Hello, everyone. Welcome to this session. I know that you guys have been attending quite a few sessions since this morning.And, I know that you have been learning a lot. But today, we will be talking about developing ethical, inclusive, and responsible artificial intelligence intelligence for human centered society. This topic is extremely important, especially given the fact that these days, we are talking a lot about, governing and putting guardrails around, not only artificial intelligence in general, but also around artificial intelligence or AI agents as we will call them throughout this presentation.

And it is imperative that we use AI and AI agents for our, like, for the growth of human society, for the development of human society. So we will be quite covering quite a few things today. And in interest of time, let's move ahead quickly. So this slide, I will skip completely, but I would just like, all of you, if you want to connect, to connect with me on LinkedIn. I have been in, industry working with AI and AI agents and building robots for last, almost fifteen years now. And, currently, I'm leading our, artificial intelligence product design and data sciences team at our AI and web three point o category defining start up. So, you guys can connect with me on LinkedIn, and we can take the conversation further there after this session if you will have any questions.

Why it is important to have a human centered AI or to build human centered AI agents? There are three layers to it. First of all, a combination of proper like, profit driven development and human centered development is really important. Yes. Profits and revenues are the reason, which because of which we do business, But it is also important to focus on the human development because, otherwise, we may not be able to keep, and control all check marks over the AI the way it is growing right now. The second layer comes when we talk about ethical development and deployment. So most of the people think that the guardrails part or governance part or regulation part for AI agents come in after deployment. Well, that is not completely true because it is important to have ethical development.

And then when you deploy an ethically developed system, then it is, like, compliant to the auditing mechanisms, which has been set by human in loop ahead of time. And then the final layer is responsible and inclusive AI agents. Now this is really important because an AI system, AI, agent is as good as the data that we have provided in in order to build that. And if we have not used the, ethical and inclusive frameworks and development methods, which will be discussed in a minute, then it is really not feasible to have a a system which will be able to make critical decisions, like, for example, in law, in health care, in criminal justice, which would be aligned with the human ethics as well as the right decision.

Moving on. So what are the bias challenges right now? The bias challenges which we are facing right now is, first of all, that the data on which we are building our AI agents or training our AI agents is reflecting societal prejudice. And then the homogeneous development teams is an other huge challenge in the tech industry. Whoever works in tech, and I'm sure that all of you or many of you are familiar with the tech, landscape, The teams which are currently developing, industrial scale or enterprise scale AI systems and as well as AI agents, they are very homogeneous when it comes to demographics.

They come from a very specific background of white males from a specific region. There are reasons for that, but that is not giving a diverse perspective when we are building those agents and systems. And then finally, inadequate testing across demographic groups. So I can give several examples for that. And one of the real life example for that is facial recognition, which has error rates of up to thirty four percent for darker skinned or black individuals, which means that when these systems or AI agents are built and they are tested, they are again tested on a very homogeneous group, which means that we don't have diversity of evaluations or diversity of probable results for these AI agents.

And other real life use case or example would be that recruitment AI. So most of the HR or recruitment AI agents, they perpetuate the same gender disparities which are present in the historic data on which these AI agents have been trained. So the entire point here is that the ethical and inclusive agent imperative doesn't come at a time when we are deploying these agents. Rather, it begins at a time when we have started building these agents and these systems and training them. And what is the proposed solution for a short term? I think comprehensive bias audits and diverse training datasets and continuous monitoring post deployment is really important. But what is even more important is having homogen, heterogeneous teams, development teams, as well as heterogeneous testing environment so that we can see that how these AI agents will work in the real life scenarios.

So these are three top pillars for responsible AI development. We can cover a lot of other things, but in interest of time, these are the three primary consideration without which we cannot even claim that our AI agent is ethical, responsible, and inclusive. So first is the ethical foundation, which means that design principles and the prevention protocols and the privacy preservation when it comes to collecting data right from the beginning, even when we have not designed or developed our AI agent is extremely important. Then comes the development or architecture part of the agent itself, whether it's a a multisystem agent or whether it's a single agent, but having bias detection systems, cross cultural or, again, heterogeneous validations or evaluations, and then accessibility standards to these agents that who would be using it and what scenarios they will be using it, and what would be their probabilistic, like, outcomes, which they would be expecting from these agents.

Taking that into development and architecture at the beginning of designing the AI agent is really important. And then finally, at the deployment stage, accountability and responsibility is a continuous process. It's important to have explainable AI frameworks, impact assessments of your AI agent, but that how it is, working and how it is impacting the outcomes in the real life scenarios after deployment. And then finally, the regulatory compliance, which is a whole ballgame altogether, and it includes a lot of regulatory bodies as well as governments, which currently is not happening. Now the technical implementation framework of agent AI agent, for a ethical and inclusive, deployment is pretty detailed. But in interest of time, I will just cover, like, few phases here.

Again, starting from the design phase, and design phase mind it that it is even before the development phase where you engage with the stakeholders, could be potential, users, could be potential, people who will be providing data inputs, could be potential organization who will be impact getting impacted by your AI agents after deployment.

Then also some regulatory bodies and taking all their inputs to see that how the actual system is working in the real world as of now without the AI agent, then using that data and all those stakeholder inputs and moving further to the development phase, And then building fair algorithms, detection bias tools, and explainable AI integration within the development of that AI agent.

Then comes the testing phase. Even before deployment, when you are testing the system, it is really important to check that the system is or the agent is applicable to different, demographics, to different, cultures, to different scenarios. For that, we usually define evaluations. We usually define probabilistic outcomes ahead of time, and then we test our AI agents to case those outcomes over and over again even before putting them into production. And then performance parity analysis happens during the time of production, and then that leads to final deployment phase. If there are any discrepancies, if there are any different results as compared to the defined evaluations, you change those.

You try to mitigate them as soon like, as well as possible, And then you finally go into the deployment phase. But even after deployment, because most AI agents, if they are not very simple, or, like, single chain AI agents, they are mostly probabilistic in nature. They don't have deterministic outcomes, which means that you can never ensure that there will be no hallucinations. There would be no slips, and that's why it is important to put those guardrails for continuous monitoring, providing the feedback loops for the refinement and for ensuring that the system is still comply with the defined evaluations and the predefined outcomes, which were established even before the production, development, and deployment of the artificial intelligence, agent.

I would like to focus quickly on a case study in health care scenario. Why health care? I think it is really important to understand that certain use cases are way more, sensitive, or critical. For example, health care, criminal justice, or law and order because these are the systems where the outcome impacts real lives, and the stakes are really high. So first of all, it's really important to identify the bias. Like, right from the beginning, when you have developed a AI agent, it's important to see that what type of outcomes it has providing, whether those outcomes after deployment match the outcomes which were there in the production phase or in the testing phase, whether there are any type of racial bias or systematical underestimating or anything like that because it is happening in real times.

For example, black patients compared to white patients with identical health conditions do not get, like, diagnosed or recognized by using these AI agents or AI systems. Then on the basis of that, now when you have identified that there are biases or there are challenges in the systems, then you do the root cause analysis that what is the reason for this bias. And for that root cause analysis, you again, unfortunately, have to go to the algorithm and the development phase to optimize and to see that what is happening in your AI agent, whether it's the logic problem, whether it's the system architecture problem, or whether it's the evaluation problem.

The evaluation in simple terms, we use that AI evaluation part in technical terms. But in simple terms, that the probabilistic outcomes which you have feed into the AI agent, whether the agent is aligning with those in the real world scenarios or not. So then after doing the root cause analysis, you redesign the model. And based on that, you redesign your agent as well. Sometimes it's only the architecture issue. At that point in time, you only have to redesign the architecture of your agent because the sequencing of the agent is not correct. But sometimes it's the redesign of the underlying AI model because it is mostly the challenge with the data or the way the data has been interpreted by the AI agent itself. And post deployment monitoring, again, after you redesign the model, re or redesign the AI agent or architecture, whatever is required based on that specific use case or scenario, when you redeploy it, even that, the ongoing monitoring is really important to see that whether the system is biased, whether, again, the system is providing the desired outcome or required outcome, which are ethical, inclusive, and which could be, or could not be challenged in the code of law.

So this type of process in the health care system, when has deployed, has reduced the bias by 84 per percent while improving overall prediction accuracy of that agent, especially in the diagnostic, mode. So how can we build inclusive AI systems? And as I said earlier that one of the biggest challenge is not the models itself. Yes. They are part of the challenge, and it's not the AI agent architecture. Rather, it's the teams which are building these AI agents and AI models. And we have seen that most of these teams don't have any type of cognitive diversity, and the reason for that is that many people come from multidisciplinary backgrounds, but at the same time, they are closely related to each other when it comes to their, like, educational or experiential background.

And that results in sixty seven percent more potential failure cases during development process because these people cannot when those teams are so homogeneous in terms of their cognitive background or diversity, they cannot even fathom all the possible salute, like, outcomes which could happen in the real world scenario.

Then the cultural representation part in the AI teams. And this is not like some DEI solution or DEI initiative. No. This is highly technical. This cultural representation is required for highly technical reasons. And their data has shown that for there is 43% better performance across global user population when these teams are cross cultural, cross racial, cross like, have different genders and people from different ethnic backgrounds as compared to having only white males from a certain group or from a certain geography.

And then finally, domain expertise integration. This is extremely important. Sometimes there are people who are extremely good AI builders, who are extremely good AI engineers, data scientists, and they're very good at building AI agents, let's say, or AI models or training them. But at the same time, if you are talking about a AI agent in a high stake environment like criminal justice, like law, like health care, like education systems where stakes are extremely high. It is really important to have domain specialists, domain specific ethicists, for example, in financial sector, and then social scientists to ensure a comprehensive consideration of overall societal well-being because, again, we are talking about building and deploying human centric AI agents here.

Another aspect which I would really like to focus on, though it may or may not be applicable to all AI agents, but it is usually going further, is applicable to most of the AI agents, is explainable AI. And here, again, there are four aspects to it. Right from the beginning, of course, that AI agent is based on some AI model, and model transparency is really important. The model doesn't need to be a black box. Rather, the model transparency is, achieved by applying explainable AI or by integrating explainable AI, which means that the model, when it shows you certain outcomes on the basis of a certain input, it also shows you that what those outcomes are based on. What are the data sources? What are some of the synthesis or some other details which has been done? Most of the, AI models right now are black boxes because they you give a prompt and you take an outcome, but they don't tell you what are the sources and what was the synthesis process or the thinking process of the model.

And that needs to change for a more accountable and AI accountable and inclusive AI agent, ecosystem going further. Then decision interpretability, which means that you can implement sharp values, attention mechanisms, counterfactual explanations, and a lot of other technical mechanisms which we can use or which we usually utilize to see that how this AI agent has reached a certain decision.

Providing information is a different task or outcome, and that has a different architecture when building an AI agent. But when an agent is making, autonomous decisions, especially, for example, in warfare where agents are taking decisions of who to kill and who not to kill, I mean, the those are high stakes situations. So the decision interpretability and accountability is extremely important. Then we it brings us to the audit trails. It is important to maintain the logs of all the previous decisions of all the previous information provided by an AI agent because that's the only way to have some accountability for that agent or whoever is monitoring that agent in case the situation goes to a point where we have to have some legal or regulatory intervention.

And then finally, it leads us to the stakeholder communication, which is extremely important. And continuous stakeholder communication will provide, like, the audience, whether those are users, whether those are, people who have deployed the AI agents, or whether those are people who are directly or indirectly getting impacted by the outcomes of those AI agents need to be updated that how this AI agent works and what decisions it's making on the basis of what data and what type of workflow.

So, again, integration of explainable AI will ensure to certain extent. Again, not completely because there would always be a component of probabilistic outcomes in these AI agents, especially when it comes to autonomous decision making. There would always be certain level of hallucinations, at least as of today, but we can ensure that we can minimize those instances as much as possible. And then, that brings us to governance and regulatory landscape. It is really important to understand that most people, when we talk about, for example, governance or ethical or inclusive AI systems or AI agents, they think that it starts after deployment. It starts with the regulatory body. It starts with the government. Well, that's really not true.

Because if your data as I we have shared earlier in this, conversation that if the data is not, good enough, if you are using the same data with those historic biases, if the teams are not diverse enough, if the architecture is not good enough, if the evaluations have not been defined or have been documented properly, the if the AI agent architecture has not been completely, like, integrated with explainable AI, then there is no way that the governance and regulatory organizations or governments can control it.

But just to give you an idea of what is happening today, in European Union, there is an AI act. They have a risk based classification system with the strict requirement, but the problem is that the implementation and conformity of these assessments on ground is not at a level or at a place where, probably their government or maybe their stakeholders would like it to see.

And then, in United Kingdom, because they have separate AI, laws, the pro innovation approach is there, but, they are also sort of sensitive to certain regulatory and, compliance and governance requirements. So, again, there is a they they have to maintain a balance between innovation and putting guardrails around these AI agents and AI systems. And then we have international standards like II, Tripoli, ISO, OECD. The challenge is that these system exist on paper today. Whether those systems are used or not in the real implementation, that's the question. Many of the organizations, first of all, do not have the resources to get compliant with these international or national or local standards. And even if sometimes they have the resources, they are they are not willing to comply with these standards because sometimes complying with these regulatory requirements and standards means that they have to look at the pace of their innovation of building these AI agents.

And, And, again, it is extremely important to, maintain a balance between the innovation pace and as well as putting the regulatory guardrails or compliance. Because in the end, if our AI agent is not working for the betterment of humanity, no matter whatever the use case, yes, probably the innovation is there. But whether it's human centered or not, that is the big question here. Going forward, the call to action for different stakeholders or for different groups who are working with building AI agents is pretty clear. For researchers, I think it is priority prioritizing fairness metrics, and then performance measures at the same time. They have to show that where the data is coming from. They need to put publish those datasets.

They need to take or bring those datasets into account, which have, like, consent into them inbuilt, not without, like, the the data privacy part neglecting the data privacy part, especially in health care or criminal justice or even in financial information. Who would like their bank information or financial information to go to an AI agent without their consent? If you put yourself in a, you know, end user or people who whose data is used for training these AI agents. For industry leaders, it is extremely important to establish ethics review boards upfront. This is not, like, imperative at the time of deploying a AI agent. Rather, it is at the time of deciding that we are we going to build an AI agent or multiple AI agents or AI orchestrated system AI agent orchestrated system, then in that scenario, they have to establish a ethics review boards upfront, and they have to invest in diverse talent pipelines so that different diverse mindsets can come into that development process in order to maintain that accountability and also to reduce the bias in the system along with the, reducing the bias of the data itself.

For educators, I think it's really important to focus as much on the AI agent or AI ethics as much on AI innovation. I have seen in the industry as well as in academia because I work on, at cross section of industry and academia that there is a lot of focus on AI innovation, But nobody, especially the big, organization where those are frontier models or whether those are AI agent tech organizations, they are not that much focused on the ethics itself.

And if those people who are building those AI agents are not focused on the ethics equally as much as they're focused on the AI agent, innovation, I think we have a huge challenge of bias here. And then finally, policymakers, they have to be, first of all, proactive to balance the innovation as well as the regulation part, and then they have to support the interdisciplinary research on AI societal's impact. Yes. An AI agent has been built. Yes. Had it has increased the speed of the process or a system or a thing which we were doing initially, probably manually or in another way. But at the same time, the what is the overall impact on the society and human well-being? That is really important for the policymakers as well as for the industry leaders or builders.

So that's how we just, like, approach development of ethical, responsible, and accountable AI agents from all different perspectives, including research, building, education, and policy making. And that brings us to end of our presentation today. I would be more than happy to stay and answer any questions, or or you guys can reach me out on LinkedIn, and we can take that conversation further.