The article argues AI coding tools amplify existing web accessibility problems, not fix them. Developers can't outsource inclusive design to a model.
ai & machine learningFriday, July 17, 2026
AI code's accessibility problem and more
Today's stories circle around the practical edges of AI: what it gets wrong, what it's good at, and how we're building around its limits. A deep look at AI-generated code reveals it's often inaccessible by default, while two hands-on pieces show local LLMs and Claude can still shine on narrow tasks. Meanwhile, Perplexity rolls out a clever skill system, and China's data annotation workforce offers a real-world view of AI's job impact.
Code and limits
Two pieces question what AI tools actually deliver, from accessibility to real-world troubleshooting.

A practical reminder that local LLMs struggle with big questions but excel at specific, well-defined tasks. Know your tool's shape.

A concrete example of Claude outperforming Microsoft's own tools on Windows troubleshooting. Niche but shows where narrow AI still wins.
Building smarter agents
Perplexity and China's data industry show two sides of making AI more capable: smarter architecture and human labor.
Perplexity's skill system uses progressive disclosure: the model loads full instructions only when needed. A smart pattern for keeping agents fast and focused.
China's data annotation boom in Guizhou offers a ground-level look at how AI creates new jobs, even as it automates others. Not a story of pure displacement.
Also today2
More roundups that day
wellness roundupvolufiline, Blue Zones, and burnout
sustainability roundupAnimation gets its own green certification
open source roundupOpen source AI hits a capability milestone
real estate roundupHousing market pain deepens across the U.S.
memoir roundupMemoir as lens on body, music, and loss
literature roundup