AI, Accessibility, and Open Source: Who gets left behind?

Open source promised that anyone could participate. AI promised that documentation would get easier to create. Accessibility research promised that better tooling would mean more inclusive software. None of these promises has been fully delivered, and when you put all three together, a pattern emerges that the FOSS community is not talking about enough.

AI-assisted documentation is being adopted fast. Teams are using LLMs to generate first drafts, summarise long pages, and auto-translate content. On the surface, this looks like progress: more documentation, faster, in more languages. But the accessibility research tells a different story. AI-generated content consistently produces heading structures that break screen reader navigation, image descriptions that describe aesthetics rather than function, and plain language scores that exclude non-native speakers and cognitively diverse readers. The tooling that is supposed to make documentation easier to produce is making it harder to use for exactly the people who most depend on documentation being right.

Meanwhile, the open source communities building these AI tools are themselves inaccessible. Contribution workflows that assume a particular development environment, community spaces that are not screen reader compatible, documentation that is the only onboarding path, and fail the people trying to walk it. Disabled contributors are being excluded from the projects shaping how AI handles accessibility.

This talk connects those three threads and asks a direct question: who is making decisions about AI-assisted documentation, and are the people most affected by those decisions in the room?

Drawing on recent accessibility research, real examples from FOSS projects, and firsthand experience as a technical writer and accessibility advocate, this talk maps where the gaps are, why they compound, and what the open source community can do to close them before the defaults get harder to change.

1. The specific accessibility failures AI-generated documentation consistently produces, and why current AI review workflows are not catching them.

2. How the exclusion of disabled contributors from FOSS AI projects creates a feedback loop that makes AI-assisted docs less accessible over time.

3. What recent accessibility research says about the gap between AI documentation output and the needs of screen reader users, cognitively diverse readers, and non-native English speakers.

4. Concrete interventions in review processes, contribution workflows, and community design that break the compounding exclusion cycle.

5. A call to the FOSS community to treat accessibility not as a post-launch audit, but as a participation requirement in the project’s building of AI documentation tools.