Short answer
Yes. While improving tone and structure, AI can change certainty, responsibility, timing, scope, negation, conditions and technical terms. The most dangerous shift is rarely a complete reversal. It is a smoother but more absolute sentence: “we plan to evaluate” becomes “we will implement”; “may be associated” becomes “causes”; or “I do not agree on the current evidence” becomes “I support further discussion.” Material writing needs a meaning comparison, not only a grammar check.
Rewriting is never just swapping synonyms
Real rewriting chooses what to preserve, what to delete, how to organise causation, who appears as the subject, and how strongly to speak. Given vague instructions such as “more professional,” “shorter” or “friendlier,” a model optimises reader experience without knowing which awkward details embody commitments the author cannot surrender.
Awkwardness can reflect genuine uncertainty, a boundary or relational tension. After a model turns it into standard business prose, the text may look better while taking a different position. Language quality and semantic fidelity are separate measures.
The problem resembles factual consistency in summarisation: generated prose can be highly fluent while containing propositions unsupported by its source. Evaluation research therefore treats source consistency as a separate dimension rather than assuming readability means correctness. Evaluating Factual Consistency of Summaries
Certainty is easily increased without notice
“May,” “tends to,” “on the current information” and “has not been ruled out” express evidence boundaries. A model seeking confident prose can remove them. In the other direction, a request to soften tone can turn a clear refusal into a reservation, leaving a recipient to believe negotiation remains open.
Compare “We expect to finish in June” with “We will finish in June”; “This factor is associated with the outcome” with “This factor caused the outcome”; or “Preliminary data support a trial” with “The data prove the solution works.” Few words change, but commitment, causation and evidential strength do.
During review, mark modal and evidence verbs: may, should, must, plan, decide, observe, estimate and prove. Asking the model to list every change to these terms is more testable than asking whether it “kept the meaning.”
Responsibility can disappear or move
“The supplier must repair the fault within five days” becomes “The fault should be repaired within five days,” and the actor vanishes. “I approved the exception” becomes “The exception was approved,” turning a personal decision into an agentless fact. “Our team did not receive the data” becomes “The data were unavailable,” obscuring who failed to supply them.
Passive voice is not inherently wrong. In a contract, incident report, decision record or action list, however, who does what by when is material. Instruct the rewrite to preserve actors, obligations, deadlines and approval states. If politeness genuinely calls for obscuring an actor, that should be an author's conscious choice rather than an automatic edit.
Inspect pronouns as well. In a long exchange, “they,” “that team” or “the client” may have several plausible referents. A model can select one to make the paragraph cohere without warning that it resolved an ambiguity.
Scope and quantifiers change the boundary of a claim
“Some users,” “in two pilot regions,” “mainly for new accounts” and “data to March” define where a statement applies. Removing them during simplification makes a local observation appear universal. Usually, at least, no more than, approximately, only, and including but not limited to are not interchangeable.
Numerical reframing needs care. Describing a rise from five per cent to six per cent as “a twenty per cent increase” can be mathematically valid while creating a different reader impression. “Within ten days” may resemble “a maximum of ten days” but omit the starting event. Turning the midpoint of a range into an expected value adds precision absent from the source.
For external disclosures, policies, research and commercial commitments, lock numbers, units, comparison bases, dates and scope words unless the author approves each change.
Negation and exceptions disappear during smoothing
Negation is small in length and large in effect. “We found no evidence of fraud” does not prove no fraud occurred. “Do not share unless the customer consents” should not become only “It may be shared with consent” if the default prohibition matters. “Not all requests require approval” is different from “Requests do not require approval.”
Models can simplify the logic of nested conditions and double negatives. For legal, safety and policy text, first represent the logic as rule, condition, exception and outcome, then improve the prose. Test boundary scenarios afterwards: What happens when the condition holds, when it does not, and when an exception conflicts with the general rule?
Do not rely only on another request to “confirm that the meaning is unchanged.” A model may remain confident in its own rewrite. Side-by-side source comparison against explicit dimensions is more reliable.
“Friendlier” can send a different relational signal
Tone is cultural, contextual and connected to power. Adding thanks and cushioning to direct feedback can make a required action seem optional. Turning a cautious complaint into forceful legal language can escalate a relationship. Standardising a non-native writer's distinctive expression can remove personal voice.
A model may introduce commitments you would not make: “We take this very seriously,” “We will ensure this never happens again,” or “Please contact me at any time.” They sound courteous but can imply admission, guarantee or continuing availability.
State that a rewrite must not add promises, apologies, legal positions or emotion. For sensitive correspondence, give precise goals: “retain the same level of directness,” “make it warmer without changing the request,” or “shorten it while preserving every qualification.”
Technical terms are not ordinary synonyms
Near-synonyms in law, medicine, engineering, finance and policy may have different definitions. Risk and hazard, rescission and termination, revenue and cash flow, validation and verification cannot always be exchanged. A model may choose familiar words to improve accessibility while destroying technical precision.
Maintain a protected glossary of definitions, abbreviations, product names, regulatory language and internal roles. The model can improve the surrounding sentences but must retain those terms. If a general reader needs help, add an explanation after the term rather than silently replacing it.
Be especially cautious when rewriting a translation. Every transformation introduces another opportunity for drift. Material text should be checked against the original language, not only against the first translated draft.
AI may add “helpful” content that the author never supplied
Asked to make an argument more complete, a model can insert examples, reasons, data or solutions. They may fit common knowledge while failing to represent the author's view or verified fact. Treat permission to edit expression and permission to extend content as different authorities.
A rewrite instruction can prohibit new facts, sources, commitments, causal claims and conclusions, and ask for gaps to be marked as questions or comments. If expansion is desired, visibly label additions and verify them separately.
NIST's Generative AI Profile discusses confabulation, over-reliance and information-integrity risks. Plausible generated material adopted without sufficient checking is exactly the type of workflow risk that calls for process controls. NIST AI 600-1: Generative AI Profile
Review material edits with a semantic-difference table
A quick read may suffice for an ordinary social message. For contracts, customer commitments, board papers, policy, performance feedback, research and public statements, compare: source proposition, rewritten proposition, actor, action, time, scope, certainty, condition, exception, number and newly added content.
Use a tool to display textual differences, ask the model to classify changes against those fields, and then let the author approve. The model's explanation helps locate issues; it is not final proof. Return to the source sentence whenever a critical field differs.
Reverse explanation is also useful. From the rewritten text alone, ask, “Who must do what, by when, under which conditions, and how strong is the evidence?” Compare those answers with the author's answers for the original. This reveals drift more effectively than “Do these mean the same thing?”
Preserve the author's version and edit trail
If only the final AI version survives, it may be impossible to tell whether a commitment came from the author or the model. Preserve source, instruction, output, human edits and final approval. Not every casual message needs a complete audit trail, but text affecting rights, money, reputation or an official record should be traceable.
Version history also protects voice. A team can identify patterns where the model repeatedly overstates or weakens particular language, then build a do-not-change list and better prompts.
Do not make “accept all changes” the default. Show differences first, highlighting deletion, numbers, negation, dates and obligation words. Interface design directly affects whether semantic error becomes visible.
Scale review to consequence
For internal brainstorming, personal notes and easily withdrawn informal text, prioritising clarity is reasonable. Customer email, project updates and performance feedback need checks of promises, tone, actors and deadlines. Contracts, legal advice, clinical explanations, financial disclosures, safety procedures and formal decisions need line-by-line comparison by a competent person; AI may need to be limited to formatting or error flags.
A short sentence is not necessarily low risk. “We accept full responsibility” can matter more than ten pages of background. Classify by the action and consequence a sentence can trigger, not its length.
Australian Government AI adoption guidance emphasises context-specific impact assessment, accountability, testing and monitoring. Rewriting is a use case in its own right; it should not escape risk analysis because it resembles ordinary word processing. Australian Government: Guidance for AI Adoption—Foundations
A safer rewrite instruction
State the intended reader and purpose. List what must remain: facts, numbers, actors, obligations, negation, conditions, exceptions, certainty and terminology. List what must not be added: promises, apologies, reasons, sources and actions. Constrain the edit, perhaps to grammar and paragraph order. Request both a revised version and a list of material semantic changes.
If the source is ambiguous, require a question rather than a silent choice. For high-risk material, work one section at a time and then recheck cross-section references and definitions after assembly.
My assessment: the author remains responsible for the proposition
AI can be an excellent editor, but it does not bear the relational or legal consequence of an email, contract or public statement. An author or approver cannot delegate responsibility to a polishing button. Authorial control does not require personally typing every word. It requires knowing which dimensions of meaning cannot change and choosing changes consciously.
If you cannot look at the final text and answer, “This is my actual position, and I accept the consequences of a reader understanding it this way,” it is not ready to send.
Meaning checklist
- Are actors and responsibilities preserved, or replaced by agentless passive language?
- Have may, should, must, plan, decide and complete been exchanged?
- Are negation, conditions, exceptions, scope and time limits intact?
- Are numbers, units, benchmarks, dates and technical terms unchanged?
- Did the rewrite add facts, causation, promises, apologies or recommendations?
- Did tone turn a requirement into an option or a cautious view into an accusation?
- Was ambiguity exposed rather than silently resolved?
- For material text, are source, differences and final approval preserved?
Conclusion
AI rewriting can quietly change meaning because clarity, brevity, professionalism and warmth all require semantic choices. Protect certainty, actors, duties, scope, negation, conditions, numbers and terminology. Treat a rewrite as a set of editorial decisions that need review, and compare semantic fields rather than style alone. Then AI can improve expression without changing your position on your behalf.
Related questions
- How Can You Separate Facts, Inferences and Advice in an AI Answer?
- How Should You Preserve Sources, Versions and Edits in AI-Assisted Work?
Continue reading: All articles in How Far Should You Trust AI?
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.