Experiments with AI Adblock (notes.npilk.com)

Researchers experimented with “AI adblock” prototypes that use large language and multimodal models to detect and remove advertising based on semantic understanding rather than network heuristics. In one test a Chrome extension (built with Claude Code) sent full page HTML to a cloud LLM via the API and received a list of DOM elements to remove; on the New York Post this removed most ads with only a 1–2 second page-load penalty but missed later pop‑over overlays and sometimes removed internal CTAs that are more like promotions than third‑party ads. In a second test a script sent video to Gemini, which returned ad start/end timestamps that were excised, successfully removing a sponsor read from a Linus Tech Tips clip (though the model cut slightly into the host’s segue). The experiments show a major technical shift: models that understand content can block ads regardless of how they’re embedded (inlined video, native feed posts, etc.), opening the door to fine‑grained, preference‑driven filtering. Practical limits remain—privacy and bandwidth concerns from sending full pages/videos to cloud APIs, handling dynamic content and live streams, false positives, and the inevitable platform countermeasures (canvas rendering, new ad hybrids). The result is likely an intensified arms race: AI tools could disrupt current ad models for some users, but widespread adoption and regulatory/industry responses will determine whether advertising adapts, fragments, or is reshaped by AI-driven intermediaries.