AI-Suggested Phrasing Drift

A simple explanation

You write a sentence. The cursor stops. Three suggestions appear. Two are clearly wrong. One is plausible — slightly smoother than what you had, slightly more conventional, slightly less specifically yours. You accept it. The next sentence repeats the pattern. By the end of the email, you have accepted nine small improvements, none of which felt like a meaningful change at the time, all of which together have moved the email an inch away from what you actually wanted to say.

AI-Suggested Phrasing Drift is the accumulation of those inches. Each suggestion is too small to refuse on principle. The aggregate is large enough that, by the third paragraph, the writing has stopped being a transcription of what was in the writer's head and become a negotiation with a model about what should be in it.

An everyday example

You start typing a sentence to a colleague: I'm honestly a bit worried about the timeline. The model offers: I have some concerns about the timeline. You take it. The next sentence — if we don't move the deadline I think the team is going to be wrecked — gets smoothed to if we don't move the deadline I'm concerned about team capacity. You take that too.

Both suggestions are correct in any measurable sense. The email reads as professional. It also reads as something the colleague will not quite be able to hear, because the original sentences carried the actual temperature of how worried you were, and the rewrites are room-temperature. The reply you receive is also room-temperature. The conversation that needed to happen has been gently averted by the helpfulness of the suggestion engine.

Why do I accept so many AI rewrite suggestions?

Because the Reward System is calibrated to take the small improvement. Each suggestion arrives with two qualities the System reads as success: it removes a friction (the cognitive cost of finding the exact phrase) and it inserts a smoothness (the social safety of the conventional phrasing). The System, optimising locally, takes the trade thousands of times before the cumulative effect surfaces as a problem.

The drift is also under-noticed because each individual suggestion is genuinely defensible. The writer can rarely point to the sentence that flattened the email. The flattening is distributed across all of them. The original message exists nowhere in the artefact — neither in the first draft, which was edited, nor in the model, which was the editor, nor in the final draft, which is the negotiated middle.

The behavioral loop

A loop that runs sentence by sentence:

1. Initial sentence — the writer types a phrase that carries the felt specificity of the thing being said.
2. Suggestion arrival — the model offers a smoother, more conventional alternative.
3. Quick comparison — the writer reads both. The conventional version reads as more professional or correct.
4. Acceptance — the suggestion is taken. The original phrasing is lost without a trace.
5. Reward signal — the new sentence reads cleanly. The System flags the trade as a win.
6. Pattern reinforcement — the next sentence begins with the same micro-pause, the same comparison, the same acceptance.
7. Cumulative drift — by the end of the paragraph, the writing has moved measurably toward the model's stylistic centre.
8. Default flip — the writer eventually begins composing in the model's voice from the outset, because the original voice is already being rewritten in their head.

Emotional drivers

Four feelings that drive the loop:

• A small relief each time a suggestion saves the cognitive cost of finding the exact word.
• A faint pride in the polish of the finished text, often misread as having become a better writer.
• An anxiety about sounding unprofessional without the suggestions, which the System uses to justify continued reliance.
• A vague late-stage dissatisfaction with the finished text, which is rarely traced back to its source.

What your nervous system does

The body learns to anticipate the suggestion. Within weeks of heavy use, the micro-pause that precedes a suggestion is sometimes initiated by the writer rather than the cursor — a brief held breath that says let me see what it offers before committing to the original phrasing. The pause is a small parasympathetic suspension, easily mistaken for thoughtfulness.

The cumulative effect on attention is more significant. Each acceptance is also a small abandonment of the work of finding the right word, and the muscle that does that work weakens with disuse. Over months, the writer becomes a person who reaches for the model when the right word does not arrive instantly, where previously they would have sat with the missing word and let it surface. The sitting-with is where most of the writer's voice lives.

The DojoWell interpretation

AI-Suggested Phrasing Drift is the substitution mechanism operating at the granularity of the sentence. The Meaning System's original ask was for language that carries the actual temperature of the writer's experience. The Reward System's substitute is model-preferred phrasing — language that has been smoothed toward a stylistic centre that registers as professional, neutral, and safe. They share a surface property: both produce sentences. They are opposite on the inside.

The deposit is real on the metric the suggestion engine was built to optimise. The text is smoother. Conventional. Free of the small awkwardnesses that signal someone was searching for a word. The residue is the missing specificity — the word that would have carried the writer's actual stance, the rhythm that would have indicated their actual urgency, the joke that would have surfaced their actual relationship to the topic. The residue does not register in the artefact, because it never made it in. It registers in the reply that fails to address what the writer actually meant.

This is why the density signature is false progress rather than residue accumulation. Every measurable surface improves. The trade only becomes visible in the second-order failures — the colleague who does not understand the urgency, the friend who does not register the affection, the reader who finishes the piece without a clear sense of who wrote it. Density is low because the deposit is being made into a stylistic centre and the meaning was distributed across the discarded specificities.

How do I tell when a rewrite improves vs flattens?

You ask whether the rewrite carries the same temperature as the original. If your original sentence had urgency, anger, tenderness, or hesitation in it, the rewrite must carry the same. If the rewrite produces a sentence that any competent professional could have written, it has almost certainly flattened. The improvement that preserves voice is rare and recognisable; the improvement that erases voice is common and easy to take.

The reliable practical test is to read both versions aloud and notice which one a friend who knows you would recognise as yours. The friend is the audit. The model is the suggestion. The judgement is yours.

Practical steps

1. Read both sentences aloud before accepting. Speed alone makes the acceptance unconsidered. Reading aloud reinstalls the choice as a choice.
2. Keep the awkward sentence when it carries the temperature. A clumsy sentence with the right feeling is more readable than a polished sentence with the wrong one. The clumsiness is often what the reader needs to feel that you mean it.
3. Set a quota of accepted suggestions per piece. Five rewrites per email is a different writing posture than fifty. The quota does not need to be principled; it needs to be visible.
4. Write the first draft with suggestions disabled. A document without the suggestion engine has a different texture. The first draft is where most of the voice is decided.
5. Track replies for misreadings. Notice when colleagues, friends, or readers respond as if you had said something slightly different than you meant. The misreadings are evidence of the drift the suggestions hid.

Reflection questions

• Which of your habitual phrasings have you accepted away? Could you reconstruct them if asked?
• When you accept a suggestion, what feeling drives the acceptance — relief, anxiety, pride, or something else?
• Where in the past month did a reader respond to a smoothed version of your message that did not match what you meant?
• If suggestions disappeared tomorrow, what would your first day of writing feel like?

Frequently Asked Questions