Use AI as a Critic Without Letting It Rewrite You

An AI can improve a draft by telling you what is unclear, unsupported, repetitive or easy to misread. It can also replace the draft with a polished average of patterns it has seen before. Those are not the same service.

The critic protocol keeps generation downstream of authorship. You establish the claim, structure, evidence and voice. The model receives a narrow review job. Its output becomes a set of suggestions to test, not prose that silently becomes yours.

Critique preserves a productive gap

A rewrite collapses diagnosis and solution. If the new paragraph sounds fluent, the writer may accept it without understanding what was wrong, which tradeoff changed or whether the replacement remains accurate. A critique preserves the gap where judgment happens.

This matters for voice. Voice is not a garnish made of unusual adjectives. It is a pattern of selection: what you notice, which claims you refuse, how much uncertainty you keep, where you slow down and which consequences you make visible. A model can imitate surface style while quietly changing those decisions.

One 2024 experiment by Anil Doshi and Oliver Hauser found that access to generative-AI ideas increased average evaluations of individual short stories in its study while making AI-assisted stories more similar to one another. That result is not a universal law of writing. It is a useful warning about a plausible tradeoff: individual improvement under one measure can coexist with reduced collective diversity.

Pass 0: decide what may leave your boundary

Before pasting, classify the draft. Does it contain private correspondence, customer data, unreleased strategy, medical or legal information, credentials, identifying details, copyrighted material you lack permission to upload or another person's confidential story? If yes, stop or use an approved environment and policy.

The AI data-boundary guide provides the decision record. Redaction is not magic. A distinctive timeline, quotation or combination of facts can identify a person even after names are removed.

Pass 1: freeze the human baseline

1. Write the claim in one sentence.
2. Name the intended reader and what they should be able to do or decide afterward.
3. List the sources that carry material factual claims.
4. Mark three non-negotiables: a distinction, phrase, example or uncertainty you do not want smoothed away.
5. Save a version before asking the model anything.

If you cannot state the claim or audience, the model will infer them from statistical patterns and prompt phrasing. That can produce a competent document aimed at the wrong job. Use a decision record when the stakes warrant a durable note about model, date, materials and accepted changes.

Pass 2: request diagnosis in separate rounds

Do not ask for “feedback” and accept whatever the model decides feedback means. Run four narrow rounds:

Separating rounds makes failure legible. If a single response mixes fact-checking, style, argument and rewriting, a correct observation can lend borrowed confidence to an invented source or harmful edit.

Use the reusable critic packet

Then add the draft below the contract. For a long document, review one section at a time while preserving the claim and outline at the top. Models have context limits and attention is uneven across long inputs. A section-level review reduces silent omission and makes each comment traceable.

Pass 3: adjudicate every suggestion

Create an accept, reject or test column. For each suggestion, ask:

• Does the quoted problem exist in the saved baseline?
• Would the change improve the stated reader job?
• Does it preserve the claim's evidence level?
• Is the model asking for a conventional phrase merely because it is common?
• Does it remove productive friction, uncertainty or specificity?
• Can the factual suggestion be verified in the underlying source?

Write the revision yourself when practical. If you request options, constrain them to the diagnosed problem and compare each against the baseline. Never treat a source title, quotation, statistic or legal requirement generated by the model as verified. Use the citation-verification protocol before publication.

Run a voice diff, not a vibe check

Compare the final version with the frozen baseline. Highlight sentences whose subject, certainty, example or conclusion changed. Counting adjectives will not reveal a shifted argument. Ask instead:

• Did “may” become “will”?
• Did a concrete actor become passive voice?
• Did an uncomfortable exception disappear?
• Did the model introduce a symmetrical structure the evidence does not support?
• Did a personal or local observation become generic advice?
• Can the author explain why every material change was accepted?

A useful critic leaves the document more defensible and the author more aware of its weak points. A ghostwriter can leave the document smoother while the author knows less about what it now says.

Choose the right critic mode

Objection mode fits an argument. Ask for the strongest charitable counterargument and the evidence that would decide between positions. Ambiguity mode fits instructions. Ask where two competent readers could take different actions. Failure mode fits a plan. Ask what dependency, incentive or edge case could defeat it. Compression mode fits a long draft. Ask which sections do the same job and what distinction would be lost by merging them.

Do not use one model response as a panel of independent critics. Variations from the same system may share training patterns, prompt anchoring and blind spots. For consequential work, involve a qualified human with the relevant domain context.

Define a stop rule

Stop when remaining comments are preference changes rather than reader failures, when revisions begin to cycle, or when the model repeatedly proposes language that restores passages you intentionally removed. More rounds do not guarantee more quality. They can converge the draft toward generic fluency.

Use the site's stop-rule guide: define the acceptance test before starting. For this protocol, a workable test is that the claim is explicit, material statements are supported, the intended reader can follow the structure and every accepted AI suggestion has a human reason.

When to skip AI critique

• The material cannot be sent to the available system under policy or consent.
• The work depends on lived, cultural or domain context the system is likely to flatten.
• You need legal, medical, safety or professional review rather than plausible objections.
• You have not produced a baseline and are using feedback to avoid deciding what you mean.
• The document is short enough that a trusted human reader can provide the missing context more directly.

Claims and boundaries

Sourced fact: NIST's generative-AI profile treats confabulation, data privacy, information integrity and human-AI configuration as risk-management concerns; one published experiment found a tradeoff between evaluated individual creativity and collective similarity in a bounded story task. Inference: restricting the model to diagnosis can preserve more human judgment than accepting full rewrites. Judgment: the author should freeze a baseline, verify claims and record material decisions. Not claimed: AI critique always improves writing, one protocol eliminates model error, AI-assisted prose is necessarily generic or the cited experiment generalizes to every genre.

Primary and official sources

• NIST AI 600-1: Generative Artificial Intelligence Profile, for generative-AI risks and risk-management actions.
• Doshi and Hauser: Generative AI enhances individual creativity but reduces collective diversity, for the bounded writing experiment and its limitations.
• Dryad research data for the Doshi and Hauser study, for inspectable study materials and data.
• How to Compare Two AI Answers, for evaluating outputs by task fit, claim support and failure cost.