Where AI tools rob you of agency (and how to take it back)

The most useful version of AI is one you use to do more of the work that only you can do. The least useful version is one that gradually takes over the parts of the work you should have done yourself. The difference isn't about how much AI you use; it's about which parts. This essay names four specific patterns by which AI tools quietly subtract from your agency, with examples of each, and what to do about them.

Pattern 1: Outsourcing the first draft

You face a blank page. You ask the AI to write a first draft. You edit the draft. The work is done.

What you lost: the first draft is where the thinking happens. By outsourcing it, you didn't speed up the work — you moved the thinking to revision, where it's harder to do well. Revisions edit what's there; first drafts decide what should be there.

What to do instead: ask the AI to react to a draft you wrote, not write one for you. The role-reversal preserves the thinking.

Pattern 2: Letting the model frame the question

You have a vague problem ("I'm not sure how to position this product"). You ask the AI to help you think about it. The AI proposes three framings; you pick the most appealing one and run with it.

What you lost: framings shape conclusions. You let the AI choose the frame, then you executed inside the frame. The conclusion was determined upstream of you.

What to do instead: write your own framing first, however rough. Show it to the AI for stress-testing, not for generation. The constraint inversion prompt pattern is built for this.

Pattern 3: Accepting the average answer

You ask the AI a question. The AI gives you a fluent, well-organized, average-of-the-internet answer. You ship it.

What you lost: average answers don't differentiate you. The reason you were the one to do the work was that you bring something specific — taste, context, position, experience — that the average doesn't. By shipping the average, you erased what made the work worth doing.

What to do instead: when an AI answer feels "fine", flag it. Fine is usually the average. Add the thing only you can add — the example from your own life, the contrarian take, the detail nobody else would think to include — or don't ship.

Pattern 4: Calibrating your judgment against the AI

You wrote something. You ask the AI to review it. The AI suggests changes. You make most of them.

What you lost: you trained yourself to write toward what the AI rewards. Over months, your sense of what's good drifts toward what the AI thinks is good — which is, again, the average of what's already out there. You stopped trusting your own taste because the AI's looked more objective.

What to do instead: keep a written record of edits you rejected. Reading it back monthly is a reminder of where your taste differs from the AI's. The differences are the parts of your voice worth keeping.

The thread

All four patterns share a structure: AI was doing the part of the work that should have been the part you were proud of doing. Once you notice the pattern, the fix isn't to use less AI — it's to use AI differently. Let the AI handle the typing, the formatting, the boilerplate, the file conversion, the OCR. Don't let it handle the framing, the first draft, the contrarian take, or the final judgment about what's good.

The thinking is yours, the models do the typing. Every one of the four patterns above is a place where the rule got inverted by accident.

What this means for tool choice

Some AI tools encourage the rule inversion more than others. Tools that aggressively push you toward "let the AI write it" produce different work habits than tools that put you in the driver's seat. oran.chat is opinionated about the second posture — model picker per question, branching to preserve your judgment, instructions that name what you don't want the AI to do. Other tools (see our comparison) make different choices.

More essays on agency and AI in Essays.