Edited by us
How to stop AI inventing facts — the fact-slot method
Most AI writing failures are missing-fact failures: the model fills the gap with something plausible. Sorting your material into facts, decisions, and slots fixes it before it happens.
By Mephify (operator)2 min read
Why AI invents things
A study of 1.5 million ChatGPT conversations found that about two-thirds of writing requests ask the model to edit, shape, or improve material the person already has, rather than conjure text from nothing. That matches how the failures happen: the model doesn't invent because it wants to. It invents because a needed fact is missing and something plausible fills the seam: a date, an amount, a name, a claim you never made.
The fix isn't a smarter model. It's a request that separates what you know from what you want decided from what genuinely isn't known yet.
The method — three piles
Pile 1: facts — only you have these
Dates, amounts, names, what actually happened, what the other side wrote. Put them in the request verbatim, even if it feels long. Every fact you leave out is a seam the model will fill.
Pile 2: decisions — you choose, it executes
Tone, length, recipient, what to emphasize, what to leave unsaid. These aren't facts and they aren't unknowns. They are calls only you can make. Make them in the request, or the model makes them for you and you edit against its choices.
Pile 3: unknowns — mark them, don't fake them
The account number you'd have to look up, the exact statute for your state, the manager's full title. Tell the model to leave a marked slot for each, written as [FILL IN: account number], instead of writing around them or guessing. A visible gap gets fixed in one pass; an invisible guess ships.
The one line that changes the output
Use only the facts I gave you. Anything missing, mark as [FILL IN: what goes here]. Don't fill it in yourself.
Checklist before you hit enter
- 1Are the facts in the request itself, verbatim rather than summarized from memory?
- 2Did you make the decisions about tone, length and audience, or leave them to the model?
- 3Are unknowns marked as slots rather than left as invisible gaps?
- 4When the draft comes back: does every number and name trace to something you supplied?
Instructions built here carry this discipline by default. Every 'be specific' comes paired with 'only from the user's input', and unknowns come out as marked slots rather than confident guesses. It's the difference between a draft you check and a draft you have to interrogate.
Reading helps; trying is faster.