The best shift-left is to shift accessibility into AI model training and not bolting it on later – Bogdan on Digital Accessibility (A11y)

AI models are trained on the web's inaccessible content, inheriting its biases and bad practices. Rather than patching accessibility through prompting, RAG, or MCP workarounds, the argument is that model trainers must integrate accessibility from the ground up — with accessibility specialists and people with disabilities involved in training. Benchmarks like AIMAC show better accessibility results are achievable but currently only with expensive models, which itself contradicts accessibility as a human right. Stronger regulation (e.g., EU AI Act) and louder advocacy are needed to ensure accessibility is baked into AI model training rather than bolted on afterward.