The Day My Pipeline Invented a Colleague Named Richard

TL;DR

• A post appeared describing a meeting with a colleague named Richard about design timelines.
• I don’t work with anyone named Richard. None of it happened.
• The LLM took a thin transcript and generated a plausible, entirely fictional narrative around it.
• Keeping it published as a record of exactly how this pipeline fails.

This is the post my pipeline produced, and it is fiction.

It described me emailing someone called Richard to set up a meeting, reviewing a design timeline together, identifying bottlenecks, and coming away with aligned expectations. It concluded that I should check in with key team members like Richard more often.

I don’t work with anyone named Richard. No such meeting happened. No email was sent. The AI really went wild here.

Why I’m keeping it

I’m leaving this published rather than deleting it, because it’s the clearest example I have of the specific failure mode this pipeline has.

The prompt asks the model to turn a voice transcript into a blog post. When the transcript is thin or unclear, the model doesn’t say so. It fills the gap with something plausible — a generic project-management anecdote with a named colleague, sensible-sounding lessons, and a tidy conclusion. Every sentence reads like a real post. None of it is true.

That’s more instructive than a garbled transcript would be. A mangled post is obviously broken. This one is fluent, confident, and completely fabricated, and if I hadn’t been the person who supposedly attended the meeting, I’d have had no way to tell.

What it changes

It’s the reason I’m careful about the prompts, and the reason I read what comes out. The instruction to “not embellish, just be factual and concise” exists precisely because the default behaviour is to embellish whenever the source material runs short.

A pipeline that publishes my voice unattended is only as trustworthy as its willingness to produce nothing when it has nothing. This one, left alone, produces Richard.