different AI ethics principles, things like data protection, transparency, fairness, accountability. And across the board, the most familiar principle was data protection and privacy. Over ninety percent of people with AI development experience said they were at least somewhat familiar with it.

Tom: That makes sense though, right? Privacy has been a thing since before AI was cool. We've had GDPR, we've had all these data breach scandals. So people know that one. But here's what got me — the principle they were least familiar with was "democracy and rule of law." Only about seventy-six percent of experienced developers were familiar with that one.

Jane: And that's a problem, because that principle is about making sure AI doesn't undermine democratic institutions. It's about who gets to make decisions about how AI is used in society. If the people building these systems don't even know that's a consideration, then we're building systems that could have really serious societal impacts without anyone thinking about them.

Tom: Right. And the study also found that this familiarity isn't just about the individual. It's about the role. Managers were the most familiar with all nine principles. Information security people were the least familiar, except when it came to privacy, which makes sense given their job.

Jane: And here's another layer — the researchers found that gender mattered too. Female participants reported higher familiarity with all the principles and said they considered ethics in their work more often than their male counterparts. That's a finding that has implications for how we think about team diversity.

Tom: Yeah, and that's not just a nice-to-have. The paper suggests that diverse teams bring broader ethical perspectives. And they also found that people in government and academic institutions were more familiar with these principles than people in multinational corporations.

Jane: Which is interesting because you'd think big tech companies would have all the resources to train their people. But the data says otherwise. The paper also did some fancy statistical modeling — LASSO regression and mixed-effects models, for those of you who like that kind of thing — and they found that the single biggest predictor of how familiar someone is with AI ethics principles is how familiar they are with AI governance initiatives.

Tom: So knowing about the EU AI Act or the NIST framework actually makes you more likely to know about the underlying principles. It's like the regulations are teaching people what the ethics are.

Jane: Exactly. And that's a really actionable finding. If you want people to understand ethics, teach them about the regulations. We'll get into what this means for how teams should actually work in a minute.

Improvements: Tom: So we've talked about what the paper found, but now I want to get into what they're suggesting we actually do about it. And Jane, this is where I think the paper really shines, because they're not just pointing out problems — they're proposing solutions.

Jane: Absolutely. And one of the most interesting suggestions is around threat modeling. You know how in cybersecurity, you have frameworks like LINDDUN that help you identify privacy risks early in development? The authors are suggesting we need something similar for AI ethics. A structured way to identify ethical risks before you even start building.

Tom: That's a great idea. Instead of waiting until the model is trained and deployed to discover it's biased, you'd have a framework that helps you think through potential ethical issues from the very beginning. Like a checklist, but smarter.

Jane: And they're also pushing for tools that can automatically translate AI ethics regulations into software requirements. So instead of a developer having to read through the entire EU AI Act and figure out what it means for their specific product, you'd have a tool that does that translation for them.

Tom: That would be huge. Because one of the things the study found was that people struggle with vague principles. They know they should be "fair" or "transparent," but what does that actually mean in code? What does it mean in a requirements document?

Jane: Right. And the paper also suggests building nudging tools into development environments. So imagine you're writing code and the tool detects that you're about to use a dataset that might have bias issues, and it prompts you to think about that. Kind of like how your word processor flags spelling errors, but for ethics.

Tom: And they're not just talking about the technical side. They're also emphasizing education. The paper found that a lot of practitioners, especially in QA and security roles, just don't have the training they need. So they're suggesting role-specific training programs, not just a one-size-fits-all ethics course.

Jane: And that's the key insight for me. A manager needs to know different things than a developer. A QA person needs different tools than a security expert. The paper is saying we need to tailor our approach to each role, and we need to make sure the people who are least familiar with ethics principles — the QA and security folks — get the most attention.

Tom: There's also a really interesting finding about regulation. Most practitioners actually see regulations as a positive thing. About sixty-six percent of them said they think AI regulation will have a positive impact. But there's a worry, especially in smaller companies, that compliance could slow down innovation.

Jane: And the paper's data suggests that the people who consider ethics most often in their work are also the ones who view regulation most positively. So the more you practice ethics, the less scary the regulations seem. That's a powerful argument for getting ethics into the workflow early.

Tom: So the message is clear: build the tools, train the people, and make ethics a part of the process from day one, not an afterthought. We'll wrap this up in a second.

Conclusion: Tom: Alright, let's bring it home. We've been talking about "Ethics Practices in AI Development: An Empirical Study Across Roles and Regions," and honestly, this is one of those papers that should be required reading for anyone in the AI field.

Jane: Completely agree. And the core message is simple but powerful. Ethics in AI isn't just about having principles written on a website. It's about making sure the people who build these systems actually know those principles, understand what they mean in practice, and have the tools and training to apply them.

Tom: And the study showed that right now, that's not happening evenly. Managers know their stuff, but the people testing and securing AI systems are being left behind. Women are more engaged with ethics than men, which tells us something about the value of diverse teams. And people who know about regulations are more likely to know about ethics, which tells us that teaching regulations is a way to teach ethics.

Jane: The paper also gives us a roadmap. Build threat modeling frameworks for ethics, build tools that translate regulations into requirements, build nudges into development environments, and build role-specific training programs. It's ambitious, but it's concrete.

Tom: And it's not just about avoiding bad outcomes. It's about building trust. The more people understand how AI works and how it's being kept in check, the more they'll be willing to use it. That's good for everyone.

Jane: Well said. So that's our take on "Ethics Practices in AI Development: An Empirical Study Across Roles and Regions." Big thanks to the authors for doing this work, and to you for listening. We'll be back soon with another paper, so until then, keep asking the hard questions.

Tom: See you next time, folks.