AI Has a UX Problem: What ISO 9241-810 Means for Trustworthy AI

AI is often discussed as a technical challenge. In practice, it is just as much a UX challenge.

If people cannot understand what an AI system is doing, trust it appropriately, or intervene when something looks wrong, the system will struggle to deliver value, no matter how advanced the technology behind it may be. That is why UX needs to sit at the centre of AI design, not at the end of the process.

This matters because many AI failures are not caused by weak models. They happen when the experience around the model is poorly designed. Users are left guessing what the system is doing, why it has produced a particular output, or whether they can rely on it at all.

Why AI Creates New UX Problems

Traditional UX design assumes that systems behave in consistent, predictable ways. If a user does the same thing twice, they usually expect the same result twice. That makes it easier to form a mental model, build trust and spot errors.

AI systems do not always behave like that.

A generative AI tool may produce different answers to the same prompt. A recommendation engine may change its behaviour over time. An autonomous system may take actions on behalf of a user without a clear step-by-step instruction. These are not just technical characteristics. They are interaction problems.

Three issues come up repeatedly:

• Opacity. Users can see the output, but not how the system reached it.
• Unpredictability. The system may respond differently to the same input, which makes it harder to build reliable expectations.
• Automation bias. People are more likely to accept a confident-looking output without questioning it, even when they should.

A loan decision system is a good example. If someone is told they have been declined, but no one can explain why, both the applicant and the reviewer are limited in their ability to challenge the outcome. That is not just a technical shortcoming. It is a design failure that affects fairness, trust and accountability.

What ISO 9241-810 Adds

This is where ISO 9241-810 becomes useful.

The standard gives organisations a framework for designing human-centred interactions with intelligent, autonomous and robotic systems. It does not replace human-centred design principles. Instead, it extends them to cover systems that adapt, act with a degree of independence, and behave in ways users may not fully anticipate.

That matters because many AI products are now expected to do more than simply display information. They may recommend, draft, classify, prioritise or even act on a user’s behalf. The design challenge is no longer only whether the system works. It is whether people can understand, oversee and use it safely.

The key point is simple: if a system has more autonomy, the interface needs more care.

Designing For Trust

Trust in AI is not created by making the interface look confident. It is created by helping users understand when to rely on the system, when to pause, and when to challenge it.

" Trust is not created by confidence. It is created by clarity."

That means design teams need to think carefully about several things:

• How clearly the system signals what it is doing.
• Whether users can see when the AI is advising versus acting.
• How uncertainty is communicated.
• What happens when the AI is wrong.
• How easily users can review, override or stop an action.

For example, if an AI tool can draft, schedule and send emails, users need to know exactly when the system moves from suggestion to action. If that transition is hidden, people can lose control without realising it. That is not a minor usability issue. It is a trust issue.

The same applies in other contexts. In healthcare, finance, hiring or public services, AI outputs often shape decisions with real consequences. Users do not just need speed. They need clarity, control and the ability to make an informed judgement.

Why Human Oversight Matters

A common mistake is to assume that human oversight exists simply because a human is somewhere in the process.

That is not enough.

Meaningful oversight has to be designed into the workflow. People need a real chance to check, question and intervene at the point where it matters, not after the decision has already had an effect.

This is especially important where AI is agentic, meaning it can take action on a user’s behalf. If a system can book, submit, send or trigger something automatically, the risk is not just that it may be wrong. The risk is that the user may not notice the problem until it is too late.

That is why human-in-the-loop should not be treated as a box-ticking exercise. It should be treated as a design requirement.

When that oversight is weak, organisations tend to see the same pattern:

• users over-rely on the system,
• users work around the system,
• or users stop trusting the system altogether.

None of those outcomes are good for adoption, safety or performance.

Accessibility Cannot Be An Add-On

AI also creates new accessibility challenges.

Conversational interfaces can be useful, but they can also introduce barriers for people with low literacy, cognitive impairments, screen reader users, and anyone who does not fit the assumptions built into a natural language interface. If an AI experience assumes everyone will interact in the same way, it will exclude people.

Accessibility in AI is not just about compliance. It is about whether the system can actually be used by the widest possible range of people. That means testing beyond the most digitally confident users and thinking carefully about the full experience, not just the output.

In practice, that could mean:

• making uncertainty visible in a clear and usable way,
• supporting keyboard and screen reader access properly,
• avoiding overly complex language,
• and designing recovery paths when the AI gets something wrong.

If accessibility is left until the end, it becomes a retrofit. In AI, that usually means a worse experience for everyone.

What Teams Should Do Differently

If organisations want AI that people can use with confidence, they need to involve UX much earlier.

The most important design decisions are often made before the interface exists. That includes the level of autonomy the system has, how much transparency is needed, what users should be able to control, and how errors will be handled. If those questions are left to technical teams alone, the result is often a product that works in theory but fails in practice.

A better approach is to build AI products around a few clear principles:

• Design for the right level of trust. Do not aim for frictionless confidence. Aim for informed confidence.
• Make autonomy visible. Users should know when the system is suggesting, assisting or acting.
• Support user judgement. The interface should help people question the output, not just accept it.
• Build in recovery. Users need clear ways to correct, reverse or challenge an AI action.
• Test with real users. Include people who are less digitally confident, not just expert users.
• Treat accessibility as essential. If some users cannot use it, the design is not finished.

These are not abstract principles. They are practical safeguards that improve usability, trust and adoption.

Why This Matters Now

The organisations that succeed with AI will not necessarily be the ones with the most capable models. They will be the ones whose systems people can actually understand, trust and use well.

That is where UX becomes decisive. It turns AI from something impressive in a demo into something reliable in the real world.

ISO 9241-810 gives teams a useful framework for that work, but the deeper message is broader: AI design cannot stop at technical performance. It has to account for human judgement, human error, human trust and human limits.

If AI is going to be used in serious contexts, the question is no longer simply whether it can do the task. It is whether it can do it in a way people can safely and confidently live with.

Key Takeaways

• AI is a UX challenge as well as a technical one.
• Users need clarity about what the system is doing, why it is doing it and when they can intervene.
• Trust should be designed, not assumed.
• Human oversight has to be meaningful, not symbolic.
• Accessibility needs to be built into AI experiences from the start.

Final Thought

AI adoption will be shaped not just by model quality, but by whether people can make sense of the experience around the model. That is the work of UX, and it is now central to trustworthy AI.