Language & writing
AI writing habits: an editing reference
A reference to keep beside a draft: editing principles, original worked examples, and an exercise in making claims precise.
Notes across the full article
Source: Stephanie Zinn, Habits of AI writing, and what to do about them, a16z crypto, August 24, 2026.
- Meaning
- Restate impressive but empty sentences. Check vague contrasts, unjustified hedges, symmetrical lists and stock conclusions. Preserve necessary qualifications. Prefer specific, plain language; collect style examples.
- Voice
- Notice interchangeable warmth, recycled phrasing, bland vocabulary, abstract nouns, vague verbs, grand metaphors and unsupported intensifiers. Choose precise, expressive words. Neutrality can suit support, safety and legal text. Voice notes still need editing.
- Structure
- Check excessive headings, forced threes, essay templates, formulaic openings, needless signposts, abrupt pivots, section previews, repetitive bullets, punchy fragments, repeated conclusions and vague morals. Do not split one idea into artificial takeaways. Keep structures that help readers, especially references. Choose format by purpose; study a model’s paragraph functions before reorganizing.
- Punctuation
- Notice repeated colons, clustered dash asides, distracting parentheses and showy semicolons. Judge correctness, rhythm and distraction, not a banned-mark list. Read aloud. Substitution preserves repetitive sentence structure; preferred-writing examples guide revision better.
These habits are not reliable proof of machine authorship. Effective communication and necessary disclosure matter more than looking human.
Five writing habits, worked through
These examples are invented, not source quotations or announcements about Lexifour.
Importance without information
Draft: “This marks a significant advance in library access.”
Revision, if these are the facts: “The branch now opens on Sundays, 10 a.m.–2 p.m.”
The hours let the reader judge the benefit. If you only know access changed, ask what changed rather than inventing hours.
An announcement before the announcement
Draft: “We are delighted to share exciting news about our timetable.”
Revision: “The evening class moves to Thursdays from 12 November.”
The change and date are useful. Genuine happiness can belong in a personal note; it should not replace the news.
A contrast that stays vague
Draft: “This is not merely a timetable change; it is a new era of possibility.”
Revision, if supported: “The later start lets students attend after work.”
Contrast can help when both sides are clear: “The room changed; the start time did not.”
Uncertainty that hides—or preserves—the claim
Draft: “It could perhaps be said that the entrance is in some respects more accessible.”
Revision, if verified: “The new ramp provides step-free access to the entrance.”
Do not make this a blanket ban on may or probably. A qualifier can accurately describe uncertain evidence. The word-choice section below gives an example.
A closing phrase that adds nothing
Draft: “When all is said and done, return the keys before six.”
Revision: “Return the keys before six.”
But “At the end of the day, lock the gate” can specify a real time. Read the phrase in context before deleting it.
Word choice and a longer worked example
“The revised lesson offers substantial improvements to the learning experience.”
You can finish that sentence without knowing what changed. Was an explanation corrected? Were examples added? Did anything become easier to use? Substantial improvements asks the reader to accept a judgment before providing something to judge.
Suppose the actual change was audio for each example. Then the sentence can say so:
- The claim
- The revised lesson offers substantial improvements to the learning experience.
- The detail
- The revised lesson includes audio for every example.
The second version gives a reader something to check. It is not automatically true because it is more specific. If you do not know what changed, ask the person who revised the lesson. Do not fill the gap with a plausible feature.
Choose the word that fits the evidence
Imagine a sign saying, “The library has expanded its opening hours.” If it now closes at nine rather than six on Thursdays, that is the information a visitor needs. If the writer only knows that the hours have changed, expanded goes further than the evidence allows. Some days might now be shorter.
The same care matters when describing a feeling. Read this invented passage:
Mara sighed when the printer jammed again. She cleared the paper, restarted it, and went back to her desk.
Irritated is a reasonable description of Mara here. Furious would make a stronger claim than the passage supports. Neither is a more sophisticated replacement for angry; each sets a different degree. “Mara seems irritated” also leaves room for the fact that we are interpreting her behaviour.
A qualifier can carry real information. Compare “The letter was written in 1820” with “The letter was probably written in 1820.” Removing probably changes the claim about how certain the date is. Keep it if the evidence is uncertain.
Give the sentence a job
Consider a note to someone borrowing your bike: “Please exercise appropriate caution regarding the braking mechanism.” That is formal, but not helpful enough if the front brake sticks. “The front brake sometimes sticks. Use another bike until it is repaired” tells the reader what is wrong and what to do.
The extra words in the second version are useful. Editing does not always mean shortening. A definition may need an example; a warning may need a reason. Conversely, a sentence announcing that a topic is important may add nothing when the next sentence already shows why.
There is no need to ban a word, a heading, or a punctuation mark. Read the sentence in its setting. What should this particular reader know after reading it?
Further reading: Stephanie Zinn, Habits of AI writing, and what to do about them, a16z crypto, August 24, 2026. Worked examples and exercises are original additions.