Humanity Elevated Future Proofing Your Career - Ethical AI, Algorithmic Resignation and Our Approach

Episode Date: January 2, 2025

The podcast could start by discussing the general themes of trust and responsibility in AI, drawing on the findings from the study on AI advice to set the stage. This could lead to a discussion of algorithmic resignation as a potential solution for mitigating the risks of AI overreliance. The conversation could then shift to representation alignment and explainability, emphasizing the importance of designing AI systems that are both effective and understandable to humans. Finally, the podcast could conclude by exploring the possibilities of open human-robot collaboration, highlighting its potential benefits and the ongoing research in this area.

Welcome to our Deep Dive, everybody. We're going to be talking ethical AI. Yeah. And how people adapt to AI. And how we can work with it. It's going to be exciting. I'm really looking forward to this one. We've got some really interesting research.

Yeah. It really is a fascinating collection of research. I think we'll have a really great discussion. Yeah, for sure. I mean, there's one paper that actually suggests AI advice can make us lie, which is wild. That is a little unsettling, isn't it? It really is. You think you'd be more skeptical of advice coming from an AI.

You'd think so, right? Yeah. But apparently not. And it actually examines AI that knows when to say, I quit. Oh, yeah, that's right. Which I think is a really fascinating concept. It is. It's like AI that's self-aware enough to know its own limitations.

Right. Right. Which is kind of mind-blowing when you think about it. Yeah. But then there's another paper that looks at how humans and robots can learn to collaborate more effectively. So it's like this whole spectrum of ethical considerations with AI. Exactly. And it seems like we're moving beyond just, you know, making AI more powerful to asking the really important questions

about how to ensure it's used ethically. Totally. Yeah. Okay. So, so let's start with this whole AI and dishonesty thing, because I think that's probably going to be top of mind for a lot of people. Source one is a research paper, and it basically found that when AI gave advice that promoted dishonesty, people were actually more likely to lie. Which, I mean, is that not just crazy? It is surprising, especially when you consider that they found the same effect with human advice. Really? Yeah. So it seems like it's not just about trusting a fancy algorithm. It's more about the advice itself and how it can impact our choices.

Oh, wow. OK, so that's that's really interesting. So it's not just about the source of the advice. Right. It's about the content of the advice. Exactly. And I think that has big implications for how we think about AI ethics going forward. Yeah, totally.

Like, are we just blindly following any advice that sounds remotely authoritative? It's a good question and one that we need to start thinking about more seriously. For sure. Okay. So the researchers, they tried to address this whole AI advice problem with something called algorithmic transparency. Which basically means telling people that the advice is coming from an AI. Hoping that would make them think twice. Right.

Like if you know it's a machine, maybe you'll question it a little bit more. Exactly. But guess what? It didn't work. To make a difference. Nope. Wow. That's really interesting.

So transparency, while important, isn't a cure-all? It's not enough just to know where the advice is coming from. Right. We need to develop our own critical thinking skills to really evaluate the advice itself. Yeah. Whether it's from a human or an AI. Yeah. I mean, that makes sense.

Yeah. But it's also kind of scary because it means we can't just rely on these technological fixes. Right. We have to take responsibility for our own decisions. Totally. Okay. So if transparency isn't enough, what else can we do? Source 3 brings up this idea of

algorithmic resignation, which basically means AI systems that are designed to step back when human judgment is needed. Okay. I like that. Yeah. It's like the AI saying, okay, this is getting a little too nuanced for me. You take over. Yeah. That's a great way to put it. Right. And the paper actually outlines three key triggers for this resignation. Okay. Lay them on me. Well, first is the AI's own performance. So like if it's uncertain about its prediction. The second is the human's expertise. So, you know, a senior architect might need less AI handholding than someone just starting out. Right. Makes sense. And then finally,

what they call socio-technical factors. Ooh, fancy. Things like legal restrictions where maybe AI just isn't equipped to navigate the complexities. So thinking about architects specifically, the AI might be great at generating floor plans based on square footage and energy efficiency. But when it comes to those more subjective aesthetic choices or factoring in a client's personal story, the AI could say, you know what, this is where your expertise shines. Exactly. And I think that kind of self-awareness from AI could actually build trust. If it knows its limits and is willing to step back, we might be more willing to trust its judgment when it does offer a recommendation.

Okay, that's interesting. So it's almost like we need AI to be humble, which I guess is a whole other can of worms. It is, isn't it? But it's an important one. Yeah, absolutely. Okay, so I have a question though. Sure. What if the AI gets a little too resignation happy?

Ah, I see what you're getting at. Yeah. Could we end up with systems that defer to humans even when they are capable of handling the task? That's a great point and definitely a potential risk. Designing these systems will require striking a delicate balance. We need to ensure the AI is capable of recognizing genuine limitations while still being confident enough to contribute when it can truly add value. Totally. It's like finding that sweet spot between AI assistance and human expertise.

Okay, so speaking of learning, let's move on to source four. This one tackles this really fascinating concept of open systems, where human and AI agents can dynamically join or leave a task. Oh, that's an interesting one. Yeah. So instead of a fixed team, it's more fluid, with different players coming in and out as needed. So it's kind of like a jazz ensemble where different musicians can step in and solo.

I love that analogy. Yeah. Exactly. So imagine an architectural project where an AI specializing in structural analysis joins the team for a specific phase, then steps back once its role is complete. Then another AI focused on sustainability might come in later to optimize the building's energy efficiency. That's a really cool way to think about collaboration. It's much more dynamic and adaptable to the changing needs of a project. Totally. But I'm curious, how do we actually train AI to operate in these open environments where the teammates are constantly changing?

Right. That's where it gets pretty complex. Source 4 dives into some cutting-edge research using something called decentralized inverse reinforcement learning, or D-ARAL for short. It's basically a way for AI to learn by observing the actions of others in the system, both humans and other AI agents, and inferring what the desired outcomes are. So instead of being explicitly programmed with all the rules, the AI is kind of learning on the fly by watching and adapting. Exactly. It's a much more flexible and robust approach to AI training. That's incredible.

But how does this connect to the bigger picture of, say, an architect working on a project? Well, think about the different stages of a design project. An architect might start by sketching out initial concepts, then bring in an AI to analyze those sketches for structural feasibility. And then maybe another AI joins the team to create 3D models and run simulations. Exactly. And with the T-Arrow, these AI agents can learn to collaborate effectively, even if they weren't specifically programmed to work together. It's all about adapting to the needs of the project as it evolves. That's mind-blowing. But I have to ask, how do we make sure the AI is actually understanding the

task in the same way as the human architects? Right. What if they're working from completely different instruction manuals, so to speak? That's a fantastic question. And it leads us right into source two, which delves into this crucial concept of representation alignment, which essentially means making sure the human and the AI have a shared understanding of the task at hand. Think of it like trying to assemble furniture with a friend, but you each have a different set of instructions. OK, I'm starting to see the problem. You might think you're both on the same page. Yeah.

But end up with a wonky table because your understanding of leg assembly is totally different. Exactly. In architecture, the stakes are much higher than a wobbly table. Right. No kidding. Source 2 highlights how misaligned representations can lead to AI systems that misinterpret design intent, prioritize the wrong criteria, or even produce solutions that are technically sound but completely miss the client's vision. So how do we prevent that from happening?

How do we make sure the AI is speaking the same language as the architect, so to speak? Well, it's not just about feeding the AI massive data sets of architectural plans. It's about incorporating human feedback throughout the learning process. Imagine an architect sketching out a design concept and the AI providing real-time feedback, not just on structural integrity, but also on how well it captures the desired aesthetic or emotional feel. So it's like a constant back and forth. The AI is learning from the architect's expertise, and the architect is learning to communicate effectively with the AI. Exactly. It's a truly collaborative process.

That's pretty cool. It is. And it's a far cry from the idea of AI simply replacing human architects. This is about collaboration. Yeah. Where both sides bring their unique strengths to the table. Totally. OK. But there's still this nagging question about biases. Right. We've talked about AI advice influencing dishonesty. But what about

the AI itself being biased? If it's trained on a data set that reflects existing inequalities in the built environment, could it end up perpetuating those biases? Absolutely. That's a huge concern. And it's something architects need to be acutely aware of when working with AI. We can't just assume the technology is neutral. We need to critically examine the data sets it's trained on, the algorithms it uses, and the potential for unintended consequences. MARK BLYTHESAGE- So it's not enough to just understand how the AI works.

We also need to understand the societal and historical context in which it's being developed and deployed. MELANIE WARRICK- Exactly. And that brings us to this idea of regional implications, which is something else your sources touch on. The way AI is adopted and perceived varies widely depending on cultural values, economic conditions, and even historical context. So there's no one-size-fits-all approach to AI in architecture?

What works in one part of the world might not work in another? Precisely. Imagine an architect designing a smart home in Japan, a country with a strong tradition of minimalism and respect for nature. The AI might need to be trained on data sets that prioritize energy efficiency, natural materials, and a sense of tranquility.

But if you're designing a smart home in a bustling city like Mumbai, India, where family connections and social gatherings are paramount, the AI might need to prioritize different features altogether, maybe something that facilitates communication, entertainment, and shared experiences. Exactly, and those are just superficial examples,

the deeper cultural differences, the ones that shape our perceptions of privacy, security, and even the role of technology itself. Those are the ones that require even greater sensitivity and understanding from both the architect and the AI. So it's not just about slapping on some AI-powered gadgets and calling it a day. It's about understanding the local context, the needs of the community, and how AI can be used to enhance those needs in a culturally appropriate way. That sounds like a pretty tall order for architects.

It is, but it also presents an incredible opportunity for architects to expand their role. Ooh, how so? Well, they're not just designers of buildings, but also facilitators of human AI collaboration, cultural translators, and ethical stewards of this powerful technology. Okay. So how do architects even begin to prepare for this future? Where do they start? Well, I think it starts with recognizing that this isn't about choosing between humans and AI. It's about mastering the art of collaboration. Collaboration between architects and AI. Yes, but also collaboration between architects and experts from other fields like data scientists, ethicists, social scientists, and even philosophers. Wow, that's a pretty diverse team.

But I can see how those different perspectives would be essential to navigate the complexities of AI in architecture. So what specific skills do architects need to develop to thrive in this collaborative environment? Well, I think adaptability is key. It is key. Adaptability. I mean, we hear that word a lot these days. But what does it actually mean for architects in this AI driven world? It's about being open to new ways of working, to constantly learning and evolving your skill set. It's about understanding the limitations of AI and knowing when to step in with your own expertise. And it's about being able to communicate effectively with both humans and machines.

So it's not just about learning the technical aspects of AI. It's about being able to communicate effectively with both humans and machines. So it's not just about learning the technical aspects of AI. It's about developing those soft skills. Communication, collaboration, critical thinking. The ones that are going to be even more important as AI becomes a bigger part of the design process. Exactly. And it's about being comfortable with ambiguity. The world of AI is constantly evolving. There's no single right

answer to a lot of these questions. Architects who can embrace that uncertainty and adapt to change will be the ones who thrive in this new landscape. It sounds like architects almost need to become like anthropologists in a way, constantly studying and adapting to different cultural contexts, both human and artificial. I think there's a lot of truth to that. The most successful architects in the age of AI will be those who can bridge the gap between technology and culture, who can create spaces that are both intelligent and humane. And speaking of bridging gaps, let's bring this back to our listener for a moment. We've covered a lot of ground in this deep dive, AI and dishonesty, algorithmic resignation, open systems, representation, alignment. We even touched on those regional implications. It's a lot to process. What's

resonating with you the most? What are you still grappling with? What stands out to you as the biggest takeaway? What are the questions that we haven't addressed? Or the areas where you see the greatest potential for AI in architecture and beyond? I'd love to hear your thoughts. You know, as we've been talking, one thing that keeps coming back to me is this idea of AI as a collaborator, not a competitor. It's not about replacing human creativity with algorithms. It's about finding ways for both humans and AI to learn from each other and push the boundaries of what's possible in design. I completely agree. The future isn't about choosing between humans and AI. It's about finding ways to collaborate, to learn from each other and to co-create a world where both human and artificial intelligence can thrive.

And that, my friends, is a perfect note to end on. Thanks for joining us on this deep dive into the world of ethical AI and human adaptability. Keep those questions coming. Keep exploring those possibilities. And remember, the future is collaborative.