Use AI where evaluation is stronger than generation

The governing rule: keep judgment you can evaluate

A useful principle from your broader operating philosophy transfers directly to writing:

Do not delegate material judgment when you cannot evaluate or challenge the result.

AI is most useful when you have enough independent understanding to reject it.

Each stage gives AI a different role and a different risk

The rule creates a simple matrix. For each task below, the first part names AI's role and the second names the main risk.

• Raw idea discovery. Socratic interviewer. Risk: leading the user toward the model's framing.
• Research. Finder, summariser, source comparator. Risk: fabrication, source flattening, false confidence.
• Thinking Essay. Challenger, dependency mapper. Risk: writing the synthesis for the author.
• Reader Architecture. Cold-reader simulator, structure critic. Risk: inventing reader beliefs.
• Drafting the final essay. Limited or none by default. Risk: voice substitution, false coherence.
• Revision. Diagnosis, counterargument, local edit. Risk: improving sentences that should be deleted.
• Fact-checking. Claim extraction, verification plan. Risk: treating model confidence as evidence.
• Source fidelity. Comparison and traceability. Risk: close paraphrase or unmarked invention.

AI as interviewer asks one question at a time

This is one of the strongest uses, because it adds friction rather than removes it.

A good AI interviewer does these things:

• asks one primary question at a time
• restates what it understood
• labels epistemic status
• identifies the strongest and weakest part of the answer
• asks for numbers, dates, sequence, mechanism, or counterevidence
• proposes interpretations only when clearly labelled as hypotheses
• stops when the conversation is circling

A bad interviewer asks a list of generic questions or supplies the answer inside the question.

AI as hostile reader infers what the draft says

Give the model the draft without your intended brief. Ask it to infer these six things:

• reader
• central claim
• purpose
• strongest evidence
• major qualification
• unresolved question

Then compare its inference with your architecture. This is more useful than asking, "Is this good?"

AI as counterargument generator gives one strong objection

Ask for the strongest case against the claim, not ten objections. Then require it to state:

• which premise it attacks
• what evidence would support the objection
• whether the objection narrows, reverses, or leaves the conclusion intact

Do not accept counterarguments invented from nonexistent facts.

AI as line editor needs a tightly constrained job

Give the editor these constraints:

• preserve facts and numbers
• preserve claim strength
• preserve point of view
• make the smallest edit
• do not add examples
• flag factual concerns separately instead of silently correcting them
• explain material changes when needed

The more open the instruction, such as "make this better", the more likely the editor is to replace the writer.