Talk To Me Petey D

Ep. 58: Six Principles of Responsible AI

• Petey D • Season 1 • Episode 58

Share episode

Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.

0:00 | 28:06

What does “Responsible AI” actually mean—and why does it matter for builders, leaders, and users of AI systems?

In Episode 58 of Talk To Me Petey D, we break down the six core principles of Responsible AI and explore how they apply across the full lifecycle of AI systems—from design and development to deployment and ongoing operations.

Responsible AI isn’t just theory—it’s an industry best practice and a critical framework for balancing innovation with ethics, risk, and real-world impact.

🔍 In this episode:

  • What Responsible AI really means (and why definitions vary)
  • Why AI systems need guardrails—not just rules
  • The tradeoffs between minimizing harm vs. eliminating it
  • How Responsible AI functions as both an ethical framework and a business risk strategy
  • The importance of ongoing monitoring and operational accountability

🧭 The Six Principles of Responsible AI:

  • Fairness – Defining and measuring equitable outcomes
  • Reliability & Safety – Ensuring systems perform as expected—even under edge cases
  • Privacy & Security – Protecting valuable and sensitive data
  • Inclusiveness – Designing for diverse users and real-world scenarios
  • Transparency – Understanding how and why decisions are made
  • Accountability – Humans ultimately own the outcomes

Whether you’re building AI systems or just trying to better understand how they shape the world, this episode will give you a clear, practical foundation for thinking about Responsible AI.

🔗 Connect with me:

🎧 Listen to the podcast:

👍 If you enjoyed the episode, like, subscribe, and share it with others interested in AI, leadership, and the future of knowledge work.

SPEAKER_00

Responsible AI refers to the practice of designing, developing, and deploying artificial intelligence systems in a way that is ethical, transparent, fair, and accountable, while minimizing harm and maximizing societal benefit. At least this is how Copilot defined it for me, but I think this is a pretty good definition of the industry best practice. So as we continue along in our series of topics in AI literacy, we'll now dive into responsible AI. Welcome to the Talk to Me PDD Podcast. I'm your host, PDD. Talk about all things technology and society, knowledge work, and of course artificial intelligence. Today is episode 58, The Six Principles of Responsible AI. If you enjoyed the show, please like and subscribe. Follow me on LinkedIn and Blue Sky. Subscribe to my newsletter at peterdepseywrights.com. Put all the links in the show notes. Really appreciate your support. Love to hear back from you if these topics are resonating, if there are other aspects of AI or knowledge work, leadership tech that you'd like me to talk about. Love to hear about it. So now, why do we call it responsible AI? Why do we need responsible AI for artificial intelligence systems? Well, especially with large language models and generative AI systems that are built on top of that, they're really open-ended. And I mean, that's one of the powerful things, and I think things that people like about it. If you go to ChatGPT or Copilot or any or Claude, any of these systems, you can type in pretty much anything and get a response back, which is very powerful and useful. But it also means there's this huge range that you may want to curate as some building an application on top of these large language models. You may not want to return certain types of information or certain types of response within your own application that the large language model or another type of AI system could do. So responsible AI is a way to put some constraints or attempt to put some constraints on the vast scope of what's possible to be returned and to make sure that's in line with a certain set of values or a certain set of profile that the application designer wants for their given application. Responsible AI, you know, is is an industry best practice now within companies that are that are building AI solutions. You may see slightly different iterations in different places, something being discussed in in academia and may have slightly different tilt there. But it's something that you need to consider if you're going to be implementing responsible or any type of AI systems that you want to be responsible. And we'll talk about why you would you would do that next. And it could be for a variety of different reasons. But let's go in and break down this definition, and I think that's a good starting point, and then we'll get into more generally the kind of the guidelines around responsible AI from there. So we'll start with with different pieces of it and look at this here. So the first part, responsible AI refers to the practice of designing, developing, and deploying artificial intelligence systems. So there are different stages where these systems are being created, and responsible AI practices can apply at all of those different stages. So even from the beginning, when you're building out sort of the architecture, thinking about what data you might want to use to build your model, all sorts of different things, different types of uh tuning. So within kind of setting up that framework, within building out your model, actually developing it, and any pieces around that that interact with the model, and then ultimately deploying it, having it go live in an application. At each of these stages, you want to be thinking about various responsible AI practices. I would also add on to this definition that when you're once you've deployed, you're not done with responsible AI. Um, it has to be part of the operational model. Um, and we'll talk about that in a little bit, but just because things have checked out with all of your responsible AI practices when you're at launch, for a variety of reasons, things could change in the future. And if you're not using an operational and monitoring metric to keep track of that, you may miss those changes and may end up causing harm or displaying information or actions that are not aligned with your intents. The next piece of the definition that we should look at is in a way that is ethical. Um, ethics may be defined differently by different people. So that's important to bear in mind here, whereas um this framework is saying that to be responsible, your AI needs to be deployed within an ethical framework, but it doesn't specify what that ethical framework is, um, which is good in a sense is that it it leaves it open to some interpretation, but one person's responsible AI ethics may be different than another. So that's important to bear in mind here that just because two different people or organizations have gone through responsible AI practice on a product or solution they're releasing that uses artificial intelligence, doesn't necessarily mean that their ethics are going to be the same and their uh solutions will perform with the same sort of ethical profile. The next piece is transparent, fair, and accountable. And we'll go into these more in detail when we get into the responsible AI principles, but these are also open to interpretation depending on um you know what you want to show to different people. This idea of fairness, again, like ethics, is not universal. Um, there are different ways that you can decide if things are fair. Um, you know, kind of in the economic sense, this is a matter of how you're allocating resources, and there are different ways to do that. Um, so it's important that your organization has a definition of what this should be for a given solution, or even more broadly within your AI, but also that your organization may have different values and different definitions of fairness than others do. The next piece is while minimizing harm. So what's really interesting here to me is this idea of minimization as opposed to preventing. And I think that speaks to the difficulty in constraining um what big AI models can do, because it's essentially impossible to um check every single possible output. Um, and since you may get different outputs for the same input, even if you did, there may be things that you missed. So ways to try and limit that harm, um, but by saying we're going to minimize harm that leaves uh the possibility that our system may still cause harm, even though it's gone through these responsible AI principles and best practices. So something to bear in mind and something to consider. Um it's a trade-off in using some of these systems potentially, and you'll have to decide if that's okay and what those potential harms are in a given situation. Uh, the next piece is and last is in maximizing societal benefit. So I think this is really interesting here because many of the companies or industry that are using responsible AI are for-profit industries where they are beholden to uh shareholders as uh above usually societal benefit. So it's interesting that this is in these responsible AI practices and in this definition of responsible AI. And I don't know that that's always going to be the possible in the design of these systems when they're coming for from for-profit industries. Um, it may be sort of a nice to have or part of that principle, um, but I think it's likely, and the reality is that many for-profit institutions that are building AI solutions are not necessarily going to be maximizing the societal benefit, even though that's part of the definition or the common definition of responsible AI. And when you're looking at building solutions that are under this responsible AI umbrella, you can think about it in a couple different ways. You can think about it sort of within that ethical framework that we talked about in ways that are beneficial to people and to society. And that and that's one lens that it can operate in and can operate in effectively. Or you can also, if you're a little more cynical or a little bit more based in the real world of companies trying to make profit, you can think about this as a framework for risk mitigation, where if your AI systems are causing harm or not protecting user information, for example, that may hurt your business in terms of profit. So you can use a responsible AI framework both in the in an ethical framework, but also as a risk mitigation framework for businesses. When we talk about responsible AI, or you hear people talk about responsible AI, um, you often hear the term guardrails and using this instead of rules or boundaries or kind of a stronger uh delineation term. And I think the reason for this is that guardrails attempt to keep you on the road. Um, if you think about guardrails on the side of a highway, but if you drive a car fast enough through it, you can go off the road. And I think that's um a good analogy here where you have these um responsible AI practices in place and you're attempting to keep um your AI solutions on the road. Um, but there are ways that they can go off the road and can sort of get around some of these um lanes that you'd like to keep them out of. Um so whether that's okay or not, I think that depends on ultimately what your application is doing and if you can tolerate those deviations. And you know, going back to the original point, if you have good operational oversight on your system and you can detect when things are going out of bounds and intervene, um, then it makes it a little bit easier to manage. Um, whereas if you're not investing in doing that, um you you may be exposed to some risk there or exposing users to potential harmful interaction. So within responsible AI, it's common commonly broken down into six principles. Uh, the principles of fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. So we'll go through these briefly and talk a little bit about what they mean and how they you may see them implemented in responsible AI practices. Um so the first one is fairness, that the idea that these AI systems should treat people fairly. And again, your definition of fair matters here. So technology is not uh neutral, there's not one definition of fairness, so this has to be um something that's decided. If you don't decide it for yourself, you'll get what is in the model, um, which again is sort of based on how these things were built and reflect the the values of the people that built them uh and the owners and all that sort of thing. So if you're not going to define your own ideals of fairness, you'll inherit them, and that may or may not be what you want. Um now, why do you have to think about these things with AI models? Is there are many different biases that that show up. That's just the reality. You you can't get around it. There will be biases in the data used to train the model, there'll be biases in in the model themselves and the way that they're they're trained and adjusted, regardless of the data that that's in them. And then also on the outcomes. And I think as an application owner, and that's who's implementing responsible AI practices, focusing on the outcomes in terms of fairness is probably the most important piece, um as opposed to sort of the individual components that you're building upon, making sure that the system that you've designed is generating outcomes that align with your ideals of fairness and making sure that you're monitoring that not just as you're building the system and designing it and getting it ready to launch, but as it as it's running, to making sure that that's a key performance indicator for the health of the system, um, that its output and whatever actions it's doing align with your fairness principles. The next one, reliability and safety, so that your system performs how it's supposed to, under both expected input and unexpected input. Because large language models and AI systems can take in anything, um, there's a lot of creative ways that people can intentionally or unintentionally get these systems to do things that you may not want them to. Um, whether that could be intentionally trying to generate harmful output or uh information about things that you don't want um kind of represented on on behalf of your organization or your product, or it could be unintentional, right? Just trying to use the system and maybe doing it in a way that wasn't accounted for, that generates some sort of output that you weren't prepared for that could be harmful to the user or maybe harmful to you and your brand within the organization. So making sure that you're testing for these conditions and then also monitoring for it when you're in production. The next one, privacy and security. A lot of this is, you know, for businesses, this is compliance. There are particular uh data privacy protections and laws that you'll need to make sure that you comply with. Also, just being a good citizen of user data, how you're how you're storing data that's going into these systems, whether you're storing kind of underlying data that you're using to train models yourself or user usage data. Um, data is incredibly valuable right now, more so than ever. Um so making sure that you're storing that in ways that are both compliant and making it difficult to um be exposed through bad actors or in other ways through through um user interaction. Um there are ways that bad actors can intentionally try to exfiltrate data while um talking with different AI systems. So something that you want to make sure that you're guarding against within um your responsible AI practices. Um, kind of an interesting area that I'll just touch on here, but won't really get into in too much detail, is this concept of different differential privacy where you may be building an AI model with user data and you can't tie any individual person and their data back to the model. So, in a sense, that that's good, right? You've um you're maintaining that user's privacy, uh, but then these systems are using to make future predictions off of um kind of archetypes and combinations of data. Um, so it's a little bit of an interesting area, and that's why this data is so valuable and why many systems want to be able to train AI tools and AI models off of your data and your interactions. So while it may not tell anyone about you specifically, it may be used to make predictions about people that sort of match or look like you, at least in terms of a data perspective. Um, so this is kind of I think an area for interpretation on whether you think this is ethical or not, or whether you want to consent to it. Um, but bearing in mind, just because an AI system does not sort of tie your personal data to you as an individual, um, it may be using your personal data to make predictions about others, and that may or may not be something that users want to consent to or are okay with. Um next up on the list of principles, we have inclusiveness. So, does this solution work well for you know people of all different backgrounds and abilities? Um, this is kind of generally under inclusive design with digital systems. Um, but also thinking about as you're building and testing systems, having diverse teams with diverse abilities and backgrounds both to build and to test these systems before they go into production is going to give you a better opportunity to think about the various scenarios and various types of people that may be interacting with your system and giving you a chance to account for those before you go into production and potentially uh don't include a certain type of user, or maybe it doesn't work as well as it should for that type of user base. Uh next up is transparency. So, do these systems make decisions or take actions in a way that are understandable to the audience that they know why a particular decision was made by the AI system? Um are those decisions then auditable? Do you have a log and can you go back and check and see why a particular decision was made? And this may not even this may not be an AI system the users are interacting with directly, and maybe behind the scenes and something like uh a loan approval system or job reviews for a job application systems looking at resumes for a particular job. Can I go in and see why were users screened through or not screened through or rejected for particular positions? So having transparency there, you should be able to go in and audit those decisions. Should also be clear what are the capabilities and limits of your system and how AI can be used there. What can AI do? Um, and some important questions here are you know, who is this transparency for? Um, not all digital systems are totally transparent to the end users. Um if you go even just run an internet search, you you can you may not know exactly why results are or where they are in a particular result set. Um, so you might understand generally how it works, um, but just because you want to have this value of transparency, ultimately the the owner of the system is deciding who it's going to be transparent to. Um so going back to let's say insurance approval that may be transparent to people working at that company fully, and maybe part of it is transparent to the end user. So, again, this is something where this is a general principle, but how you specifically implement it, that's something that the application designer is going to you know get to decide. Um and also here, I think one of the things to be careful of is there are systems that are based on various AI solutions that are sometimes sold with the promise of being able to do things and predict things in the future that aren't technically possible. Um so I would say that those would fail under this principle of transparency. So even if you're not necessarily designing systems on your own, if you're thinking about buying them from a third party or implementing them in your organization, um this principle of transparency should apply to how those decisions are being made. And then you may have to kind of suss out and determine is this really something that can be predicted based on the data that's available? Um, and just because somebody claims it's possible with an AI system um doesn't mean it it works in reality. So it may make some decision based on the data, um, but whether that is a reliable prediction or not, or reflective of the real world, um that's often something that you as a stakeholder will have to determine. And last, in In the six principles is accountability. Um, I think there are a number of examples of where um companies making AI solutions are trying to skirt this a little bit, where you'll have disclaimers like their um solution is for quote unquote entertainment purposes only, or we're looking at health-related applications as wellness apps, so they don't fall under the same strict regulation and controls that a health application would. Um so I think those are ways to sort of skirt this principle of accountability. Uh, ultimately, you know, humans are responsible for these systems. It doesn't matter kind of what decisions the the AI makes. Ultimately, um, I think for us to have the best solutions for people and that societal benefit, uh, there needs to be accountability. We can't just say, oh, well, you know, the AI said to do that and something bad happened, but you know, what are you gonna do? Um, so having organizational and people actually responsible for the results of these systems is important. Um having oversight by people, and this goes back to kind of that operational aspect, making sure you have KPIs, you're monitoring this system, and you have people that can go in and intervene if, or even automation that can intervene if these KPIs fall out of a healthy range. Um, so those are the six principles, just a quick quick overview of that. Um a couple other topics that I'll add on here before we wrap up is with all of these things, it's it's important to think about PCP or business continuity practices. So you often will have to think about this in people-supported applications, right? Like let's say there's the natural disaster in some area where people are working supporting a service. Do if I have a business continuity plan, I'm going to need to have a plan that I can follow where maybe people in a different area can come online and help out or work in a different time of day. So if things go wrong and they inevitably do, um, then I'm not scrambling at the last minute to try and figure out what to do. That I have a plan in place that I can uh put into action to mitigate any particular outages. So why would you want to do this as well with AI systems? Because those, I mean, those are automated, you can have various ways to have digital failover and all that sort of thing, where um in theory that shouldn't happen as often. Um, but if you're really taking responsible AI practices seriously and you're treating them as an operational health metric, um, you need to have a plan in place that if they fall outside of what's acceptable, what do you do? Are you gonna take the system offline? Do you have a backup? Um, do you have, let's say, a rules-based system that you can put in place temporarily to fill in for some of these things? So deciding what you're gonna do in advance, um, because if you don't do that, I think the temptation is just to kind of leave things be while you try and figure them out, which is potentially exposing your users to harm, exposing your organization uh to risk. Um, and you want to make sure that you're not doing that just because you haven't put together a plan and decided where um where you're willing to make trade-offs and where you need to kick in with another solution. Um and why do you need to to opera operationalize responsible AI? Well, these systems are built on data and different data comes in, uh, both in terms of in the models themselves and in the types of of input that users are putting in. Um, and then if you're you know pulling in any sort of organizational data or you have um retrieval augmented generation or different ways of adding content in, that content is likely going to change, and you may get different types of output than you have tested for before launch, um, and getting different types of results. So it's natural that these systems will change some over time, that you may have models that are trained on particular sets of data, and then that data is changing and it no longer matches as well. Um so this is a natural part of maintaining um AI systems or data science-based models, whether it's generative AI, large language models, or other types of things. So that applies here as well. Um, so yeah, that's that's a quick overview of responsible AI practices, the six principles. Uh hopefully that that helps get a better sense of sort of kind of how industry is thinking about these things. Um, again, the six principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. And then understanding that within those six principles are on top of them, these um kind of specific definitions that are ethics-based, um, you need to define those yourself. You need to define what fairness means for your organization, what accountability means, all those types of things. Um, and then live with that as an organizational operational practice and make sure that you're maintaining your AI systems once they've launched and that they're within these boundaries. And if they're not, that you have predefined actions that you can take so you can avoid causing more harm and risk there. So that's it. More in the AI literacy uh series here. Um, hopefully you enjoyed the the show. Uh, you know, please like and subscribe to the Talk2Me PDD podcast. Follow me on LinkedIn, Blue Sky, subscribe to the newsletter on peterdempseywrights.com, and I'll have all the links in the show notes. Uh, good luck with your responsible AI journey, and we'll see you on the next time.