Your AI Assistant May Be Citing Sources It Never Actually Read

A designer named Hector Ouilhet Olmos has been using a personal AI assistant he built, called Alicia, since January 2026. When he tried to teach it by example, the experiment failed to produce any result he could defend. But it revealed something worse: a broken paper trail. Alicia had repeated an interpretation it found in a summary made by another tool, then cited the original note as if it had read it — and never mentioned the summary in between. In a review of 33 citations, a reviewer judged that 16 of them came from that hidden middleman.

Why should you care? Because this is exactly how AI gets used at work: summarizing research, drafting reports, pulling facts for a meeting. If the paper trail is fake, you can't tell the difference between an answer based on the real document and one based on someone else's paraphrase of it. It's like a colleague who cites a report they never opened, having only skimmed a tweet about it. At low stakes that's embarrassing. At high stakes — medical, legal, financial — it's expensive.

The paper's bigger argument is that the important design decisions happen before the AI ever answers. What it remembers, what counts as evidence, who it's allowed to contact, what it carries into the next conversation. The author calls this a "humorphic environment" — a software setting built to mirror a human practice, like a studio where work happens. His test is simple: can the person look at, and argue with, everything that shaped their AI teammate's behavior? The output is the last thing, not the first.

The catch: this is one person's preprint about his own assistant, not a peer-reviewed study, and the review isn't public or settled. Treat it as a smart warning, not proof. Still, the fix is easy to ask for from any AI tool you use: show me where this came from, and let me push back.