misinformation at scale

On its own a hallucination is a single mistake. The systemic risk appears when models make producing convincing false content nearly free. What once took a person hours — a plausible fake article, a fabricated review, a flood of on-message comments, a deepfake script — can now be generated by the thousand in minutes, each variant fluent and tailored. The bottleneck on misinformation used to be human effort, and that bottleneck has largely fallen away.

This changes the threat in kind, not just degree. Low-quality machine text can pollute the very web that future models train on, a feedback loop where errors get recycled as training data. Trust erodes in a second way too: once anything can be faked cheaply, real evidence becomes easier to dismiss as fake. There is no clean technical fix — defenses span provenance and watermarking, platform moderation, and old-fashioned source-checking — so a healthy default is to treat unattributed online content as unverified until shown otherwise.