"AI is inevitable" is bullshit (yatil.net)

A web-accessibility specialist argues that the idea “AI is inevitable” is misleading hype that’s already harming accessibility practice. The piece criticizes common pro-AI signals — GitHub bot commit counts, venture-backed company PR, and platform features — as poor proxies for quality or real impact. LLMs can be helpful for fast, low-stakes tasks (captions, draft translations) but regularly produce errors, inconsistent outputs (different image descriptions for the same product), and brittle fixes (e.g., ARIA attributes overwritten by successive models). The author also highlights economic distortions: many AI firms run on VC money and opaque monetization, which biases them toward pushing products even when they don’t work well for accessibility. For AI/ML practitioners, the post raises three technical and operational warnings: automated quantity ≠ quality (commits aren’t validated or merged), human review currently masks model failures but risks eroding reviewer expertise over time, and model inconsistency can introduce accessibility regressions. The recommended response is pragmatic: prioritize human training, bake accessibility into design from the start, and treat LLMs as transitional tools that need rigorous review — not as a replacement for domain expertise.