How to make ChatGPT not sound like AI

Works for ChatGPT's custom instructions; the same rule file works in Gemini and Claude. Updated August 2026.

ChatGPT's writing has a recognizable accent, and readers have learned it: the "it's not just X, it's Y" sentence found in roughly 6 percent of real messages, the em dashes, the bold-everything formatting, and a vocabulary of words like "delve" and "seamless" that run at many times their human frequency. You can remove most of the accent in five minutes of setup, and the good news from the data is that instructions actually work now: since a fix OpenAI shipped in late 2025, custom instructions are respected far more reliably than they used to be.

Step one: install standing instructions

Open ChatGPT, then Settings, then Personalization, then Custom Instructions. In the box that asks how ChatGPT should respond, paste a rule file. You can write your own, or take ours: the HumanSounding instructions file is free, fits the box, and encodes the measured tells rather than folklore. It also tells the model to consult our weekly-updated trending list when it can browse, so the rules stay current as models change.

Step two: know the five rules doing the heavy lifting

If you'd rather keep instructions short, these five earn their space. Ban the contrast reflex: no "it's not just X, it's Y" in any variation; state what the thing is. Cap em dashes at one per 300 words. Ban the top slop words: delve, seamless, leverage, meticulous, pivotal, transformative, tapestry, underscore, showcase, and their friends; require the plain alternative. Forbid summary endings; the draft ends on its last new fact. And demand varied rhythm: mixed sentence lengths, an occasional one-line paragraph, because uniform cadence is a measured fingerprint that survives every word-level fix.

Step three: prompt for specifics, not style

Settings fix the mechanical accent. What they can't supply is content worth reading, and vague drafts smell like AI even with clean vocabulary. The single most effective prompt habit is feeding the model concrete facts to build with: numbers, names, dates, the thing that actually happened. "Write a status update saying the invoice fix shipped Tuesday and refunds are back under ten minutes" produces human-sounding text almost automatically. "Write a status update about our progress" produces the vernacular, because vagueness is where the vernacular lives.

Step four: check the result before sending

Even with instructions installed, drafts drift. Run anything important through a tell check: our free checker highlights every recognizable pattern with a suggested fix, entirely in your browser, and its one-click mechanical fixes handle the word swaps and dash surgery for you. What it flags but can't fix, the structural habits, take a two-minute human pass.

An honest closing note

This setup makes text read naturally to humans, which is the goal that matters. It does not make text undetectable to stylometric analysis; research classifiers identify model output at high accuracy even after paraphrasing, and any tool promising otherwise is overselling. And where AI disclosure is required, in classrooms, journals, and some workplaces, disguising authorship is the wrong move entirely. Sound like yourself; don't pretend the tool wasn't there.

Sources: Washington Post message dataset (via Decrypt) · TechCrunch on the custom-instructions fix · GPTZero AI vocabulary · Wikipedia: Signs of AI writing · Sun et al., CMU stylometry

Part of HumanSounding, the live index of AI writing tells. Figures quoted from their sources without adjustment; the trend data behind this site updates weekly.