Can we really trust AI?

AI is scaling faster than trust can be built. That mismatch may prove more important than many of the technical questions dominating current discussions about artificial intelligence. Organisations can deploy new systems in weeks. Models can change between one release and the next. Yet confidence in the institutions building, deploying and governing those systems develops much more slowly.

In Episode 17 of The AI Values Podcast, Edosa Odaro is joined by guest co-host Tolu Adebekun, Data and AI Leader, to examine why trust in AI cannot be separated from trust in employers, technology companies, governments and other institutions. Their discussion moves beyond whether individual models are accurate or useful. The deeper issue is whether people believe the organisations behind those systems are acting in ways that deserve confidence.

For leaders, that changes the nature of the problem. Trust cannot be installed through a communications campaign after deployment. It has to be created through decisions, incentives, accountability and evidence over time.

AI trust is really institutional trust

AI is often discussed as though trust is a property of the technology itself. A system is trusted because it is accurate, secure, or reliable.

The reality is more complicated.

People encounter AI through institutions. Employers decide where it will be deployed. Technology companies decide how models behave. Governments establish rules and oversight. Banks, schools, hospitals and other organisations increasingly use automated systems to influence decisions that affect people's lives.

Adebekun argues that this makes AI trust an extension of a much older problem.

"AI is really an extension of how we trust wider society."

For organisations, this distinction matters because dissatisfaction with AI may have little to do with model performance. Employees may distrust an AI deployment because they believe it is being introduced primarily to reduce headcount. Customers may question automated decisions because they do not trust how a company uses their data. Citizens may doubt regulatory assurances because confidence in public institutions has already weakened.

The technology inherits that history.

Odaro therefore argues that the trust deficit often sits around the system rather than inside it.

"It's not really the technology that's struggling with the trust deficit. It's actually the institutions and the kind of conditions around the tech."

This creates a much broader governance responsibility. Improving an algorithm does not automatically repair institutional trust. Leaders also need to examine the incentives, processes and behaviours surrounding its use.

Adoption is moving faster than consent

One of the difficulties with AI is that many people do not experience adoption as a voluntary decision.

They may choose whether to use a chatbot personally, but they have far less control over whether their employer, bank, school, healthcare provider or government uses AI behind the scenes.

Adebekun captures that tension directly:

"Do we really have a choice but to trust?"

The question is uncomfortable because AI increasingly enters people's lives through systems they did not choose. Organisations face competitive pressure to adopt it. Employees are told that failing to use AI may leave them behind. Customers may be evaluated by systems whose operation they barely understand.

This means trust is being demanded before it has necessarily been earned.

Traditional approaches to organisational change often respond to this problem through training, awareness and communication. Those interventions can help people understand a system, but they cannot by themselves establish confidence in the decisions that system produces.

As Odaro observes, organisations can fall into the trap of treating distrust as a knowledge deficit.

"It almost feels like, I will tell you how to trust this thing."

That approach risks missing what people are actually asking. They want to know whether the decision is fair, whether the outcome is reliable, who is accountable and what happens when the system gets something wrong.

Trust therefore depends less on persuading people to feel confident and more on giving them reasons to be confident.

Technology scales faster than trust

Trust develops through experience. People observe whether an organisation keeps its commitments, responds well when something goes wrong and acts consistently with what it says it values. That takes time.

AI operates on a very different timeline.

New capabilities can move from experimentation to widespread deployment rapidly. An organisation can introduce a system across thousands of employees before confidence in that system has had time to mature.

Odaro identifies this as one of the central governance problems:

"The time it takes for trust to be built is completely different from the amount of time it takes for technology to scale."

The consequence is a widening gap between capability and confidence.

Organisations can be technically ready to deploy AI long before their employees, customers or other stakeholders trust what that deployment means for them.

This is not necessarily an argument for slowing every implementation. It is an argument for recognising that trust requires its own infrastructure.

That includes clear accountability, meaningful transparency, evidence that systems perform as expected, credible mechanisms for challenge and visible responses when problems occur.

Without those conditions, AI adoption may appear successful initially while confidence erodes underneath it.

Trust is built through outcomes, not statements

Corporate values statements and AI principles can signal intent, but trust is ultimately shaped by behaviour.

Adebekun points to a broader decline in confidence in organisations. Employees have seen profitable companies make large-scale redundancies. Consumers have watched technology platforms optimise for engagement while concerns about addiction and distraction grew. Citizens have witnessed failures in public administration and questioned whether institutions consistently act in their interests.

Against that background, asking people simply to trust AI is unrealistic.

Adebekun puts the issue starkly:

"There is a lot of trust issues because I feel like companies, institutions, they've proven over the past 20, 30 years that they're not just doing things in our best interest. Their best interest is their bottom line."

Whether every institution fits that description is less important than the perception it reveals.

People assess AI deployments through what they already believe about the organisation behind them.

This means leadership behaviour before an AI implementation can influence whether that implementation is trusted. An organisation that has treated employees transparently during previous changes may have more credibility when introducing automation. One that has repeatedly prioritised short-term savings at employees' expense may encounter scepticism regardless of how well the technology performs.

Trust is cumulative.

AI governance therefore begins before the AI system arrives.

The systems carry the priorities of their makers

Discussions about advanced AI frequently drift towards the idea that systems are independently choosing how to behave.

Odaro draws an important distinction.

Current systems operate within layers of instructions, controls and incentives created by the organisations building and deploying them. Users may provide prompts and organisations may create custom configurations, but those sit above deeper system-level decisions established by the platform provider.

That makes the incentives of the organisations creating AI systems a governance concern.

"These systems don't yet have a mind of their own, they adopt the mind of their maker."

The phrase captures a difficult reality for organisations purchasing AI services.

A model may be technically external to the organisation, but the choices embedded within it can still shape internal outcomes. Providers determine acceptable behaviour, guardrails, commercial structures and many aspects of how systems respond.

For boards, this means vendor assessment cannot be limited to capability, security and price. Leaders also need to understand the incentives and governance arrangements of the organisations supplying increasingly consequential technology.

The more authority an organisation delegates to an external AI system, the more important those questions become.

Autonomy changes the risk

AI capability and AI autonomy are not the same thing. A system can be extraordinarily capable while remaining limited to answering questions. Risk changes when that system is permitted to act.

Odaro separates these two concepts in the discussion. One model provides an answer. Another is given authority to take action based on the answer.

That shift matters because organisations are no longer simply receiving information from AI. They are increasingly exploring agents capable of executing tasks, making changes and interacting with other systems.

The governance question therefore becomes one of agency.

How much control is the organisation willing to surrender? Where must human approval remain? Which decisions are too consequential to delegate?

Adebekun pushes the concern further:

"Do you think we can control it though? Can we actually control it?"

Even before hypothetical AGI, organisations already need to decide what systems are permitted to do and what happens when their behaviour conflicts with user expectations.

Trust in autonomous systems therefore depends heavily on the conditions under which authority is granted.

Regulation cannot carry the whole burden

AI regulation matters, but Episode 17 raises doubts about whether regulation alone can generate sufficient trust.

The technology is complicated, fast-moving and already embedded in systems that many regulators are still learning to understand. The same challenge has appeared with earlier digital technologies, where institutional understanding often lagged behind the systems being governed.

Odaro argues that organisations should therefore avoid relying on a single mechanism of control.

"I think what really is needed is a collection of instruments in some ways of control."

Those instruments can include regulation, technical standards, internal governance, testing, audit, transparency, human oversight and market accountability.

The underlying principle is that no single safeguard should carry the entire burden.

A good analogy for this is the crash of the Boeing 737 Max. This shows what can happen when automation, organisational incentives, training, governance and oversight fail to work together. This demonstrates that technical capability can become dangerous when institutional conditions around it are inadequate.

That lesson is directly applicable to AI. A technically capable system cannot compensate for weak governance surrounding its deployment.

Trust failure can remain hidden

One of the most significant risks for organisations is that trust can deteriorate without producing an obvious crisis.

An AI system may continue delivering operational savings. Usage figures may remain strong. Leaders may believe implementation is succeeding.

Meanwhile, employees develop workarounds. Customers question outcomes. Teams increasingly rely on unauthorised tools. Regulators begin asking harder questions. Confidence declines gradually.

Odaro describes these as signals of an emerging trust deficit.

This matters because trust problems can eventually become value problems.

An organisation can spend heavily on AI and still lose much of the expected benefit if employees resist adoption, customers avoid automated interactions or stakeholders repeatedly challenge decisions.

The business case therefore needs to account for trust as part of sustainability.

Deployment is not the finish line.

As Odaro explains:

"It might seem to work on day one. But what happens on day two and day three"

That is the question leaders need to carry into AI governance.

Short-term technical success does not guarantee that an implementation remains acceptable, trusted or valuable over time.

Listen to Episode 17

Going deeper: trust is becoming AI infrastructure

Organisations often think of AI infrastructure in technical terms: data, models, platforms, security and computing capacity.

Episode 17 points towards another form of infrastructure that is easier to overlook.

Trust.

Without it, technically successful systems can become institutionally fragile. Employees may refuse to rely on outputs. Customers may demand human alternatives. Regulators may increase scrutiny. Leaders may hesitate to act on recommendations. Teams may create workarounds that weaken governance further.

Building trust therefore requires more than explaining AI. Organisations need to demonstrate that accountability exists, that incentives are aligned, that consequential decisions can be challenged and that someone remains responsible when something goes wrong.

They also need to recognise that trust is relational. People are not merely deciding whether they trust an algorithm. They are deciding whether they trust the organisation deploying it, the provider that created it and the governance system responsible for constraining it.

This becomes especially important as AI systems gain greater autonomy. Delegating more authority to machines does not reduce institutional responsibility. It increases the importance of being clear about where authority came from, what limits apply and who remains accountable.

Organisations racing to match competitors on AI capability should therefore ask whether their trust systems are advancing at the same pace.

If they are not, the technical deployment may still succeed.

The institution surrounding it may not.