If you don’t know about the gorilla, take the selective attention test.
If you do, take the sequel test.
The short version
When you’re immersed in your draft, you can’t see it the way your reader will.
Ask AI to roleplay your reader, and it will point out what you missed.
Four steps: build a profile, get AI into character, filter the reactions, fix the gorillas.
I don’t think the interesting part of AI is that it can do something for you. The interesting part is what it allows you to see what you couldn’t see.
Because you’re so immersed in the content when you’re writing, it’s easy to have a perspective problem like the one in the video. You need a way to jump into the reader’s point of view.
AI can read from whatever perspective you tell it to. My favourite way of doing this is using in-character roleplaying agents. You tell AI, “Be my reader,” and this turns out to be surprisingly useful.
What you might find
AI readers can show where:
you made a thinking typo: errors in logic, mixed up dates, forgotten URL etc
it’s possible to misinterpret something
the draft raised a question that it didn’t answer
Below is how I do this. If you want to test it, I put full example prompts at the end.
Step 1 - Build an audience profile
I capture and these things about a main reader and give them to AI:
their role in relation to my content
any easy characterizations (she’s like Hemingway in a workplace)
specifics that matter (he asked me to research X)
mood or bias (oppositional, receptive, nitpicky, friendly, impatient)
I resist the temptation to add a lot of detail because it doesn’t seem to add value. AI test readers can’t and don’t need to match reality exactly.
(Doing this has the secret added benefit of getting me to think more about who I’m writing to.) You can also save audiences to reuse later if you write AI skills or build your own workflows.
Step 2 - Get your agents into character
Tell AI who to be
Say “you” as if you were talking to the reader. “Let’s roleplay. You are…” “You just received this document” then attach the document.
Add constraints
I usually force AI readers to stick to,
“Where did you have questions?”
“Was anything confusing?”
“Do you have enough information to make a decision or take action?”.
This helps limited, focused responses from multiple of AI readers using different models and levels of reasoning.
Step 3 - Filter reactions
AI will generate more than is relevant. I don’t overanalyze reader reactions—it would be expecting too much from AI. Just grab insights and resolve them in the draft. For more sophisticated AI users, you can pass results to a single strong model to consolidate reactions and trim AI noise.
Step 4 - See if you find any gorillas
AI should effortlessly deliver you some gorillas. It makes you wonder, “How did I miss that?” It’s really fun to catch those things and fix them before you send the document. Once you see it, you can’t unsee it and probably won’t make the same mistake twice.
This technique turns the weakness of artificial intelligence to your benefit. Models are trained on mass data and have painfully conventional, average opinions. If they’re confused or raise convincing objections, your real readers will too. There’s nothing worse than writing a good document and having people get stuck on an error that takes attention away from key ideas.
Case in point
(I’m thinking about the famous Starbucks cup that made it into a scene from Game of Thrones, which was almost more interesting than the episode itself. An AI audience would have found the Starbucks gorilla for sure.)