Why we still don't trust AI

Roughly 51 per cent of UK adults now use AI tools for personal advice. Not to automate work tasks, not to speed up research, but to get guidance on their health, their relationships, their decisions. They do this not because they prefer a chatbot to a human being, but because the human being is not available. The GP appointment takes weeks. The therapist has a waiting list. The chatbot answers in seconds.

That figure was cited from The Independent during the episode. Before treating it as background noise, consider what it says about the world AI is being built into. At scale, AI is stepping into a gap that healthcare systems, public services, and employers have collectively left open. And it is doing so without anyone formally deciding that was the right role for it to play.

In Episode 4 of The AI Values Podcast, Edosa Odaro and Lindley Gooden speak with Dr Naeema Pasha, Visiting Fellow at Henley Business School, Founder of The World of Work Institute, Behavioural Scientist, and author of Futureproof Your Career. The conversation centres on the question behind that number: who is AI actually building the future for, and who gets left outside it when nobody thinks to ask?

The word that changes so much

Dr. Pasha opens the episode with a word she returns to throughout: fairness. Not ethics. Not compliance. Not responsibility. Fairness. She is precise about what she means. Fairness for people using AI. Fairness to people in society. And fairness for people whose lives are shaped by AI decisions they may not even know are being made about them.

The distinction matters more than it first appears. Ethics is often treated as a framework to satisfy. Fairness is something people feel in their daily lives. And once people stop feeling it, something more damaging than a governance failure occurs.

"Once trust starts being eroded, people feel disempowered. Once people feel disempowered, they can either lose motivation, disconnect from society."

Pasha connects this to the roughly 10 million people in the UK who are currently not in work or education. Disengagement at that scale is not a coincidence. It is linked to broken trust in institutions, in systems, and increasingly in the technologies those institutions deploy. When AI removes the human contact point without replacing the sense of being seen and heard, it does not address the problem. It deepens it.

AI as an architect of boundaries

One of the sharpest ideas in Episode 4 is how Pasha describes what AI is doing to human connection. Not by deliberate design. By structural effect. She calls AI an architect of boundaries.

When an algorithm surfaces content based on what you have already engaged with, it is not being neutral. It is reinforcing the walls between you and everything unfamiliar. Your recommendations pull you deeper into existing tastes. Your workplace tools are trained on data produced by organisations with particular cultures. Predominantly Californian. Predominantly from narrow demographics. And they encode those cultures into the norms they reward.

"What it does in the workplace is that because of pattern amplification, the kind of things that are recommended, it's creating boundaries, it's creating a particular way of behaving, the kind of language you're looking for."

In a meeting with people in the room, you see differences. You see someone push back. You see the moment when someone's face shows they disagree. The friction is visible, and it is productive. When human contact is increasingly mediated by AI tools, that friction disappears. And with it, so does the kind of thinking that difference creates.

Pasha calls the result groupthink. Not the groupthink of a difficult boardroom, but something quieter and more systemic. The slow convergence of perspective that happens when the tools shaping how you work all point in the same direction.

The fork in the road between access and quality

One of the more honest moments in the episode is the conversation about AI and mental health support. Pasha acknowledges something that most AI governance discussions tend to skip: for a significant portion of the population, an AI chatbot is not the inferior option. It is the only option.

Young people who do not feel they will be understood by a therapist. People from communities who have learned to distrust institutions that have historically judged them. People who cannot get an appointment at all. For them, a non-judgemental, immediately available, always-patient chatbot represents a genuine step forward from having nothing.

"Something like a non-judgmental bot giving support might be good... the research shows the output of coaching bots is good and comparable to human but not as comparable to very high quality human coaching."

That is the fork in the road. High quality is not the same as accessible. And accessible, even if imperfect, has real value for people who would otherwise go without. The implication for leaders is not that AI in wellbeing or healthcare is the wrong call. It is that deploying it without asking who you are serving, and what they are getting instead of what they actually need, is not a strategy. It is an abdication.

The business case hiding in plain sight

Pasha led a major piece of research at Henley Business School called the Equity Effect, examining race equity in the workplace. The central finding was striking: organisations that actively tackled race inequity had 58 per cent higher revenue than those that did not.

The research then asked the same senior leaders a follow-up question. Do you recognise the importance of race equity in your organisation? Yes, overwhelmingly. Will you act on it? It came near the bottom of the priority list.

"I've got the business argument. I could have gone in for race equity with a social argument or a moral argument or even a legal argument. But I went with the business argument. And I didn't expect to find those results."

The gap between recognising that something matters and acting on it is a pattern this podcast has returned to in every episode. In Episode 1, it was the question nobody thought to ask before the investment was made. In Episode 2, it was the AI project that worked and was cancelled because the results were inconvenient. Here it is an equity case backed by a 58 per cent revenue premium that still ends up at the bottom of the to-do list.

The lesson is not that leaders are indifferent. It is that the structures they operate within make inaction the path of least resistance. And until those structures change, data alone does not move anything.

Mind your language

The episode closes with a call to action that sounds deceptively simple. Pasha's closing advice to AI developers, business leaders, and everyone using these tools is to pay attention to language. Not policy language. The language of everyday conversation.

Language is where bias enters systems. It is where the assumptions of the people who built the tools get encoded at scale. It is where division gets normalised or where connection becomes possible. The AI systems being trained on the data we generate will reflect the language we use. If that language draws lines between people, the systems will draw them too.

"Consider the language that you use. Watch the language that makes division. And then I think the AI will follow that."

For organisations, this is practical rather than philosophical. It means asking what language your internal tools are amplifying. What kind of communication your AI-assisted processes reward and what they filter out. Whether the norms being embedded in your systems are the norms you would consciously choose if you were asked directly.

Language is not soft. It is architecture. And right now, in most organisations, nobody is reading the blueprints.

Listen to Episode 4

‍

Going deeper: why fairness is a governance question, not just a values statement

Pasha's work is useful for leaders precisely because she does not present fairness as a moral aspiration. She presents it as an organisational design question with measurable consequences. The 58 per cent revenue premium is not an argument for being good. It is an argument for being serious about what actually drives performance, and for asking why that evidence still ends up near the bottom of the priority list.

The broader implication of her argument is that the governance questions organisations most want to avoid are often the ones with the most direct connection to outcomes. Who is this system serving? What does it take from the people who carry its costs? Does the language it uses reflect the full range of people it affects? These are not philosophical questions. They are diagnostic ones. They are the questions that would surface the problem before it becomes a headline.

The challenge is that the systems of incentive, reporting, and accountability inside most organisations do not reward asking them. The CEO who publicly commits to fairness still operates on quarterly targets, investor expectations, and a board whose definition of risk does not yet factor in the cost of 10 million disengaged people or the long-term damage of systems that embed existing inequalities at scale.

What Pasha is pointing at is a structural mismatch: the timeline of social consequences is longer than the timeline of business reporting. AI moves fast. Trust, once broken at the scale AI is now operating, takes years to rebuild. The organisations that handle this well are the ones that build the question "who does this serve?" into the design process rather than the post-launch retrospective. That is what fairness as governance looks like in practice. Not a policy document. A question that gets asked before the decision is made.