Understanding AI · Workbench
AI writing has a reputation for being generic. Ask an AI for a strategy memo and you can get a memo that could belong to any company. The reputation is fair when the prompt is the only thing you give the AI, because the AI fills every gap with the most common answer. This page shows how to put yourself into the work: your problem, your thinking, your notes and your judgment. At the end you build a prompt from your own material.
Guess first
Tap an answer
In a 2024 experiment, about 300 people each wrote a short story. Some writers were given story ideas from ChatGPT and some were not. Hundreds of readers then rated the stories. Compared with the stories written without AI, what happened to the stories written with AI ideas?
The mixture
An AI model learns from what millions of people have written, so the AI is good at the common pattern: how a memo is usually structured, which points usually come up, and which words usually follow. When a request leaves a gap, the AI fills the gap with that common pattern. Five people who send the same short prompt get five drafts that look much alike.
You hold what the AI cannot know: the problem in front of you, the conversation you had on Tuesday, the client who always asks about cost, and the way you speak. Each of those details narrows the gaps the AI would otherwise fill with the common answer. The draft that comes back is a mixture of the two, and the share that comes from you is what makes the draft yours.
The diagrams are our illustration of the idea, not measured data. The experiment in the question above measured the left-hand effect: stories written from one ChatGPT idea were more alike than stories written without AI, by 10.7% of the full range of the researchers’ similarity score (Doshi and Hauser, 2024). The Pick a Number page shows the same bunching in its simplest form, when chatbots asked for a random number answer 7 far more often than people do.
How to put yourself in
Each ingredient below replaces a gap the AI would otherwise fill with the common answer. Start with the first two, and add the others as the work matters more.
Define the problem yourself before you open the AI. Name who the work is for, what decision or action the work should lead to, and the constraint that makes your situation different from the textbook case.
Generic: “Write a plan to improve customer retention.” Yours: “Our renewal rate on small commercial accounts fell from 88% to 81% after the price change in March. Brokers say clients are shopping on price. I need a two-page plan for the regional VPs that I can defend on Thursday.”
Spend ten minutes writing or talking through the problem before you ask for anything. Record a voice memo on a walk, or type a page of messy notes with half-ideas, doubts and the option you dislike. Paste the raw version into the AI and ask the AI to organize your thinking, not to replace your thinking.
The rough page holds your priorities and your doubts, which are the details a short prompt leaves out.
Give the AI what happened, not a summary of what usually happens: your meeting notes, the transcript of a discussion with your team, the client’s email, the numbers from last quarter. Ask the AI to quote the record when the AI makes a claim, so you can check each claim against what was said.
Check your organization’s AI policy first, and remove names and client details the tool should not see. The Safe AI Usage guide covers what to keep out.
Treat the first draft as raw material. Work through several rounds, and in each round tell the AI exactly what is wrong and what you would say instead.
The AI still does a large share of the work in this mixture. The AI structures your notes, suggests options you missed, argues the other side, and drafts quickly. Your share is the problem, the record and the final judgment.
Your turn
Pick a real piece of work due this week. Fill in the four boxes with your own material, then copy the prompt into the AI tool you use. Rough notes are better than polished ones here. Your answers stay in this browser and are never sent to us.
After the first draft comes back, run the four rounds of edits above. Keep the rounds in the same chat, so the AI can see your earlier corrections.
Where this works
Each example contrasts the prompt alone with the prompt plus the person who asked.
Prompt alone“Write a quarterly IT update for the board.” The draft lists the usual themes of cyber risk, cloud and AI, in an order any company could use.
Prompt plus youPaste the three decisions the board must make, last quarter’s minutes and the question the chair asked you in the hallway. The draft now answers this board.
Prompt alone“Draft a proposal for a claims process review.” The draft reads like a template.
Prompt plus youAdd the transcript of the discovery call and the client’s own words for the problem. Ask the AI to use those words in the opening paragraph.
Prompt alone“Suggest priorities for our team next year.” The AI returns a list of common priorities for a team like yours.
Prompt plus youRecord the team’s whiteboard discussion, paste the transcript and ask the AI to group what the team said and to name the disagreements the team has not resolved.
Prompt alone“Write an email declining a vendor’s renewal.” The draft is polite and could be from anyone.
Prompt plus youGive the history in a few lines, say what relationship you want to keep, paste one email you wrote to this vendor before, and edit the result until the email sounds like you.
Use it this week
The next time you open an AI tool, stop before you type the request and add one thing the AI could not know without you: the real problem, your rough notes, the record of a conversation, or a sample of your own writing. Then edit the draft in rounds until you would sign the draft as your own.
AI fills every gap with the common answer, so the work becomes yours when you fill the gaps first.
To see why AI favours the common answer, try Pick a Number Between 1 and 10. To see where AI helps and where AI misleads, try The Jagged Frontier Quiz. Working well with AI is one part of digital intelligence, and the 30-day course builds the rest of that skill one day at a time.