Why AI Shorts generators lie about real people and books — and how it's fixed
Large language models have an unpleasant habit: ask them about something they don't actually know, and they will very often not say "I don't know" — they'll invent an answer, confidently and smoothly. For topic mode, that's a specifically dangerous trap: a topic often names something concrete — a book, a person, an event — and if the model doesn't actually know that specific thing, it can quietly swap ignorance for fabrication, with nothing in the tone giving it away.
The real incident that triggered this
That's exactly what happened on a topic about a specific book: the model didn't know the book, but instead of an honest "I don't know," it confidently narrated an invented plot — characters, setup, resolution — as if it were the real content. To a viewer who actually knows that book, it would read as a flat lie delivered with an expert's tone. And this isn't about one particular model — it's how language generation behaves by default unless something architectural stops it.
Classify first, write second
Vertix's fix isn't asking the model to "not make things up" — that's an unreliable instruction — it's not giving it the chance to in the first place. Before a script gets written, a separate, cheap classification call sorts the topic by type: a general explanation of something external, an inner feeling, a breakdown of a specific work, a story about a real person, a real event, pure fiction, or a how-to. Separately, it also flags whether the model actually knows the named thing at all.
The topic's type isn't a cosmetic label — it's mixed into the prompt and switches on different honesty rules. For a specific work, recounting the plot from memory is explicitly forbidden. For fiction, the opposite: the restrictions on invention are lifted, because there's nothing to lie about — it's already meant to be made up.
What happens when the model doesn't know
If a topic is about a specific work, person, or event, and the classifier honestly determines the model doesn't know it, the bot doesn't silently write a script and hope nobody notices. It says so directly and offers a choice:
- "I'll describe it myself" — you give a short description, and it becomes part of the topic the script is written from
- "My own full script" — you write the script yourself, with no model involved at all
- "Write it anyway" — the model writes something, but under an explicit ban on faking familiarity: no invented plot, characters, or dates
- Cancel
It survives rewrites too
The classifier's verdict — whether the model knows the thing or not — carries over into regeneration. If you tap "rephrase," the second and third versions of the script are written with the same topic type and the same doesn't-know flag, not as if it's suddenly familiar. That didn't use to be the case: a second attempt used to lose the flag and get written as commentary on something familiar, without the honesty rules the whole mechanism exists for. Now the verdict holds across every attempt.
Why this isn't about hiding that it's AI
The goal here isn't to disguise that a script was written by a model — that's a separate, and honestly not very interesting, problem. The goal is that what the model says is either true or honestly flagged as unknown, not fiction delivered with the intonation of fact. The difference is simple: a viewer who knows the topic better than the bot does shouldn't run into a confidently delivered lie.