Culturally Intelligent AI: Rhea Sharma on Building AI That Understands Culture and Context

Understanding Culturally Intelligent AI

Q1. Let’s begin with the fundamental idea. What does Culturally Intelligent AI mean to you. And how is it different from simply making AI more diverse or inclusive?

RS: To me, culturally intelligent AI goes beyond having diverse datasets or diverse teams. It means designing systems that recognize people interpret information, communicate, make decisions and experience technology differently depending on their cultural, social and linguistic context. Diversity is about who is represented. Cultural intelligence is about whether the system actually understands the context behind that representation. I’ve lived through experiences that no demographic checkbox could fully capture, and that has shaped how I think about this distinction.

Q2. AI systems learn from data, but human behaviour is deeply influenced by culture, language, geography, social norms and lived experience. Why has cultural context been relatively difficult to incorporate into AI?

RS: Cultural context isn’t a simple variable you can add to a dataset. It’s shaped by language, geography, family structure, social norms and lived experience, and those factors change how the same words or behaviors should be interpreted. During my worst years, my silence would have looked like nothing on a chart, but it was everything. AI systems need better ways of recognizing context instead of assuming that whatever is easiest to measure is what matters most.

Insufficient Data or Deeper Challenges

Q3. Is the problem primarily one of insufficient data. Or are there deeper challenges in how AI systems are designed and evaluated?

RS: Data is part of the problem, especially when certain communities are underrepresented or their experiences are poorly captured. But the deeper issue is who decides what data matters, what outcomes are important, and how success gets measured. I’ve rarely seen people like me in the rooms where those decisions are made, and that matters because bias can enter long before a model is trained — through problem definition, data collection, labelling, product design, evaluation criteria and deployment decisions.

Q4. How can AI developers distinguish between genuinely understanding cultural context and simply learning cultural stereotypes?

RS: A stereotype assumes who you are before you’ve said anything. Real understanding leaves room for context. From a product standpoint, that means systems shouldn’t infer someone’s needs just from their demographic identity. Developers need to test whether a model can tell the difference between a cultural pattern and an individual preference, and whether it can ask for clarification instead of making an assumption.

In the 1:1 conversations I’ve had with over 800 women, one pattern keeps showing up: career gaps. Hiring algorithms tend to treat a gap in someone’s work history as a flat negative signal, no matter what caused it. For a lot of the women I’ve spoken with, that gap wasn’t a lack of ambition or capability — it was caregiving, family expectations, or a cultural obligation that isn’t optional where they come from. A system that flags the gap as a red flag without asking why is stereotyping — it’s decided the meaning of that gap before knowing anything about the person. A system that’s genuinely culturally intelligent would recognize the pattern and still leave room to ask what actually happened.

AI System Biased?

Q5. Could an AI system become more biased precisely because it is attempting to account for cultural differences? How should that risk be managed?

RS: Yes, absolutely. Trying to account for cultural differences can create new stereotypes if organizations treat a culture as one fixed set of behaviors. The safer approach is to involve the communities being affected, test assumptions with real users, evaluate performance across relevant groups, and build in a way for people to correct the system when it gets their context wrong. The goal isn’t to make assumptions more sophisticated — it’s to make the system better at recognizing when it doesn’t know enough.

AI, Bias and Representation

Q6. Much of the conversation around AI bias focuses on biased datasets. What other sources of bias should organisations be concerned about?

RS: Bias can come from who’s allowed to build the system, not just the data it’s trained on. It enters through product requirements, team composition, incentives, labelling decisions, evaluation criteria, procurement choices, and even which problems an organization decides are worth solving in the first place. A technically diverse dataset can’t fully make up for a development process that keeps excluding certain perspectives. I built my company after being told I had too much baggage to be taken seriously, and that experience shaped how strongly I believe the people closest to a problem need a real voice in designing its solution.

Q7. What happens when communities are underrepresented in the data used to build AI systems that ultimately make decisions about them?

RS: They get decided about instead of asked. When a community is underrepresented, a system can perform well on aggregate metrics while still failing that community in specific, serious ways — especially in healthcare, employment, financial services or other high-impact settings. Representation can’t stop at collecting more data. Organizations need to understand whose experiences are missing, test performance across those groups, and monitor what happens after deployment.

Performs Equitably

Q8. How should organisations determine whether an AI system performs equitably across different cultural, linguistic and demographic contexts?

RS: They need to stop relying on one overall accuracy score, because it can hide real differences between groups. Organizations should disaggregate their evaluation results across relevant demographic and linguistic groups, test culturally specific scenarios, and look at where error rates diverge. They should also involve representative users in testing rather than relying entirely on internal teams. A model can look excellent overall and still consistently fail the people who are least represented in its data.

Evaluation and Benchmarking Frameworks

Q9. Do today’s AI evaluation and benchmarking frameworks adequately measure cultural intelligence? If not, what is missing?

RS: Not really — most benchmarks measure what’s convenient rather than what’s meaningful. They’re built around aggregate performance, accuracy, safety or general task completion, which matters, but it can miss whether a system behaves appropriately across different cultural and linguistic contexts. We need more culturally grounded evaluation suites, disaggregated performance metrics, and participatory evaluation where the communities being assessed help define what good performance actually looks like.

I see this most clearly in clinical trials, which is where I spend most of my working life. Most trial recruitment and eligibility-screening tools get evaluated on overall accuracy across the study population — did it correctly flag eligible patients, yes or no. What almost never gets measured is whether that accuracy holds up within specific groups: by race, primary language, or the socioeconomic makeup of a given site’s community. A screening algorithm can look completely reliable in aggregate and still be quietly filtering out eligible patients from the exact populations a trial is supposed to reach. What’s missing isn’t a better overall benchmark — it’s a standard practice of testing performance group by group, and involving the communities being recruited in deciding what data even gets collected at screening.

Explicit Design and Governance Requirement

Q10. Should cultural intelligence become an explicit design and governance requirement for AI systems, much like privacy, security and explainability?

RS: Yes. We take privacy seriously because we learned the cost of getting it wrong, and we should apply the same seriousness to how AI systems treat people. Cultural intelligence should be built into the development and governance process from the start, not added as an optional layer at the end — through representation requirements, subgroup testing, documented limitations, community input and post-deployment monitoring. The exact requirements should vary by risk and use case, but the principle should stay the same: if a system can materially affect people, context can’t be an afterthought.

Designing Around Real Human Experiences

Designing Around Real Human Experiences

Q11. You emphasise designing technology around users’ lived realities rather than expecting users to adapt to technology. What does that principle look like in practice?

RS: It starts with accepting that the person using a product may not experience the world the way the product team imagined. In practice, that means doing research before designing the solution, observing how people actually use the technology, testing with different user groups, and designing for real constraints — language, accessibility, time, trust and social context. I know what it feels like to be rebuilding your life at midnight, exhausted and still trying, and that’s the question I keep coming back to: are we designing for the user we imagined, or the person who will actually use this?

Q12. How can product teams bring lived experience into AI development without reducing individuals or communities to demographic categories?

RS: By sitting with someone’s actual story before sorting them into a category. Lived experience should inform research and product design without becoming a shortcut for assumptions. Product teams can use qualitative research, interviews, participatory design and representative testing alongside quantitative data. I’m much more than any box I’ve ever been put in, and AI should be designed with that same humility — demographic categories can help identify patterns, but they should never replace understanding the individual.

Designing and Testing AI Systems

Q13. What role should communities themselves play in designing and testing AI systems intended for them?

RS: They should help write it, not just review it afterward. Community participation should start before the product is built and continue through testing and deployment — through advisory groups, participatory research, user testing, feedback mechanisms and clear processes for acting on what those communities report. I only trust rooms where I actually have a seat, and if a community is only brought in after the important decisions are made, that’s consultation, not real participation.

Q14. How can organisations incorporate cultural intelligence into product design without making the development process excessively complicated or expensive?

RS: Talk to real people early. Upfront user research and targeted testing is far cheaper than discovering a major cultural failure after launch. Organizations don’t need a completely separate process — cultural considerations can be built into existing product research, QA, model evaluation and governance workflows. The key is making it part of the process from the start rather than a correction at the end.

What I’ve seen at Diverse AI is that women tend to create this change themselves. When one woman finds another woman who’s been through the same pain or difficulty, support starts spreading person to person — it becomes a chain, not something we have to engineer or pay for. That’s how most of our product feedback and community input actually reaches us. So cost was never really the barrier. The real work, and the thing that takes time, is building the trust that makes a woman willing to be that first link in the chain.

Technology Development Process

Q15. What are some warning signs that an organisation is talking about inclusive AI but has not actually embedded inclusion into its technology development process?

RS: Watch what survives when the budget gets cut. If inclusion disappears when timelines tighten, if diverse testing gets postponed until launch, if community feedback is collected but never acted on, or if no one is accountable for subgroup performance, inclusion is probably still a talking point rather than a practice. The real test is whether it shows up in the roadmap, budget, evaluation criteria and post-launch monitoring — not just in the messaging.

AI, Healthcare and Equity

Q16. Your background spans healthcare operations, clinical research and AI/ML. What does healthcare reveal about the consequences of AI failing to understand cultural and social context?

RS: Healthcare makes the consequences especially visible, because a system can be technically accurate and still fail the person in front of it. I’ve sat in hospital rooms as a patient, not just as a director, and I know what it’s like to be treated through the lens of a chart while the person behind that chart is carrying a much bigger story. In healthcare AI, cultural and social context shapes how patients communicate symptoms, access care, follow recommendations and engage with providers — so clinical performance can’t be considered separately from the patient’s experience and circumstances.

Q17. Healthcare involves highly personal decisions and significant differences in language, trust, communication and access. Where could culturally intelligent AI make the greatest difference?

RS: One of the biggest opportunities is in who gets recognized, understood and invited into care in the first place. Culturally intelligent AI could help identify communication barriers, support multilingual interactions, personalize patient education, and flag where standard assumptions don’t fit a particular population — but those systems need strong clinical oversight and validation, because personalization without safeguards can also introduce bias. I know what it’s like to feel overlooked by a system that was supposed to help me, and I don’t want technology to automate that experience.

AI in Healthcare

Q18. What are the biggest risks when AI is deployed in healthcare settings without adequately accounting for underrepresented populations?

RS: The biggest risk is that a system can look successful at scale while repeatedly failing a smaller population — in screening, diagnosis, triage, patient communication or access to care. Underrepresentation can lead to weaker model performance for particular groups, and aggregate metrics make that hard to see. The fix isn’t just “more data.” It requires representative validation, subgroup performance testing, clinical oversight, post-deployment monitoring, and clear escalation when the system behaves differently across populations.

Q19. Can culturally intelligent AI improve clinical outcomes. And, also the broader patient experience and relationship between people and healthcare systems?

RS: Yes, because trust is part of the healthcare experience. A patient can receive technically correct information and still disengage if the system doesn’t understand their language, circumstances or concerns. AI can support more personalized communication, clearer explanations and better navigation of care, as long as it’s designed around the patient’s context. I healed when someone finally saw more than my chart, and that’s why I think patient experience needs to be part of responsible AI design, not an afterthought.

Getting AI Inclusion Right

Q20. Where do you believe healthcare organisations are currently getting AI inclusion right, and where is there still significant work to do?

RS: We’re getting better at language access — recognizing the importance of translation, diverse datasets and responsible AI governance. Where there’s still real work to do is making sure those principles translate into actual clinical workflows and measurable outcomes across different populations, not just policy statements.

Honestly, I haven’t seen one that’s fully getting this right yet — and that’s exactly why we’re doing this work. Every organization I’ve come across is missing some piece of what it takes to build genuine trust with the people it’s supposed to serve. That gap is what we’re trying to close.

Underrepresented and Immigrant Communities

Q21. What unique challenges do underrepresented and immigrant communities face when interacting with AI-powered systems?

RS: It’s not just about language. It involves trust, cultural norms, family structure, stigma, unfamiliar institutions and different expectations about how authority works. There are experiences our cultures don’t always give us permission to name — divorce, addiction and depression weren’t conversations that had a natural place at my dinner table. AI systems interacting with these communities need to account for those contextual barriers without turning them into stereotypes, which requires listening, representative testing and giving users a way to clarify or correct the system.

Q22. Language translation alone does not necessarily provide cultural understanding. What additional layers are required?

RS: Translation converts language. Cultural understanding requires paying attention to context, tone, social norms, intent and the circumstances in which communication happens — the same phrase can mean very different things depending on who’s saying it, to whom and why. That means culturally aware systems need evaluation that goes beyond translation accuracy, including contextual scenarios, native-speaker testing and community feedback.

Digital and Institutional Barriers

Q23. Could AI become a tool. Especially, for reducing digital and institutional barriers for immigrant communities. Rather than creating another layer of exclusion?

RS: Yes. AI could help people navigate unfamiliar systems, understand complex information, access services in their preferred language, and get support when they don’t know where to begin. But accessibility can’t just mean putting a chatbot in front of an existing institutional barrier — the system needs to be accurate, transparent about its limitations, and connected to real pathways for human support. That’s the version of AI I want to build: technology that reduces the number of moments someone is left thinking, “I don’t even know who to ask.”

Q24. How should technology companies engage with communities whose experiences have historically been missing from mainstream technology design?

RS: Build the relationship before you need something from us. Companies should invest in sustained user research, community partnerships and participatory testing before a product launches, not approach communities only when they need validation or a testimonial. I’ve been used for a headline before, and I know what that feels like. Representation shouldn’t be performative — people should be able to see where their input actually changed a decision.

The Business and Technology Case

Q25. Some organisations may view culturally intelligent AI primarily as an ethical responsibility. Is there also a compelling business case for building technology that understands cultural context?

RS: Yes. This isn’t a niche market — it’s a market that has historically been underserved. When technology works better for different populations, organizations can reach more customers, reduce friction, improve trust and avoid costly failures. For companies operating across geographies, languages and demographic groups, cultural intelligence becomes a direct product advantage. Every woman my company serves is evidence that there’s real demand for technology built around realities mainstream products have often overlooked.

I’ve personally spoken 1:1 with over 800 women through this work. That number, on its own, tells you the demand is real — women kept showing up, wanting to talk, wanting to be heard, because nothing else was built for what they were actually dealing with.

Q26. Could culturally intelligent AI become a competitive differentiator as enterprises increasingly deploy AI across customer-facing and employee-facing experiences?

RS: Yes, especially as AI gets embedded in more everyday interactions. As basic AI capability becomes commoditized, differentiation will shift from whether a company has AI to how reliably that AI works for different people. Customer-facing systems that misunderstand cultural context can damage trust, and employee-facing systems can reinforce organizational bias if they aren’t evaluated carefully. People remember who saw them, and they remember when technology repeatedly made them feel invisible.

Cultural Representation, Bias Testing, and Contextual Intelligence

Q27. What should technology leaders be asking their AI vendors. Especially, about cultural representation, bias testing and contextual intelligence. That too, before deploying their systems?

Q28. Do you expect cultural intelligence to become a significant capability differentiating AI models and applications over the next five years?

RS: Yes. As AI becomes more capable, the question shifts from “can the system do this” to “can it do this reliably for different kinds of people and contexts.” Models that can operate across languages and cultures without making harmful assumptions will be more useful to global organizations, so cultural intelligence could become a real differentiator in both model development and application design.

The Future of AI

Q29. AI is becoming increasingly multimodal and agentic. Will cultural intelligence become more important? Especially, because AI systems will interact more directly with people and make more autonomous decisions?

RS: Yes. The more autonomy a system has, the greater the consequences of misunderstanding context. A chatbot giving a bad recommendation is one thing — an agent making decisions, taking actions or coordinating services on someone’s behalf raises a much higher bar for contextual understanding. When a system starts deciding for you instead of just talking to you, it needs mechanisms for uncertainty, clarification, human oversight and culturally informed evaluation. My own story took years to tell, and AI should never assume it understands a person after a few data points.

Q30. What would a genuinely culturally intelligent AI system look like five years from now?

RS: It would feel less like a system making assumptions and more like one that knows when to listen, when to ask and when to defer. It would perform consistently across relevant cultural and linguistic contexts, communicate uncertainty, adapt to individual needs without stereotyping, and give people real control over how it uses their information. To me, that would feel like something finally saying, “I see you,” without making you repeatedly prove why you deserve to be understood.

Largely Generic AI Systems

Q31. What technological, organisational or policy changes are necessary to move from today’s largely generic AI systems toward more culturally aware AI?

RS: Change needs to happen at three levels. Technologically, we need better representative datasets, culturally grounded evaluation, subgroup performance testing, and systems that can recognize uncertainty instead of confidently guessing. Organizationally, we need more diverse development teams, real community participation, and accountability for what happens after deployment, not just during model development. And at the policy level, high-impact AI systems need clear requirements around transparency, testing, documentation and monitoring for disparate outcomes. Put people who’ve actually experienced these gaps in the room where decisions get made — lived experience isn’t the only form of evidence, but it’s one technology has historically undervalued.

Q32. Finally, if you could change one assumption that the technology industry currently makes about artificial intelligence and human diversity, what would it be?

RS: I’d change the assumption that people have to fit neatly into categories before technology can understand them. I rebuilt myself through divorce, addiction, an eating disorder and depression, and I built a company through those experiences too — healing and building happened at the same time. People are more complex, resilient and capable than the categories we usually build around them. If AI is going to be part of people’s lives at scale, it needs to reflect that complexity instead of reducing people to whatever’s easiest to measure.

---

Editorial Angle

The next frontier of AI may not simply be larger models or better reasoning. It may be the ability to understand cultural context without turning diversity into another form of algorithmic bias.

Some Points to Ponder:

– Why the future of AI requires cultural intelligence.

– Why generic AI can fail diverse populations

– Cultural context versus cultural stereotyping

– Inclusive AI design

– AI bias beyond datasets

– AI and healthcare equity

– Technology for immigrant and underrepresented communities

– Human-centred AI

– The business case for culturally aware technology

– The future of responsible AI

Finally:

“AI should not simply understand data. It needs to understand the human context in which that data exists.”

---