The Washington Post's own AI invented quotes and ran commentary as the paper's editorial position. By December 2025, leaked Slack messages from its standards editor described an AI-generated podcast doing exactly that, plus misattributing sources. The technical failure was small: a missing citation-verification step. The reputational damage was global. Publishers are already cornered. AI Overviews now show on 48% of Google searches, and publisher search traffic fell about 33% in the year to November 2025 — with news executives expecting a further 43% drop by 2029. But "build our own AI" is exactly where the Post got burned, and mid-tier publishers can't replicate the six-engineer ML teams the FT, Bloomberg, and NYT built before the cliff. The real work is unglamorous: cleaning decades of archive, resolving entities so "Mr. Musk," "Elon Musk," and "the Tesla CEO" collapse to one node, enforcing citations on every answer, and putting an editorial review queue in front of the masthead. It's two plays, not one — run your own engine for retention, and opt into leakage-capture deals (ProRata's 50/50 pool for 2,200 publishers, Cloudflare Pay Per Crawl) for the AI engines you'll never own. And the fear that a chatbot cannibalizes subscriptions? Ask FT's data showed the opposite — retention lift, not erosion. If you shipped an archive chatbot tomorrow, what stops you first — the engineering, or the WaPo-style fabricated quote nobody caught in review? #PublisherAI #GraphRAG