I'm Ekta Soni. I'm working as a head of change and transformation in worker groups group. I'm executive coach and a UN woman UK delegate.The topic that I'm going to speak today is AI is not biased, leaders are, rethinking responsibility in AI. So open your eyes right now, an algorithm just made a decision about your life, your loan, your CV, your medical risk score, your insurance quote. You weren't in the room when it was built. The question is, who was? Who was in the room? Where we begin, a two story. Let me tell you about a resume that never got right. In 2014, Amazon built an AI to find the world's best engineers, reigned on ten years of hiring data. And in 2018, Amazon quietly killed the project. The AI had been penal penalizing resumes that contained the word woman, as in women's chess club captain.

The headline blamed the algorithm. I want to ask a very different question. Just a bit about myself, from the boardroom to the AI debate here. I have twenty five years of experience. I'm working in strategy tech and MA, integration space. Worked with Wipro, HCL, Alton Cognizant, Emirates Airlines. Few of the organizations, very closely work with CXO advisors, advisory board, in HSBC, GSK, BP, JLR, Barclays, Deutsche Bank. I have scaled businesses. I've run four divestment programs, recognized by the industry as a computer weekly top 50 influential women in UK last year and have been in the, winner of six more awards. I'm part of, UN Women UK delegate European Women on Boards champion for women in AI leadership and tech women for boards. A little book little fun facts about me, I have traveled to 48 countries. I'm a yoga teacher and instructor, Reiki teacher. So and I know four languages.

My day job teaches me to shift change. The other things teach me to slow down and ask better questions. Both are about to matter. Alright. Starting with when AI fails, we point to the data, to the algorithm, to the math. We blame the data. The training set was skewed. We blame the model. The algorithm learned bad patterns, and we blame the math. It's a black box. We can't explain it. Convenient explanations. Right? They put the problem somewhere technical, somewhere distant from us. Is that close to the fact? Not really. But what if AI isn't biased and leaders are? Long before a moral is trained, humans decide what problems matter, which data counts, and whose voice gets a seat at the table. That is where the bias is born. So three decisions made by leaders that shape every AI system as of today.

Firstly, which problems get solved, which data gets prioritized, and which and whose voices are in the room. When we come when we talk about the problems that get solved, a fraud detection model is built because someone funded it. AI use cases reflect a leader's priorities, not a neutral universe of needs. And which data gets prioritized? What counts as a good data? Who decides that it is a good data? What sources are trusted? Which populations are sampled? These calls are made in a conference room long before the actual training begins. And whose voices are in the room? Who reviews this design, challenges the assumptions, name the edge cases? Nobody else sees bias compounds in the homogeneous rooms and shrinks in the diverse ones, and that's not fair. This isn't hypothetical. It's already shaping your life.

I will walk through four of the use cases, one in the hiring industry, one in the health care, financial services, Apple card credit limits, and how criminal justice is also impacted. So four industries, but one pattern, and that's what we're going to talk about today. In each one, the bite bias wasn't in the code. It was in the room much before the code was written. So let's start with the hiring one, Amazon 2014 to 02/2018, the AI that learned to reject women. Goal was to automate hiring at scale. Method, freeing ten years of resumes. Problem, those ten years were heavily 60% of were male resumes. What did the AI learn? Penalized the word women, downgrade graduates of women colleges, and reward male coded verbs, executed, captured. The leadership lens, free signatures sit underneath this failure. Who decided to use the ten years of biased data as ground truth? Not the algorithm, a leader who decided not to audit before deployment.

Again, not the algorithm, a leader. And who decided this trade off was acceptable. Again, my friend, not the algorithm, a leader. The algorithm that decided who deserved care in the health care sector, a widely used US health care algorithm decided which patients qualified for extra care. It uses health care costs as proxy for the medical need. The flaw? It was a leadership flaw. It was a proxy itself. Black patients historically had less spent on their care for the same conditions, and the algorithm read that as they need less? No. The cost 46% fewer black patients flagged for additional care than should have been, and that's just not right. Third case study, financial services, Apple card. 2019, same household, same income, but 20 times more credit cap. It's the algorithm, Apple's official response. So same joint taxes, same accounts, same assets, his limits versus hers.

The regulator investigated, and the bank apologized. No one ever explained what the algorithm actually did. Who signed off on a credit mode model nobody should could explain to a customer. That signature is where the bias becomes a policy. The fourth case study, criminal justice. The algorithm that your future. Justice use compass, risk scores to inform bail, sentencing, and parole. A higher score can mean more years in prison. So if you see, the statistics show that the black defendants versus white defendants, in black defendants, more likely to be falsely flagged as high risk, and in white defendants, more likely to be falsely flagged as low risk. Is that fair? No. The algorithm was great. It was also unjust. Four industries, four failures, one signature beneath all of them. Hiring, leaders choose the data. Health care, leaders choose the proxy. Financial services, leaders choose the metric. And criminal justice, leaders choose to deploy.

If a human signature precedes every failure, responsibility lives upstream of the model and not in the AI. Who is actually in the room? That's the leadership gap. If you see today's statistics, 22% of AI and data professionals worldwide are women. Only 22%. That's the World Economic Forum, statistics from 2024. 14% of authors at top AI conferences are women. That's it. And less than 10% of senior AI leadership roles are held by women. Why? And one in four AI ethics team report no women of color representation. This is this is what we are this is the world that we are living in. If leadership shapes AI and mostly this look the same, I have no grudge against Davidson Johns of the world, and I'm I I have very close friends who are Davidson Johns. So no no personal, you know, comments. But do you see? The bias is structural and not technical, and we can't blame AI for this. Why it matters? It matters.

Diversity isn't a checkbox. It's a design input. If leadership is homogeneous, we'll see same blind spots embedded in every model, end cases, anything outside their room. Bias debated as PR, not a design failure. Speed mistaken for progress, and failures arrive as scandals, not bugs. But if leadership is diverse, more questions get answered before shipping. Eth cases surface. Someone has lift them. Ethics becomes a habit and not a panel. Systems work more for people, not for fewer people, and risk shrinks and trust scales and adoption follows. And that's the reason why lot of AI projects are failing and not being adopted very well. So we need to come completely look at from a very different lens. What good looks like when the room changes, the model changes.

A black female researcher couldn't get a facial recognition system to detect her face unless she wore a white mask. She didn't blame the algorithm. She asked who built it. She published gender sheets, showed IBM, Microsoft face miss misclassified darker skinned women up to thirty four point seven percent of the time versus less than one percent for lighter skinned men. Within eighteen months, those companies retrained their models. IBM later exited the market. One leader, one question, and one industry shift. Before thirty four thirty four point seven percent error rates for dark skinned women, whereas after this, you know, this change in the policy, less than 5% within eighteen months of audit. This is how we need to work. This is how we need to raise our voice, find a seat in the table. Quick round, myth versus reality. Three things people say, three things that aren't quite new. Myth, AI is neutral. It's just maths. No. Maths is neutral. The choices about what to count are not.

Myth, more data fixes bias. Truth, no. More of the same data multiplies the same bias faster. And the myth, third myth, this is an engineering problem. Absolutely not. It's a leadership problem with engineering symptoms. So drop the comfortable myths, pick up the harder questions, and let's address that. So what you leave with today, four takeaways. Everyday decisions shape AI outcomes. Procurement choices, vendor selections, performance reviews, each one quietly trains the system. Secondly, influence AI without writing a code. Question the brief. Demand the failure modes. Ask who isn't represented in this test data because that's very important. We can't leave anyone out of this game. And thirdly, step into AI leadership roles. Don't expertise. Domain expertise, sorry, is the credential we keep undervaluing. Translate it, name it, and claim a seat. No one else will claim a seat for you. And fourthly, diverse leadership, responsible AI.

Not because the diversity is fair, because it builds systems that work for more of the world, not just the one section of the world. The playbook. How to influence AI without writing a line of code? First, ask the missing question. Who does this model fail for, and why is this acceptable? Naming the gap reframes the room. Secondly, demand the data lineage. Where did it come from? Who consented? And what's missing? Provence is policy. Thirdly, bring the user in early. Insist on lift experiences input at design, not as a focus group after launch. It's not a survey that we need to do at the end because the again, the results will be skewed. Fourthly, refuse the false trade off. Speed or fairness is a manufactured choice. Build covenants into the velocity, not after it. Remember, you don't need to build the model.

You need to shape the build brief, fund the work, and define what done looks like, and done looks like for everyone, not just for one section again. Step into leadership. You're already qualified. You just haven't claimed it. Translate. Reframe your domain expertise. Whether it's risk, HR, compliance, customer ops, health care, every domain is now an AI domain. Your context is demo it. Position. Get visible on AI governance. AI ethics board, vendor reviews, model risk committees. The seat exists. They're quietly being filled right now. Find your seat and amplify. Sponsor the next woman in. Leadership scales when we open doors behind us. Mentor one, site one, recommend one, and repeat. We need to bring everyone together. It's a journey that we need to lead together. The next breakthrough in responsible AI won't come from a model. It will come from a leader. So be the leader.

Bring one with you. Thank you so much, and that's it from my side. Any questions? Happy to address, and take it on. Otherwise, I'm happy, to give back few minutes of your time, and thank you so much. Honored and humbled to be here and joining you all. Thank you. Okay. There's one question. Do you have any tips on convincing the leadership of new ideas? Yes. Awareness is the starting point, and we need to start bringing, you know, very early in the game, you know, bring every teams and every focus. Everyone should be brought in together, and we shouldn't leave any any as as I said earlier, you know, we need to have a diverse group of people. We need to bring everyone together in the deciding board in the decision making, process. That's how, leadership should take a call early on.

What is the most challenging AI project you have worked on? Alright. So I am working on a AI project where I'm trying to get the adoption rate higher. And these these are the issues and the challenges that I face that when the decision was made that we are going to implement this particular tool, lot of people were not involved in the beginning, and that completely changes the game. And that is where we need to make sure that, you know, there is a representation. Every seat in the table is diverse and not the same group of people, you know, striped t shirts and striped shirts and, you know, collared men white men. I mean, I'm not sorry. I might sound like being a feminist. I'm not. But I just want a fair game and a fair chance and opportunity to everyone, and that's where I'm coming from. So my biggest challenge is now AI adoption.

We have bought the tool, but people are not ready to, you know, kind of take a they are not ready to buy in. They are not ready to, kind of because it doesn't work for them. It doesn't fit their bill. And it is only for the leadership team. They wanna see a dashboard view, but people are doing their BAU activities. And, for them, it is an extra chore of work. So yeah. Varunam, I mean, some tools that can help help you drive the people side of change. Let's connect. Absolutely. Yes. We should. I am also doing this lot of research. I'm doing my PhD in AI and leadership and, trying to, you know, kind of bridge the gap and, bring everyone on the same journey. So thank you so much for asking these questions, Varna. Really appreciate, and we should definitely connect. Any other questions that I can take? Thank you. Thank you so much.

I'll stop my screen, and thank you so much.