What AI can learn about trust from banking

At the UN AI for Good Summit last week, Amazon CTO Werner Vogels made a simple observation:

"Without trust, there is no AI."

His argument wasn't that trust is disappearing. It was that our traditional mechanisms for creating trust are failing at scale.

To explain why, he traced the evolution of trust itself.

In the eighteenth century marketplace, trust was largely physical. A Dutch trader knew who they were dealing with because they saw them every day. Reputation was local, visible and personal. Presence itself was evidence. Trust was grounded in human relationships.

As trade expanded beyond local markets, that model no longer worked. You couldn't personally know every trading partner across borders and oceans. So trust migrated to institutions. Exchanges, banks, regulators and clearing houses became trusted intermediaries. The institution provided assurance where personal familiarity could not.

The internet changed that model again.

Trust became distributed. Today we routinely trust people we have never met. We get into taxis with strangers because an app tells us they are highly rated. We stay in someone else's home because hundreds of reviews suggest it is safe. We buy products from unknown sellers because thousands of previous buyers appear satisfied.

We have become remarkably comfortable placing trust in systems rather than people.

Yet, according to Vogels, AI has disrupted each of these trust models simultaneously.

The old signals of personal trust can now be synthesised. A video call, a voice message, even the reassuring phone call from your boss can be convincingly deepfaked.

Institutional trust can be forged too. Authentic-looking documents, credentials and communications can be generated at negligible cost.

And distributed trust is increasingly vulnerable to manipulation. Millions of fake reviews, synthetic content and coordinated influence campaigns undermine the very signals that digital platforms rely upon to establish credibility.

If trust was once expensive to fake, AI is making it increasingly cheap.

Vogels' answer was what he described as verifiable trust.

Rather than trusting people, institutions or crowds by default, we need mechanisms that verify what goes into AI systems, what comes out of them and, increasingly, what autonomous agents actually do.

That means governance. It means automated compliance. It means encoding rules and policies into machine-checkable logic. For higher-risk decisions, it may mean requiring multiple agents to agree before actions are taken. Most importantly, it means applying controls proportionately, according to risk.

Listening to this, I was struck by how familiar the debate sounded.

I spent fifteen years working in banking, an industry that has experienced its own crisis of trust.

For decades, banks acted as trusted intermediaries for increasingly complex and global financial activity. Cross-border trade, lending, settlement and investment all relied upon institutions that customers, businesses and governments believed would operate safely and predictably.

Then came the financial crisis.

What followed was not simply a market correction. It was a collapse in confidence in the institutions themselves.

For some, the answer was disintermediation. Technologies such as blockchain and later cryptocurrencies emerged from the belief that trust should not depend on trusted intermediaries at all. The premise was straightforward: if institutions can fail, perhaps we should remove the need to trust them in the first place.

In many ways, that mirrors today's conversation around AI.

When trust in existing structures weakens, the instinct is often to decentralise.

But banking's story suggests the answer is more nuanced.

Over time, the banking sector regained much of the trust it had lost. Not by abandoning institutions, but largely by reforming them. Banks simplified business models, reduced risk exposure, strengthened governance and accepted greater scrutiny. Regulators expanded their oversight. Capital requirements increased. Risk management became far more central to decision-making.

Trust was rebuilt through stability rather than disruption.

Even physical presence retained value. At a time when many consumers moved to digital services, some institutions maintained trust precisely because they retained a visible presence on the high street. For certain demographics, the branch remained an important signal of accountability and permanence.

The lesson for AI may be that the evolution of trust does not require the abandonment of institutions.

It may require better ones.

Perhaps the answer to restoring trust in AI is not solely decentralisation, open models or community moderation. Perhaps it is stronger institutions that understand AI deeply, can apply risk-based oversight intelligently and can intervene where harms are greatest.

Vogels' risk framework feels particularly relevant here. Not every AI use case requires the same level of control. A tool helping someone draft meeting notes carries very different risks from a system influencing healthcare decisions, legal outcomes or financial advice.

The future of AI governance is unlikely to be one-size-fits-all.

Just as banking eventually learned to focus its strongest controls on its highest-risk activities, AI may need to do the same.

The question is not whether we need trust. The question is where trust should reside.

I agree with Vogels, that for centuries, trust has moved from people, to institutions, to distributed networks. AI may be forcing us into a fourth era, one where trust is not assumed, but continuously verified.

And if banking teaches us anything, it is that when trust is broken, rebuilding it rarely comes from removing all intermediaries. More often, it comes from creating institutions worthy of trust in the first place.