Who wins when we filter the open web through an opaque system? (hidde.blog)

At TPAC this year a breakout session flagged a growing risk: as users rely more on LLMs and agents to “browse” the web, we stop visiting source sites and instead filter the open web through opaque systems. That shift threatens publishers’ ad-driven business models (reducing traffic and raising hosting costs from aggressive crawlers), user privacy (centralized profiles built from queries), and democratic information ecosystems—independent media could lose revenue and visibility. More subtle but scarier is content distortion: LLMs can reword or skew facts, monetize transactions by adding hidden margins, or steer users toward preferred outcomes. Technically, the danger isn’t just hypothetical. Crawlers optimized for LLM training change traffic patterns; training and moderation choices embed biases or “ideology dials” into outputs; and data-poisoning can backdoor models—Anthropic showed 250 malicious documents can backdoor models from 600M to 13B parameters. Detection is hard when training corpora are opaque, and optimization efforts (SEO for models, ideological content farms) can amplify manipulation. The upshot: society must decide who controls these filters and under what rules. Possible responses include transparency about training data, non-commercial or “sovereign” models, and technical safeguards against poisoning and covert monetization to keep the open web genuinely open.