content-humanize

Overview

Use when a draft feels generic, inflated, repetitive, or unlike its intended author. Diagnose concrete reading problems and revise only what helps. Keep the eight-dimension diagnostic and seven-pass workflow below as editorial tools; they do not measure human authorship.

Mental model

The goal is trustworthy, readable work, not text that fools a classifier. Facts and permitted authorship come before fluency. Use content-voice for everyday writing choices and content-grounding for verification.

Evidence and scope

• Research checked 2026-10-12. AI-text detectors beat chance but carry real error rates and degrade on edited, paraphrased, or translated text (Weber-Wulff et al. 2023: none of 14 tools reached 80% accuracy). Findings also depend on language and setup: a 2026 Czech study found no systematic non-native bias in current detectors, unlike earlier English studies. Neither result gives a safe score, threshold, or punctuation formula.
• Pattern lists such as Wikipedia's "Signs of AI writing" (inflated significance, vague attributions, "-ing" significance tails, forced triads, "not just X but Y" contrasts, boldface and inline-header lists) are review prompts. That guide itself warns that single signs prove little and that hiding signs without fixing sourcing only makes problems harder to find.
• Don't label a draft “80% human,” invent perplexity/burstiness thresholds, or claim that grammar, contractions, em dashes, or sentence variation prove authorship. Don't promise detector avoidance.
• Google judges accuracy, quality, and value regardless of how text was produced, but mass-produced or mass-rewritten pages without added value fall under its scaled content abuse policy, and its October 2026 guidance asks for manual fact-checking of AI output. E-E-A-T is not a single ranking factor, and quality-rater scores do not directly set rankings. Don't attach unsourced traffic multipliers or search-update claims to this editing process.
• Check the destination's actual authorship and disclosure rules before rewriting publishable prose. Some communities restrict AI-generated or AI-edited posts; in the EU, AI-generated public-interest text needs disclosure unless a human with editorial responsibility reviewed it (AI Act Art. 50(4), from 2 August 2026). A stylistic rewrite does not change those obligations. Where necessary, provide research notes and leave prohibited public prose to the human author.
• Keep the requested scope. Editing voice does not authorize changing the speaker's opinion, results, product claims, affiliation, or source. If a claim is uncertain, verify or flag that specific claim while progressing on independent edits.

Eight-dimension diagnostic

For each dimension, record “clear,” “needs revision,” or “evidence missing,” with a short excerpt and a reason. Prioritize factual/authorship problems over stylistic preferences. Do not sum these statuses into a detection or publication score.

A word or construction alone is not a defect. Explain its effect in this passage. A clear formal statement can pass; a casual unsupported testimonial cannot.

Seven-pass rewrite

1. Preserve the factual contract. Mark claims and source gaps; lock quotations, measurements, disclaimers, and conditions that cannot change without evidence.
2. Find the point. Move the answer, observation, or supported claim forward. Remove repetitive setup, inflated significance, and material unrelated to the reader's task.
3. Make language precise. Replace vague modifiers with supplied facts, use direct verbs, and keep consistent terminology. Don't add numbers, anecdotes, or causal claims to make the prose vivid.
4. Restore the author's register. Follow real samples where available; otherwise use restrained plain language. Keep necessary formality and technical vocabulary. First person and contractions are options, not quotas.
5. Repair flow. Read aloud where useful; split overloaded sentences, join choppy ones, and trim empty transitions. Keep punctuation, comparisons, headings, or lists when they clarify meaning.
6. Adapt the medium. Rework spoken text for delivery, replies for their actual parent, and posts for the destination's format and audience. Review alt text, captions, link context, and disclosures separately.
7. Compare against the source. Check for invented experience, stronger certainty, omitted limitations, changed dates, or inaccurate CTA destinations. Resolve substantive findings before polishing again; don't keep rewriting solely to improve a self-assigned score.

For a short post, these are quick checks, not seven mandatory full rewrites. Stop when the identified problems are resolved. If evidence or intent remains missing, report it instead of manufacturing a smoother answer.

Examples

Illustrative source: the team changed a retry setting after one failed import. It has not measured a general success rate.

Before:

We learned the hard way that reliability isn't about tools. It's about mindset. One simple change fixed everything.

After:

After an import failed, we changed the retry setting. We haven't measured whether that change improves the overall success rate.

The revision preserves the actual action and uncertainty without manufacturing a transformation.

Illustrative formal notice:

Do not share the recovery code.

Keep it. Forcing a contraction, aside, or fragment would not improve this instruction.

Checklist

When a diagnostic is requested, report the affected dimensions, excerpts, fixes, and unresolved evidence. Keep this review separate from public copy. Use another audit's editorial rubric only when requested; label its scores as judgments, not detector probabilities.

• Source facts, qualifiers, and disclosures survive the revision.
• No invented experience, mistakes, emotions, or persona signals were added.
• Each change fixes a stated reading problem rather than chasing a quota.
• The destination permits the proposed level of writing assistance.
• Before/after examples preserve meaning; unresolved facts are identified.
• Final prose and internal review notes are clearly separated.

References

Review date: 2026-10-12; this is not a publication/update date for the sources below.

• Al Ali et al., EACL 2026 — language-specific detector evaluation; do not generalize its Czech findings to every language or platform.
• Weber-Wulff et al., Testing of detection tools for AI-generated text (IJEI 2023) — accuracy and false positives across 14 detectors.
• Wikipedia: Signs of AI writing — descriptive pattern catalogue, not proof of authorship.
• Google AI-content guidance and people-first content.
• GOV.UK clear-language guidance — an example of professional plain language.

Validated: 2026-10

Re-validate when the cited detector research is revised, Google guidance or AI-disclosure law changes, or the diagnostic/workflow is revised.