There is an objection that surfaces whenever a duty of care toward vulnerable users is proposed, and it deserves a direct answer rather than a dismissal. The objection is that serving vulnerable users well is a cost paid by everyone else, that designing for the edges means a slower, more expensive, more compromised product for the ordinary majority. If that were true, ethical AI for vulnerable populations would be a genuine trade-off, to be weighed against other goods.
The evidence indicates it is mostly not true. The Foundation considers this important to set out clearly, because the trade-off assumption, left unchallenged, quietly defeats the case for inclusion before it is fairly heard.
What a duty of care actually asks for
Look concretely at what designing for vulnerable users requires: clear communication; accessibility; robust performance across varied conditions, devices, and inputs; genuine, usable recourse; and conservative behaviour when the stakes are high. Now ask, of each item, whether it is narrowly an accommodation for a minority, or a general property of a good system.
Clear communication helps every user. No one prefers a system they cannot understand; clarity is not a concession to low literacy, it is a virtue for everyone. Accessibility features, captions, adjustable text, voice control, high-contrast modes, predictable navigation, are used constantly by people who would never describe themselves as disabled, in bright sunlight, in noisy rooms, while tired, while multitasking. Robust performance across varied conditions is, simply, robustness. Genuine recourse builds trust for every user, not only the one who needs to appeal. Conservative behaviour at high stakes protects everyone the high stakes apply to.
Each item, examined, turns out to be a general improvement. The list of what a duty of care requires is, very largely, a list of what makes a system good.
The curb-cut pattern
There is a well-known pattern in design history: an accommodation built for a specific group at the edge becomes an improvement relied on by the whole population. The curb cut, designed for wheelchair users, is used by anyone with a stroller, a suitcase, a delivery trolley, or a bicycle. Captions, designed for deaf and hard-of-hearing viewers, are used by vast numbers of people watching in silence or in noise. The pattern is consistent enough to be predictable: designing for the edge improves the centre.
It holds in AI systems too. A system made clear enough for a user with low digital literacy is clearer for everyone. A model made robust enough to perform well for an underrepresented group is more robust in general. A recourse process made genuinely usable by someone with little time, knowledge, or standing is more usable for everyone. The accommodation, built well, becomes the general standard.
The costs that are real
Honesty requires acknowledging the costs that do exist. Disaggregated evaluation takes more effort than a single aggregate metric. Involving affected communities in design and testing takes time and care. Building genuine recourse is more work than omitting it. These are real.
But they are not a tax paid for the benefit of a minority. They are the cost of building the system properly, and they are repaid in the currency that matters to any system: robustness, trust, and reach. A system evaluated only in aggregate is not cheaper; it is less well understood, and its hidden failures are a deferred cost that comes due later, in the world. The effort of inclusive design is largely the effort of not shipping latent failure.
The trade-off framing, reconsidered
So the objection should be answered plainly. Designing AI with a duty of care toward vulnerable users is not, in the main, a trade-off against quality for everyone else. It is a route to quality. The clarity, robustness, and contestability it demands are the properties of a system that serves all its users well.
There will be specific cases with genuine tensions, and those deserve honest, case-by-case judgement rather than a slogan. But the broad framing of inclusion as a cost imposed on the majority does not survive contact with the evidence. Far more often, the system that serves its most vulnerable users well is simply the better system, and the organisation that learns to build for the edge has learned to build well for the centre too.
