Why the Right Mental Model Matters More Than the Right Words
A 1972 experiment by Bransford and Johnson has been stuck in my head, because it maps perfectly onto how we work with AI.
Read this:
"The procedure is actually quite simple. First, you arrange things into different groups. Of course, one pile may be sufficient depending on how much there is to do. If you have to go somewhere else due to lack of facilities, that is the next step. Otherwise, you are pretty well set. A mistake can be expensive as well."
Makes no sense, right? Now reread it knowing it's about doing laundry.
Every sentence clicks. "Arranging things into groups" is sorting clothes. "Lack of facilities" means no washer at home. Same words, totally different comprehension. The only thing that changed was the schema you brought to it: the background knowledge your brain uses to fill in gaps.
Schema cuts both ways
Researchers call this schema activation. When the right one fires, comprehension is effortless. But when the wrong one fires, you confidently misunderstand everything. A study by Oded and Stavans showed exactly this: readers who were nudged toward a false schema wrote summaries that completely missed the author's point, and had no idea they'd gotten it wrong. The text still mostly made sense through the wrong lens, so they never questioned it.
If I'd told you that paragraph was about cleaning out a garage, you'd probably buy it at first. Arranging things, one pile might be enough, sure. But the details that don't fit? You'd just gloss over them. That's the danger: a wrong mental model doesn't feel wrong. It feels close enough.
This is what happens when we prompt AI
Every prompt activates a schema in the model. The context we give (or don't give) determines which region of its training data it draws from, how it fills gaps, and what assumptions it makes. Right framing, sharp output. Wrong framing, and you get something that reads fine but is fundamentally pointed the wrong way. No framing at all, and both you and the model are guessing, neither of you aware of it.
The best prompting isn't about clever wording. It's about giving the model the equivalent of "this paragraph is about doing laundry" before it starts writing.
Preview, Predict, Purpose
The video that inspired this post recommends a pre-reading routine that translates directly to AI work:
- Preview the problem space before prompting. What files matter? What does the existing code do? What are the constraints? Don't just start typing.
- Predict what good output looks like. If you can't picture a correct answer, you won't catch a confident wrong one.
- Purpose: are you asking it to generate, review, explain, or transform? That shapes both the prompt and how you evaluate the response.
Next time an AI response feels off, before you rephrase your prompt, ask: did I activate the right schema? Sometimes the fix isn't better words. It's a better mental model before you type anything at all.
> # Why the Right Mental Model Matters More Than the Right Words
A 1972 experiment by Bransford and Johnson has been stuck in my head, because it maps perfectly onto how we work with AI.
Read this:
> *"The procedure is actually quite simple. First, you arrange things into different groups. Of course, one pile may be sufficient depending on how much there is to do. If you have to go somewhere else due to lack of facilities, that is the next step. Otherwise, you are pretty well set. A mistake can be expensive as well."*
Makes no sense, right? Now reread it knowing it's about **doing laundry**.
Every sentence clicks. "Arranging things into groups" is sorting clothes. "Lack of facilities" means no washer at home. Same words, totally different comprehension. The only thing that changed was the *schema* you brought to it: the background knowledge your brain uses to fill in gaps.
## Schema cuts both ways
Researchers call this *schema activation*. When the right one fires, comprehension is effortless. But when the *wrong* one fires, you confidently misunderstand everything. A study by Oded and Stavans showed exactly this: readers who were nudged toward a false schema wrote summaries that completely missed the author's point, and had no idea they'd gotten it wrong. The text still *mostly* made sense through the wrong lens, so they never questioned it.
If I'd told you that paragraph was about **cleaning out a garage**, you'd probably buy it at first. Arranging things, one pile might be enough, sure. But the details that don't fit? You'd just gloss over them. That's the danger: a wrong mental model doesn't feel wrong. It feels close enough.
## This is what happens when we prompt AI
Every prompt activates a schema in the model. The context we give (or don't give) determines which region of its training data it draws from, how it fills gaps, and what assumptions it makes. Right framing, sharp output. Wrong framing, and you get something that reads fine but is fundamentally pointed the wrong way. No framing at all, and both you and the model are guessing, neither of you aware of it.
The best prompting isn't about clever wording. It's about giving the model the equivalent of "this paragraph is about doing laundry" before it starts writing.
## Preview, Predict, Purpose
The [video that inspired this post](https://www.youtube.com/watch?v=mbCAOQfBrlk) recommends a pre-reading routine that translates directly to AI work:
1. **Preview** the problem space before prompting. What files matter? What does the existing code do? What are the constraints? Don't just start typing.
2. **Predict** what good output looks like. If you can't picture a correct answer, you won't catch a confident wrong one.
3. **Purpose**: are you asking it to generate, review, explain, or transform? That shapes both the prompt and how you evaluate the response.
Next time an AI response feels off, before you rephrase your prompt, ask: *did I activate the right schema?* Sometimes the fix isn't better words. It's a better mental model before you type anything at all.
Top comments (0)