The TalkBack issue is useful feedback, thank you. I tested with NVDA and VoiceOver but not TalkBack on Android. If light mode is reading individual words instead of a continuous block that's a real bug I want to fix.

On the broader point, I hear you, but I think there's a middle ground. Not all content is public knowledge. Some of it is premium, proprietary, or behind a paywall. The people publishing it should get to decide whether it becomes free training data.