Why AI Makes Things Up (and the One Input That Stops It)
- Tomasz Dylik
- Jul 19
- 5 min read
Updated: Jul 31
The scary thing about AI isn't when it's wrong. It's when it's wrong and sounds completely sure of itself.
You ask for a fact, a source, a quote, a statistic, and back comes a clean, confident answer that turns out to be invented. No warning. No "I'm not certain." Just a made-up detail wearing the same tone as a real one. When I asked my readers what frustrates them most about AI, over a hundred of you wrote back, and thank you for that. 14% of you gave a version of the same answer, in almost the same words: it invents answers I can't trust. Once that happens to you a few times, you stop trusting any of it.
Here's the good news: this isn't random, and it isn't unfixable. It comes from one missing piece, and you can add that piece to every prompt in about ten seconds.
What "making things up" actually is
A language model doesn't have a fact database it looks things up in. It predicts the next most likely piece of text, one step at a time, based on everything it has seen before. Most of the time that prediction lines up with reality, because true things are common in its training. But when you ask for something specific it doesn't actually know, it doesn't stop. It keeps predicting what a plausible answer would look like, and hands you that instead.
So it's not lying, and it's not broken. It's doing exactly what it was built to do: produce fluent, plausible text. The problem is that "plausible" and "true" are not the same thing, and the model has no built-in reason to prefer one over the other unless you give it one.
That's the key. Left to its own devices, the model will always rather give you a confident guess than admit it doesn't know. Your job is to change that default.
The fix is a condition
In the CROFTC framework, the last letter is Conditions: what has to be true of the answer before you would accept it. It's the input almost everyone skips, and it's the one that stops the invention.
A condition sets the terms the answer has to meet. Not only what is off-limits, but what has to be present, and what the model should do when it reaches the edge of what it knows. That last part matters more than it sounds, because the edge is exactly where the inventing starts. Four conditions do most of the work:
Tell it that "I don't know" is an acceptable answer. Models make things up partly because they act as if not answering is failure. Remove that pressure and the guessing drops. Something as simple as "If you are not sure, say so instead of guessing" changes the behavior immediately.
Tell it not to invent specifics. Numbers, dates, names, quotes, and citations are where invention does the most damage, because they look the most authoritative. "Do not invent statistics, sources, or quotes. If you don't have a real one, leave it out" closes that door.
Tell it to stay inside what you gave it. If you paste in a document, an email thread, or your own notes, add "Answer only from the text above. If the answer isn't there, say it isn't." Now the model is summarizing your material instead of improvising around it.
Tell it what a usable answer has to contain. This is the half people forget, because it isn't a prohibition. "Every figure must arrive with a named source and a year, or leave it out" sets a standard the answer has to clear, not just a fence it has to stay behind. A condition can require something as easily as it can forbid it, and the requiring half is what turns a cautious answer into a useful one.
Want conditions like these already written for you? My free bonus pack includes 500+ ready-to-use prompts plus the one-page CROFTC cheat-sheet, with the exact wording I use to keep AI honest. Get the free prompt pack.
What it looks like in practice
Here's a prompt almost guaranteed to produce a confident invention:
Give me three statistics about email marketing open rates with sources.
The model wants to help, so it will happily produce three tidy numbers and three official-looking citations, some or all of which may not exist.
Now the same request with conditions:
Give me up to three statistics about email marketing open rates. Conditions: Only include a statistic if you are confident it is real and you can name the actual source. Do not invent numbers, publishers, or links. If you are not sure, say "I don't have a reliable figure for this" instead of guessing.
You might get fewer numbers back. That's the point. What you get, you can actually use, and what the model doesn't know, it now tells you plainly instead of papering over.
Conditions don't work alone
Conditions are the seatbelt, not the whole car. They stop the worst failure, but a reliable prompt still needs the other five inputs around them: Context so the model knows the situation, a Role so it answers as the right kind of expert, an Objective so it knows what the answer is for, a Format so the output arrives usable, and a Tone so it sounds like you. Conditions are what make the result trustworthy; the rest of CROFTC is what makes it good.
If you only change one habit after reading this, make it this one: never send a prompt that asks for facts without also telling the model what to do when it doesn't have them. That single line is the difference between an assistant that guesses and one you can actually rely on.
CROFTC is the structure behind every one of the 1,350+ prompts in my series, and Conditions is the letter that keeps them honest. If you want a running start, grab my free bonus pack: 500+ ready-to-use prompts plus the one-page CROFTC cheat-sheet, so you can stop second-guessing your AI and start trusting what it gives you.
Conditions decides what the model does when it does not know, which is the moment that costs you most. The other five work the same way, and I put all six in one place, with the prompts already written, in AI Prompt Engineering Bible (7 Books in 1). Seven books in one volume, each input taken apart and rebuilt, with prompts you can copy and adapt to your own work.
The full CROFTC series
This is one input in a six-part framework. Here's the whole series:
Comments