AI can support game accessibility at two points where teams and players need better evidence. During development, a multimodal system can inspect a playable prototype, compare settings and interactions with the Game Accessibility Guidelines, and flag cases for human testing. After release, language models can map community reviews and developer documentation to the same guidelines, creating structured reports that players can compare and researchers can update. This talk draws on my analysis of 448 Can I Play That? reviews and a 121-guideline audit of the five games nominated for The Game Awards 2025 Innovation in Accessibility award. I will explain how the two workflows can share one evidence format, where automated annotation can save time, and where it can go wrong. AI should not decide that a game is accessible. It can make the evidence easier to collect, check, and publish, while disabled players remain part of the evaluation.
1. Attendees will be able to distinguish feature presence, practical usability, and pre-purchase documentation instead of folding them into one accessibility score. 2. They will see how a multimodal model can collect evidence from a playable prototype and turn it into cases for an accessibility specialist to test. 3. They will get a reusable format for mapping reviews and developer materials to the Game Accessibility Guidelines across game versions. 4. They will know which safeguards an AI-assisted audit needs, including source evidence, confidence records, human review, and testing with disabled players.
