YouTube seems to have recently rolled out dynamically generated AI titles and thumbnails for some videos. I remember one of the harshest critics of large language models describing them as “stochastic parrots”—and in this case, I couldn't agree more.

Most of the AI-generated titles and thumbnails appear optimized purely for click-through rates, often at the expense of accuracy and clarity. It’s a textbook example of short-term engagement gains undermining long-term usability and trust.

I’ve noticed videos I’ve already skipped being resurfaced repeatedly, each time with a even more extreme or obnoxious thumbnails and titles—forcing me to manually mark each one as “not interested” just to remove it from my feed. On the flip side, if I want to revisit a video and search for it by the title I saw, YouTube often fails to surface it—because that title was dynamically generated and not part of it's meta data.

This shift raises broader concerns about the direction of algorithmic content platforms—where even basic affordances like video titles are no longer stable, and user agency continues to erode.