Are we offloading too much of our thinking to AI?

TL;DR Highlight

Essay urges forming your own hypotheses before asking AI, not outsourcing all thinking.

Who Should Read

Developers using AI for research, design, and code review who worry their own judgment is getting dull

Core Mechanics

Offloading thinking reached trivial decisions
- The author sees growing tendency in herself and others to hand research, reasoning, and answers to AI.
- She says Ken Liu's 2012 story "The Perfect Match" and its assistant Tilly depicted this accurately.
Microphone Man outsourced everything
- At a San Francisco startup event, a man recorded all his conversations and had AI summarize and analyze them.
- He said Claude Fable is better at critical thinking than him, so he lets it do all his thinking.
Search and AI skip different steps
- Search still required you to break down questions, evaluate sources, and synthesize.
- AI performs those intermediate steps and returns a finished answer in minutes.
Hypotheses first, AI second worked well
- In Portugal, the author and her sister guessed, debated, and built theories before asking AI.
- AI then confirmed many theories, added missed explanations, and omitted some possibilities they still found plausible.

Community Reactions

Junior who said 'I don't know'
- One commenter's junior couldn't explain a computation in a design review.
- It was AI-generated and wrong, and he couldn't tell.
Arguing on behalf of agents
- A commenter described a production-fix debate over connection pooling and threading where everyone just defended their agent's advice.
- Admitting they didn't know enough made the commenter fear looking incompetent.
Scoping AI use deliberately
- A 20-year developer commenter uses AI for debugging, review, typing out finished designs, and idea suggestions, not learning or deciding.
- Another commenter guesses a calculator answer first and does the same with LLMs.
Objection: hallucinations ignored
- One commenter argued the essay assumes AI answers are correct.
- They said AI can invent answers, which should be addressed first.
Consulting cleanup work growing

How to Apply

Write your 3-line hypothesis before asking
- Today, before asking AI about a design or root cause, write your expected answer in 3 lines.
- Then compare with AI's answer and mark matches, gaps, and disagreements.
Split AI use into allowed and restricted
- Allow AI for debugging, review, and implementing designs already decided.
- Write in team rules that design decisions and learning tasks are done by you first.
Make authors explain AI output in reviews
- In design reviews, ask authors why AI-generated code or calculations were done that way.
- If they can't explain, have them understand it and resubmit.

Try It

copy & run

# Paste into Claude or ChatGPT as is (prompt version of the essay's 'hypothesis first, AI second')

You are a partner who verifies my thinking, not replaces it.
Read my hypothesis first, and before generating a new answer, respond in this order.

Question: {{question}}
My hypothesis (3 lines): {{my hypothesis}}

1. Which parts of my hypothesis look right
2. Weak points or likely errors in my hypothesis
3. Explanations I missed
4. Two facts I should verify myself

Mark anything you are unsure about as uncertain.

Next step: after reading the AI's answer, pick one point you disagree with and write a rebuttal.

Terminology

Deep ResearchA feature where AI searches, reads, and compiles on its own into a long report. Like handing research to a junior analyst.

METR Time HorizonA metric of the length of software tasks, in human working time, that an LLM completes with 50% success.

tokenmaxxUsed in a comment for management forcing engineers to maximize token usage.