Who’s in control? AI is creating new inclusivity challenges in public services

Digital exclusion in the UK is more widespread than one may think. Nearly 8 million people still cannot complete foundational digital tasks (e.g. connecting to Wi-Fi), with many people still relying on family members, phone support or workarounds to complete everyday tasks. While AI has the potential to improve citizen experiences, it can also further alienate and even harm those already struggling.

For years, digital inclusion has focused on access, connectivity and basic digital skills. AI changes the conversation. The challenge is no longer simply about ensuring people can access digital services; it’s also about whether they can use AI-enabled services safely, confidently and independently.

Meet Kirsty (not her real name), a recent participant in our user research aimed at delivering a new government service. She’s retired, married and often describes herself as someone who “isn’t particularly confident” with technology.

Her belief has been shaped by small but memorable setbacks. She recently tried to buy two tickets online to see a show but was disappointed to learn upon arrival that she’d actually only purchased one. Experiences like these have made her cautious online and worried she’ll do things wrong. Whenever possible, she’ll opt to speak to someone over the phone or lean on her children to help. Her experience isn’t unusual; she is part of a significant minority feeling disenfranchised by digital technology – and the type of user we must keep front of mind when using AI to improve and deliver new services.

The arrival of AI introduces new complexities around the adoption and use of digital technologies. It’s changing how people work, access information, engage with the NHS and risks widening the divide between those who will benefit from its potential to improve their lives and those who cannot.

There is, however, another side to this story. AI can also be a great enabler.

For people who find traditional digital services difficult, conversational interfaces may feel more natural than websites, forms and search boxes. Instead of needing to understand where to click, which page to visit, or how to phrase a search query, users can describe what they need in their own words. For someone like Kirsty, that could make a meaningful difference.

Voice interactions may remove further barriers. They can help people who struggle with typing, spelling, reading dense content, or navigating complex page layouts. The interface feels less like operating a system and more like having a natural conversation. For some people, AI may be the most accessible interface we’ve ever built.

However, accessibility alone doesn’t guarantee that Kirsty will get the right information she needs. A service can feel easy and still be unsafe, unreliable, or inappropriate for those who depend on it most. That is where the challenge begins.

Using AI safely requires a different kind of digital literacy. In the past, someone might have needed to know how to connect to Wi-Fi, fill in an online form, search for information or upload a document. Those skills still matter, but AI adds another layer.

People now need to know when to trust a response, when to cross-check it with other information, and when to seek human support. This means being aware of the technology's limitations, applying appropriate judgement, and knowing when to step away from the tool entirely. That is a high bar.

We are asking people to spot biased algorithmic decisions, often without making it clear that an algorithm was involved. We are asking them to maintain healthy scepticism while presenting AI outputs through interfaces that can sound authoritative. We are also asking people to identify the limits of systems that are often designed to feel seamless.

A clear example of this is how popular search engines have interwoven AI summary content into search results. At the start of this year, the provision of incorrect health-related advice delivered by AI summaries in this way was described as ‘dangerous’ by experts. In these circumstances, content has been extracted from elsewhere (often from accurate government or third sector sources) and repurposed with factual errors.

There’s a very real risk that this information could cause harm, especially among those with less experience accessing information online. Various studies have highlighted that when hallucinated content or inaccurate information is present, people are not very good at detecting it. We can’t assume users will catch all mistakes.

We should, therefore, be careful not to frame this only as a user education problem. Public services also have a responsibility to design AI systems that make uncertainty visible, provide routes to verification and avoid placing too much burden on those least equipped to carry it.

Taken together, these examples show how AI is reshaping digital exclusion, rather than entirely solving it. Knotty, emerging challenges need to be tackled with policy and digital working collaboratively. I also believe there are some guiding principles that should point us in the right direction to deliver better outcomes for citizens, especially for those who are underserved.

Start with the hardest cases

This requires us to start where it’s hardest: with edge cases and extreme users. This should not be a tick-box exercise late in the delivery process or to rectify half-baked live services rushed out. Our focus needs to be on underrepresented users, who are more likely to have extreme needs or to interact with services in ways that differ from expectations. AI magnifies their risks, which means our research, testing, and governance must prioritise them rather than treat them as anomalies.

Make uncertainty visible

We also need to design systems that help people reason about what they’re seeing. Not generic warnings like ‘AI can make mistakes’ (which is very easy to miss), but actionable transparency: confidence ranges for responses, accessible source links, cues that encourage second checking, and interfaces that slow users down when it matters. If AI is going to sit at the centre of large-scale services (in both the public and private sectors), it must emphasise its fallibility.

Ensure there is an accessible and appropriate alternative

AI does not need a parallel human-led route in every context. But the more consequential the service, the stronger the safeguards and routes to support or redress should be. When the stakes include healthcare, welfare payments or employment decisions, alternative routes are necessary. We need to explore how to harness AI’s benefits of scale to serve more people, while providing alternatives appropriate to the context – including human support when needed.

We started with Kirsty, someone who represents the millions still struggling to navigate digital systems. Inclusion in the age of AI doesn’t mean everyone has to be an AI expert. The focus should be on designing systems that remain safe, understandable, and accessible for people with varying levels of confidence, capability, and support.

If AI becomes a gateway to public services, we have a responsibility to ensure people aren’t excluded by the technology that was supposed to help them.

If you're thinking about how to bring AI into your organisation responsibly, our strategic AI transformation service can help you ask the right questions from the start.