AI in HR: Diversity + Ethics (EN)

00:00:00: AI is increasingly influencing HR decisions.

00:00:03: But what does that mean in terms of diversity, fairness and ethics?

00:00:08: Let's take a closer look!

00:00:17: Welcome to HR.

00:00:18: What's Next?

00:00:19: Quick insights new perspectives and practical ideas on the future of HR.

00:00:25: I'm Hinata your host for this episode.

00:00:27: let's get started.

00:00:29: Artificial intelligence is rapidly transforming HR But beyond efficiency, it raises critical questions around fairness diversity and ethics.

00:00:40: In this episode we want to explore how AI systems influence decision-making what risks they carry And How HR can ensure responsible and inclusive use.

00:00:52: Joining me today is Dr Oksan Bok Associate Research Director at the Technical University of Munich's Institute for Ethics in AI.

00:01:02: Oksan will help us unpack how organisations can navigate AI in HR responsibly, especially when it comes to diversity and ethical considerations.

00:01:13: Great to have you with us!

00:01:15: Thank you very much for having me.

00:01:18: Oksana.

00:01:18: – To start with diversity in general what exactly does that mean?

00:01:22: And why is this so important in AI-driven HR?

00:01:28: So diversity is the concept of something is different from the rest of the group.

00:01:35: It can be everything and nothing, when you have a group that's mostly male having one female it's diverse but it could gender or age professional background ethnicity religion anything.

00:01:51: now usually we talk about diversity in context of AI marginalized populations especially historically marginalized population.

00:02:00: So we talk about people that were not recognized at their value of human beings, similarly to other human beings.

00:02:08: And I mean... That's a good conversation to have in HR where that can translate as the bias and outcome for an AI system because it is trained a certain way therefore might be biased.

00:02:22: It also depends on teams you have either developing or using an AI Because diversity means diversity of perspectives So it means having less blind spots.

00:02:32: It means more different opinions that might clash, but when we talk about ethics... ...we also like the idea of having diverse opinions which then leads us to the best answer possible.

00:02:45: We often hear that AI can be biased From a diversity equity and inclusion perspective.

00:02:52: how can organizations actually mitigate these risks?

00:02:57: So there are many ways that you can try to mitigate risks.

00:03:02: when it comes to bias.

00:03:03: First, I think the important thing is to understand what an AI is.

00:03:07: In this context i'm assuming we're talking mostly about LLMs because those are types of AI You would see them most in whether its recruitment hiring onboarding conversation whatever.

00:03:19: and so An LLM Is a system that is trained on a data set That might or may not be representative And therefore, its answer being probabilistic based on what it has been trained on will be biased towards what is originally in the dataset.

00:03:36: Now how do we fix that?

00:03:37: Because we know that as a problem.

00:03:40: Auditing... It's very good way to fix this.

00:03:43: Just audit try test run the AI Run it Try and see outcomes again.

00:03:50: have diverse team testing so they can see maybe the different biases in the outcome, the different blind spots and outcomes.

00:03:59: You can also look at where data comes from.

00:04:01: what really is that trained LLM you're using?

00:04:05: make a conscious decision about which AI are going to pick for usage in your practice?

00:04:13: And if you develop design an AI yourself then have very diverse team on set.

00:04:20: when it comes to conceptualization, thinking what problem are we fixing?

00:04:24: How are we fixiing then?

00:04:26: which model am I picking.

00:04:27: Which weight am i giving ?

00:04:29: To which value in the decision making of DAI?

00:04:33: An example that can be good for this is usually the Schufa.

00:04:37: You know the schufa?

00:04:38: yeah It's a Germany and its credit scoring system.

00:04:43: If imagine an LLM was used to evaluate someone's credit score But the parameters that make that decision don't take into account parental leave.

00:04:56: And LLM might assume someone just stopped working for two years without any good reason, and so they may not be as reliable.

00:05:06: but really if you have in a team developed by parents probably flag it early on and say okay we need to tell the LLM that here, We will have some kind of flag as parent-to-leave.

00:05:22: So you see?

00:05:22: That's the importance Of having diverse teams, diverse data sets, diverse everything And then testing testing testing.

00:05:30: This raises a really important question in terms of You know should AI ever make decisions?

00:05:37: I mean especially In an area like HR because It is about people's lives or The shoe for example is also something to do with very sensitive data and life circumstances.

00:05:50: Where should we actually draw the line?

00:05:53: I think, very early on!

00:05:56: So...I would give a low amount of responsibility for an LLM.

00:06:02: The first answer to this was no, an AI shouldn't have decision-making power.

00:06:08: A human has a decision-making power for many reasons, including the right to redress.

00:06:15: So if a decision is made that as you were saying impacts people's lives then this person needs have the ability go against whoever make their decisions and question it.

00:06:27: The AI lack of transparency would be first issue.

00:06:29: second issues that an AI doesn't think It does not reason or common sense its probabilistic machine.

00:06:37: it's a stochastic parrot.

00:06:39: It doesn't have common sense really, that is why this So having an AI in decision place for me would be a no-no.

00:06:47: Having human being supported by an AI as yes.

00:06:50: Now talking about this I see two big things we should do if we had AI implemented As partnering or teaming with the humans.

00:07:00: First literacy What I just told you about the stochastic parrot, what i told earlier about how an LLM works.

00:07:07: This should become a knowledge for anyone using AI so that they can see the limitations of LLMs.

00:07:13: and The second thing is always human oversight.

00:07:17: You should Always have A Human That Makes Decisions Never An AI Even Though An AI Can Support Your Decision

00:07:25: Yeah?

00:07:25: I mean also in terms Of support yes because You know, you need to move into that idea of also trust.

00:07:35: And it's very natural for us have trust in humans and HR is a human business right?

00:07:43: In every organization.

00:07:45: how can organizations then ensure the transparency and explainability when they're using AI systems especially in the area where we said about people lives?

00:08:00: So let's talk maybe a bit about different use cases because I think that is the best way to look into potential ethical risks, especially linked with lack of transparency.

00:08:13: If you are looking at recruitment using LMS to support recruitment You can be transparent about how you use it or which LLM you picked, arguing here is the data set that this LM was built on and therefore this why we picked.

00:08:32: So this intentionality I think needs to be transparent if your using an EAI.

00:08:41: That's what I was saying at the beginning.

00:08:43: Do some tests, see is that working?

00:08:46: And if you do some bias audits in your own company for the application of an AI and your processes... ...and then report this as a trust-building transparency.

00:08:57: So we're not just using AI mindlessly.

00:09:00: We are actually showing how to use it.

00:09:04: You can also have human oversight.

00:09:11: take the place of someone.

00:09:13: It's about enhancing a human's ability at this stage, or at least the productivity and so it is also important to say that AI will not make decisions which just be followed.

00:09:25: A Human Will Use an AI That might support its decision making And then from that The Human Is Still Responsible.

00:09:32: So trust has built on the fact you can go with somebody.

00:09:35: Tell them hey I disagree With You.

00:09:37: How Do We Fix This?

00:09:39: What About When It Comes To AI actually interacting directly with people.

00:09:44: This happens a lot at the moment in recruiting, for example or in feedback grounds sometimes and internal processes.

00:09:51: what ethical risks do you see here?

00:09:55: So there are many ethical risks.

00:09:59: first it's the way we perceive AI.

00:10:03: I think that's an important point.

00:10:06: We have a cognitive bias, even if we know that it's not true.

00:10:11: We still have this tendency to believe that AI understands us when really doesn't understand anything and the only answer through This is literacy in education Really?

00:10:22: It's about knowing the limitations of an AI knowing how it works And even though you had this feeling You know The truth right now.

00:10:30: after That you have many other issues I can see.

00:10:32: so by yes For me, at least the scariest thing.

00:10:39: We all know this story about Amazon that years ago decided to have this AI in hiring and the AI ended up having a very homogenous type of people also because it was trained on Amazon's data who they hired.

00:10:53: so its not out-of-nowhere but identified big bias already then increased by the AI bias very similar white men from Ivy League.

00:11:08: That is what was mostly hired by this AI algorithm, they of course took it down and then tried to fix that.

00:11:15: Another issue I would see privacy Very big concern.

00:11:18: when we talk about LLMs if you use the LLM's that are on the market They're the big ones.

00:11:23: If We Talk About Cloud, ChatGPT, Gemini...they are linked to a cloud And they belong another company.

00:11:30: So if you put data on there, this data is going to be used most probably to retrain the AI and under GDPR You need to always have a right to remove that data But at this stage not doable.

00:11:43: This is not enforceable.

00:11:45: so We have a big privacy issue here And also more simple privacy issues in terms of employee onboarding .This is also data encryption issue.

00:11:55: Maybe if we can have those LLMs on internal servers and not on a cloud, that would fix it.

00:12:01: And the final problem will be transparency.

00:12:04: You were talking about AI in decision-making.

00:12:07: An AI is intransparent especially an LLM there neural networks behind.

00:12:11: It's very hard to know exactly how decisions was made by an AI.

00:12:16: This lack of transparency Is big problems for people who leave through their decision that was made by this AI, because there is no redress possible.

00:12:26: And that's why we always need human supervision—human in the loop —that will in the end be the person making any kind of

00:12:33: changes.".

00:12:35: I mean... That's so much right for people to also take in!

00:12:38: Because it's a big responsibility….

00:12:40: To wrap up –I would say like what?

00:12:43: or i would ask you?

00:12:44: What would be let's say your top three recommendations AI, diversity and ethics.

00:12:55: First be mindful always be mindful.

00:13:00: do your best to pick the right tool or to develop the tool the best way possible And that requires diversity.

00:13:09: it requires many opinions It requires expertise uh...it requires understanding what you're doing here.

00:13:16: So be mindful all the way when you want to integrate AI.

00:13:19: A second point is be transparent about it.

00:13:22: Say clearly what you're picking, why your picking it.

00:13:25: It's also about keeping you accountable a little bit.

00:13:28: but its' also the trust.

00:13:29: as we were saying Trust is big part of our sector.

00:13:33: Its' about humans.

00:13:35: And final one is Literacy.

00:13:38: Educate your teams educate everyone educate everything About What You are Implementing with them for Them And why and how it works.

00:13:51: I think today, It should be mandatory that everyone has some class on what is an element?

00:13:56: How it works because if you're going to use it You need to know exactly what it does in what doesn't do also to reduce the biases as talking about earlier.

00:14:04: Thank-you oxan.

00:14:05: i mean these are really important points.

00:14:10: i Do have one last question for you before we finish Today slightly beyond What We've Been Discussing so far Everyone is talking about the future of work and whether it's autonomous cars or robots working in factories, humanoids.

00:14:31: How do you see the role of digital co-workers or AI agents evolving?

00:14:38: Here I'd really love to hear your personal opinion.

00:14:42: You're right.

00:14:42: there are a lot things happening Now, I know that the fear behind most of those questions and this topic is the big replacement here.

00:14:51: Right?

00:14:52: We're afraid that AI's going to take over And there are not gonna be jobs for us anymore.

00:14:58: The truth Is that AI due to its lack Of basically everything That it isn't just probability Cannot do much of our job At least not full jobs Can do some bits & pieces of it.

00:15:10: And thats where actually i see future be Those bits & piece.

00:15:15: Let's delegate them.

00:15:16: There is a lot of research on human-machine teaming and collaboration, how do we integrate AI in all workflows to make our life simpler but also spend more time?

00:15:25: maybe the harder tasks?

00:15:27: The higher cognitive task or the higher human tasks... ...the feeling tasks!

00:15:32: We talk about it in healthcare as well As where can AI remove some burden especially organizationally but then can give more space for doctors, nurses to actually care for people.

00:15:45: So that's in my opinion where we're going.

00:15:48: definitely because also what you see happening is a lot of those less fun jobs being replaced.

00:15:58: You were talking about autonomous cars, that requires a bit more than just an autonomous car.

00:16:02: it requires the entire ecosystem to allow us and we know where not ready you are talking about trust or when they're there yet!

00:16:09: The population doesn't trust them but... It's also because of technology fully implemented.

00:16:16: its going be a process for humans to be enhanced by integrating technology as teams, team members.

00:16:28: Thank you for being here Oksana.

00:16:29: it was really great to share your perspectives with us today.

00:16:34: thank you very much.

00:16:35: I had a lot of fun.

00:16:37: thanks for listening to HR.

00:16:38: what's next?

00:16:39: if you enjoyed this episode make sure to subscribe.