The /humanize Command — Remove AI Fingerprints From Text
How /humanize scans for AI writing patterns across vocabulary, structure, and rhythm, then surgically removes them while keeping the author's message and voice intact.
AI-generated text has a detectable signature. The vocabulary clusters around the same dozen words — "delve," "robust," "seamless," "leverage," "tapestry." The structure repeats: three parallel bullet points, sentences that all open the same way, transitions that announce themselves. /humanize is designed to remove that signature without rewriting the content underneath it.
The command takes a file path as its argument. Before making any change, it saves a timestamped backup next to the original file and creates a working copy at a temporary path.
What the scan catches
The scan runs against three categories of patterns. The first is banned vocabulary — a specific list of words and phrases that appear at high frequency in AI-generated text and rarely in natural human writing. English and Russian each have their own lists. In English: words like "delve," "tapestry," "testament," "meticulous," "pivotal," "leverage," "utilize," "robust," "seamless," "comprehensive," "cutting-edge," "empower," "elevate," "streamline." In Russian: equivalent patterns including overly formal phrasing, textbook tone, and generic filler that adds length without adding meaning.
The second category is structural tells: several consecutive sentences with the same subject or opening word, over-use of list formatting where prose would work better, and formatting artifacts like excessive bold text or headers on short sections that don't need them.
The third category is communication artifacts: sycophantic openers ("Certainly!", "Great question!"), knowledge-cutoff disclaimers woven into the prose, and restated conclusions that summarize what was just said rather than advancing the argument.
How fixes are applied
Changes go in three passes to prevent fixes from introducing new problems. The first pass handles vocabulary and formatting: banned words are replaced with natural equivalents, formatting artifacts are removed. The second pass handles structure and rhythm: consecutive same-subject sentences are varied, list blocks that work better as prose are converted. The third pass is an anti-AI audit of the result: the edited text is re-scanned to confirm no new patterns were introduced.
The core constraint throughout is surgical editing, not rewriting. The author's structure, argument, and intent are preserved. Only the patterns that mark the text as AI-generated are removed. If a sentence has an AI-pattern word but the surrounding paragraph reads naturally, the word gets replaced, not the paragraph.
Output
The scan produces a diagnostic report with counts by category before changes are applied. After the three-pass edit, the corrected file is saved at the original path. The backup remains in place. The report shows what changed and why — replacements are listed with before/after so the author can review any correction that feels wrong.
When to use it
/humanize is the single-purpose version of Phase 2 inside /dr-edit. Use it when a draft needs AI pattern removal but the facts have already been checked and structure is sound. It is also useful as a final pass on human-written content that unintentionally picked up AI-style phrasing during revision.
For fact verification without pattern removal, see /factcheck. For the full three-phase editorial pipeline, see /dr-edit. For context on the content workflow these commands support, read what Datarim is.