One Polluted Page Is Enough: Evaluating Web Content Pollution in LLM Recommenders

Name: One Polluted Page Is Enough: Evaluating Web Content Pollution in LLM Recommenders
Uploaded: 2026-09-29T12:08:13.522847+00:00
Duration: 415 s
Description: The episode discusses a paper evaluating web content pollution in LLM recommenders using a benchmark called FORGE. The researchers found that even one polluted page can fool models across various commercial and open-weights AI, showing that deceptive volume bypasses human intuition. Defenses like skepticism prompting and consensus filters were shown to be limited by utility costs or backfire.

The episode discusses a paper evaluating web content pollution in LLM recommenders using a benchmark called FORGE. The researchers found that even one polluted page can fool models across various commercial and open-weights AI, showing that deceptive volume bypasses human intuition. Defenses like skepticism prompting and consensus filters were shown to be limited by utility costs or backfire.

Open the full episode