AI lives in a world where nobody leaves

You ask AI for a brief on a client. It cites a director who no longer works there. That’s not a hallucination: that director existed and that role was real. But the system simply has no way of knowing that chapter ended. What can we do about it?

At LLYC we learned this firsthand while building an internal AI assistant trained on more than 14,000 articles from our corporate blog. We thought that volume would make the corpus practically complete. The system knew with precision when a professional joined the firm. It had a much harder time knowing when that professional had left.

That wasn’t a system failure. It was a corpus failure. And the corpus was us. The lesson was immediate: an AI trained only on what you celebrate becomes a flattering mirror. For it to be genuinely useful, you need to train it on what you’d rather not say out loud too.

Because even as professional communicators, we document asymmetrically. We celebrate arrivals and quietly manage departures. New hires, launches, promotions, and new clients generate articles. Departures, cancelled projects, discarded tools, and abandoned practices almost never do. The legal maxim “if it wasn’t written down, it didn’t happen” ¹ coexists with the practice of not writing down anything that marks an ending.

The training process itself amplifies the bias. Recent research on reinforcement learning from human feedback (RLHF) shows that majority preferences dominate while minority ones nearly disappear ². The result isn’t a faithful portrait of real opinion — it’s a polished version of what gets the most applause. The historian Michel-Rolph Trouillot had already observed this in colonial archives: “facts are not created equal” ³. The production of documentary traces is, simultaneously, the production of silences.

And this is where internet search falls short. A model cannot autonomously search for what it doesn’t know exists. It’s a form of survivorship bias applied to AI: we only see what survived the documentation process, never what fell through the cracks ⁴. If a methodology became obsolete without anyone writing “this is no longer used,” the system treats it as still current. The absence of updates is read as persistence, not abandonment.

You can feed your assistant 14,000 articles, half a million documents, or the entire intranet. It doesn’t matter. If you haven’t told it the endings, it’s missing exactly the information that makes the difference between a useful answer and a dangerously plausible one. Who left. What got cancelled. Which tool you stopped using. What no longer applies.

This isn’t a technical limitation that someone will eventually fix. It’s the half of the conversation that the AI era is forcing us to have. We’d better get on with it.