AI & Technology

Too often, AI narrows rather than widens access. Here’s how the tech industry can fix it

By Gauthier Van Malderen, Founder and CEO of Perlego

Almost every AI product launch comes with the same promise: this technology will democratise access. Access to knowledge, to services, to opportunity. It’s a compelling pitch, and in many cases a sincere one. 

But a growing body of evidence suggests that when accessibility isn’t built into AI products from the outset, the opposite happens. The tools work well for the majority, and create new barriers for the people who already face the most friction.

The promise vs the reality

Higher education offers one of the clearest examples, because the impact can be measured. When Stanford researchers tested seven widely used AI-writing detectors a few years ago, the tools wrongly flagged essays by non-native English speakers as AI-generated 61% of the time on average. Essays by native speakers were classified accurately.

Those tools were adopted to protect fairness in assessment. For one group of students, they did the reverse – and the cost of a false accusation of cheating fell on those least equipped to challenge it.

The same pattern is visible well beyond education. In September, Citizens Advice called on essential service providers to guarantee a “right to talk to a human” after finding that AI chatbots were leaving people unable to resolve billing, debt and housing problems. The charity linked the issue directly to digital exclusion, which it estimates affects at least 14 million people in the UK.

Regulators elsewhere have reached similar conclusions. Dutch authorities have warned that organisations using chatbots must always offer people the option of speaking to a person, and that chatbot design must be fair and accessible. Meanwhile in financial services, the FCA’s Consumer Duty is cracking down on the need for firms to evidence delivery of good outcomes for customers in vulnerable circumstances, whatever channel they use.

The “retrofit” problem in AI design

One reason we keep seeing challenges around accessibility in AI is that it often comes too late in the design process. When accessibility features are added after an AI product’s core architecture is built, they tend to behave differently from the rest of the product. They are applied inconsistently, maintained less carefully, and often hidden in settings menus that the people who need them aren’t able to find.

The nature of AI itself makes the question of accessibility harder than it is for the likes of traditional software. The Web Content Accessibility Guidelines (WCAG), the international standard most accessibility law is based on, were designed largely with fixed content in mind – a web page written once, checked once, and that looks the same for every visitor.

By contrast, an AI response is different every time. One answer might be clear and well structured; the next might include a table that a screen reader can’t make sense of, an image with no description, or a dense paragraph of jargon. An audit can check the interface, but it can’t check responses that haven’t been generated yet.

Accessibility must therefore be built into how the AI produces its responses. The most useful question to ask of any AI product is not “does it support accessibility?”, but “at what stage of the design process did accessibility enter?” That answer reveals more about a product’s true inclusivity than any feature list ever could.

What the data shows, and what it doesn’t

Once they’re live, most organisations deploying AI tools have limited visibility into how different groups of users actually experience them. They may not know whether accessibility features are being used, where users with additional needs are dropping out of a journey, or whether the AI produces equally good results for everyone.

That gap matters, because data averages hide problems. And, crucially, teams can only fix what they can see. If the dashboard reports high accuracy and satisfied users on average, that’s what gets optimised, and nobody goes looking for the students or customers the tool

is failing. The absence of this data isn’t neutral: it builds the majority experience into the product, while the minority experience goes unmeasured and unimproved.

The obvious objection to the alternative is privacy. Organisations can’t, and shouldn’t, ask every user to disclose a need or disability. But they don’t need to. They can track how often accessibility features are used, compare outcomes for users who turn those features on with those who don’t, monitor how often people ask for a human, and test regularly with a variety of users. What matters is deciding in advance which outcomes will be measured, and for whom.

The business and regulatory case

It’s tempting to treat accessibility in AI as a future regulatory issue, but the reality is that much of the regulation is already here. Existing accessibility and consumer protection rules weren’t written with AI in mind, but none of them make an exception for it.

Take the European Accessibility Act. Since June 2025, it has covered consumer services from online shopping to banking to e-books, including services UK businesses sell into the EU. Add an AI assistant to a banking app, and that assistant becomes part of that regulated service, held to the same accessibility requirements as everything else in it.

In the UK, public sector bodies – including many universities – must meet WCAG 2.2 at level AA across their websites and apps. That puts the risk in procurement: buy an inaccessible AI tool, and its accessibility gaps become yours.

And in financial services, firms cannot meet Consumer Duty requirements to show that customers in vulnerable circumstances receive outcomes as good as everyone else’s without group-level data on how their AI performs. That puts them in danger of not meeting regulations.

These examples alone evidence why treating accessibility as a compliance checkbox, rather than a design principle, simply does not work when it comes to AI. A product can pass an audit at launch and fail the next user because its outputs keep changing. And as we’re increasingly seeing, those failures don’t stay hidden for long.

What building accessibility in actually requires

There’s no single fix, but there is a clear pattern to what works. For teams building AI products, four principles can make all the difference.

First, treat WCAG as the floor, not the finish line. Build to it from the first design stage, rather than auditing against it at the end. And because WCAG covers the interface rather than the AI’s behaviour, test the AI outputs too for clarity, consistency and compatibility with assistive technology.

Second, involve a variety of customer types, including those with disabilities, in design personas and user testing from the start – not just in a final audit. From a business

perspective, fixing problems for customers after launch is far more costly than designing properly from the beginning.

Third, measure outcomes, not just features. Decide which outcomes matter, and track them across different groups of users from day one.

And finally, always keep a human route. When AI can’t help, reaching a person should be easy.

It’s an exciting world where AI can genuinely widen access to knowledge, services and opportunity. But it won’t do so by default. The organisations that build accessibility into their AI from the start will be the ones that help to deliver on that promise, for everyone.

Related Articles

Back to top button