Hire Great Thinkers, Let AI Write: Why It Works
Writing used to be a decent proxy for thinking. AI broke that link. Here is how to hire for the thinking and let the model do the typing.
6 min read · Last updated: October 25, 2026
Hiring great thinkers and letting AI write means judging people on their decisions, not their prose. Polished writing used to signal clear thinking. Language models now produce clean prose for anyone, so the prose proves little. What still separates people is what they decide to say, to whom, and why.
Came here from Instagram, LinkedIn, X or YouTube? Here is the full source list you came for. Everything below is built from these, nothing else.
The idea
- "Hire Great Thinkers. Let AI Write." on Medium (the essay that prompted this post. Medium blocked my fetch, so I used only its title and one-line summary, not its body)
The research
The practitioner view
Key takeaways
- Writing was a proxy for thinking. AI made the proxy cheap to fake.
- In an MIT Media Lab preprint, essay writers using an LLM showed the weakest brain connectivity of three groups, and the lowest sense of ownership over their essays.
- Chris Lema lists roughly 30 strategic choices that happen before a draft, and says AI skips them.
- Test candidates on decisions: audience, angle, what to cut, and what would change their mind.
- Do not use this for roles where the prose is the product, or where nobody can check the facts.
What is the real-world comparison for writing as a proxy?
A proxy is a cheap signal you use because the real thing is hard to measure. Writing worked that way for hiring: a tidy cover letter or a clear design doc suggested a tidy mind behind it. Think of a firm handshake standing in for trustworthiness. It held up until someone could fake it for free.
That is what happened. Producing fluent text used to take effort and practice, so fluency correlated with having done the thinking. A language model, meaning software trained on large amounts of text to predict the next word, now produces fluent text on demand.
The signal did not get weaker because people got worse at thinking. It got weaker because it stopped costing anything.
Why did AI break the link between writing and thinking?
AI separates the typing from the deciding. A model can draft a clean paragraph about any topic without having formed a view on it. The decisions that make writing worth reading, such as who it is for and what to leave out, happen before the draft, and the model does not make them for you.
Chris Lema, in his March 3, 2026 post, makes this point directly. He says AI output is clean but "Perfectly forgettable," and that roughly thirty strategic choices come before drafting, including the specific audience, the core angle, the structure, your own voice and the distribution channel. His argument: speed is not the bottleneck. Knowing what to say, to whom, and why is.
Note that Lema also sells products built around this idea, so read his framing as a practitioner's view, not neutral research.
What does the research say about thinking when AI writes?
A 2025 MIT Media Lab preprint, "Your Brain on ChatGPT" by Nataliya Kosmyna and colleagues, compared three groups writing essays: one using an LLM, one using a search engine, and one using no tools. According to the arXiv abstract, brain-only writers showed the most extensive neural connectivity, search users moderate, and LLM users the weakest.
LLM users also reported the lowest sense of ownership over their essays and had trouble accurately quoting their own work.
| Group | Brain connectivity (EEG) | Ownership of essay | Quoting own work |
|---|---|---|---|
| No tools | Most extensive | Not reported as lowest | Not reported as a problem |
| Search engine | Moderate | Not reported as lowest | Not reported as a problem |
| LLM | Weakest | Lowest | Had trouble |
Two cautions. This is a preprint, not a finished peer-reviewed result. And the authors say LLM convenience "may" carry cognitive costs, which is a hypothesis, not proof that AI makes people less intelligent. Use it as a reason to test thinking directly, not as a verdict.
How do you hire for thinking instead of writing? (step by step)
You test decisions out loud or on a constrained page, then let the candidate use AI for the polish. The goal is to see what they choose, not how well they phrase it.
- Give a messy brief. Two paragraphs of real, conflicting context. Not a clean prompt.
- Ask for decisions before a draft. Who is the reader? What is the one claim? What are you cutting?
- Ask what would change their mind. A real thinker can name the evidence that would make them wrong.
- Then allow AI for the draft. Compare the draft to their decisions. If the draft drifts from what they chose, they did not own the thinking.
Each step produces something an AI cannot fake in a live conversation: a position you can push on.
What does a thinking-first scorecard look like?
A small rubric keeps the interview honest. Score the decisions, ignore the polish.
BRIEF -> candidate decisions (scored 0-2 each)
audience: who exactly reads this? [ ]
one claim: what must they believe after? [ ]
cut list: what did you leave out, why? [ ]
falsifier: what evidence would flip this? [ ]
DRAFT -> AI-written, candidate edits (scored 0-2)
matches decisions? [ ]
catches a wrong or invented fact? [ ]
TOTAL /12 (below 6: strong prose, weak thinking)
The last line is the point. A candidate can hand in a beautiful draft and still score low, and the rubric will show it.
When should you use this approach, and when not?
Use it for roles where judgment drives the outcome and drafting is overhead: product, engineering leadership, strategy, support leads, technical writing review. Avoid it where the prose itself is the product, or where no one on the team can check what the AI wrote.
| Situation | Hire for thinking, AI drafts | Why |
|---|---|---|
| Internal docs, specs, status updates | Yes | Decisions matter more than phrasing |
| Engineering and product roles | Yes | Judgment is the job |
| Brand voice, fiction, speechwriting | No | The prose is the product |
| Regulated or high-stakes text | Only with a human checker | Fluent errors read as correct |
| Teams with no domain reviewer | No | Nobody can catch bad output |
The risk is simple. AI produces confident text whether or not it is right, so someone on the team must be able to tell.
What happens to developers and juniors?
The same shift applies to code. If typing code or prose is cheap, the scarce skill is deciding what to build and checking it. I covered that in what coding skills survive AI in 2026 and in code was never the hard part.
For people learning, the lesson is not to avoid AI. It is to form your own view first, then use the model to speed up. How to use AI to learn programming faster shows that order in practice.
FAQ
Is writing still a good signal of thinking?
It is a weaker signal than it was. Fluent text is now cheap, so a polished sample proves little on its own. It still helps when you can ask follow-up questions live, because then the candidate has to defend the choices behind the words.
Does using AI to write make you think less?
The MIT Media Lab preprint found the weakest brain connectivity in LLM users among its three groups, but its authors frame the result as possible cognitive cost, not proof of lower intelligence. It is early evidence from one preprint, so treat it as a caution, not a conclusion.
How do I test someone's thinking in an interview?
Give a messy brief and ask for decisions before any draft: the reader, the one claim, what to cut, and what would change their mind. Score those answers, then let them use AI for the draft and check whether it matches their choices.
Should I ban AI in writing assignments?
Banning it mostly measures who ignores the rule. A better test is to allow AI and grade the decisions and the corrections. That reflects how the work will actually be done.
If you want to practice the thinking side, try the free tools on Art of Code, and see why AI is not ready, and what that means for you or browse open roles on /jobs.
Written by Adithya Guttha, Founder of Art of Code.
Sign in to like and save posts.
Keep reading
Discussion
Comments (0)
Want to join the conversation?
Sign in to commentNo comments yet. Be the first to comment!