What should I tell AI to humanize text?
Tell the AI to vary sentence length, cut hedging and transition words, use concrete specifics instead of abstractions, and write in your voice from a sample you paste in. But know the limit: research shows some models keep producing tells even when explicitly told not to — em dash rates stay as high as 9.1 per 1,000 words under direct suppression.
Why — the first-principles explanation
AI text reads as AI for a mechanical reason: the model picks high-probability words. Given "in today's fast-paced," the likeliest next word is "world." Human writers are noisier — we pick the odd word, run one sentence long and clip the next, drop a detail nobody asked for. Averaged over a paragraph, the model's output is smoother than any real person's. That smoothness is the tell. So "humanizing" isn't adding personality on top; it's adding back the roughness the model optimized away.
The second cause is training, not prediction. Models are heavily exposed to structured, markdown-saturated text, and then tuned by human raters through RLHF who reward writing that feels clear, thorough and well-organized. That preference bakes in a register — signposted, tidy, symmetrical. Research on this describes em-dash-heavy prose as reading as "precise, articulate, and structurally aware," exactly what raters reward. Which is why the fix isn't asking for "a more human tone" and getting the same essay with contractions bolted on. You have to name the specific artifacts.
And there's a hard ceiling. One study measured em dash frequency under explicit instructions to stop, and found suppression resistance varying wildly by model — from 0.0 per 1,000 words in Llama up to 9.1 in GPT-4.1 under suppression. The paper reframes this as a fingerprint of the fine-tuning procedure rather than a style choice you can prompt away. So prompting helps; it does not fully work.
Be careful about why you're doing this. "Make it sound like me" is legitimate editing. "Beat an AI detector" is a different goal with a bad risk profile — detectors are unreliable in both directions, and if you're submitting work under an academic or professional integrity policy, evading detection is the offense, not the detection. Human editors post-editing AI drafts removed only 23% of em dashes present, which tells you something: even people trying to de-AI text leave most of the fingerprint behind.
An example that makes it click
Think of a piano roll versus a pianist. The piano roll hits every note exactly on the beat, exactly the right volume. Technically flawless — and instantly identifiable as not a person, because a human rushes slightly into the loud part and leans on one note too long.
Asking AI to "sound more human" is like asking the piano roll to play with feeling. It'll add a little rubato where it thinks feeling goes — evenly, predictably, in all the expected places. To actually get a pianist, you have to hand it a recording of you playing and say: match this.
How to do it
- Paste 300–500 words of your own real writing and say: "Rewrite the draft below to match the voice, rhythm and vocabulary of this sample. Match my sentence length variation." A sample beats any adjective.
- Ban the specific artifacts by name, not by vibe: "No em dashes. No 'delve', 'moreover', 'furthermore', 'it's important to note', 'in today's world', 'navigate the landscape', 'testament to'. No sentence starting with 'Additionally'."
- Force rhythm: "Vary sentence length deliberately. Follow at least one 25+ word sentence with a sentence under 6 words. Do not make every paragraph three sentences."
- Demand specifics: "Replace every abstract claim with a concrete number, name, date or object. If you don't have one, cut the sentence rather than hedging."
- Kill the scaffolding: "No introduction restating the prompt. No concluding paragraph that summarizes. Start at the first real point and stop at the last one."
- Remove hedges: "Delete 'can be', 'may help', 'often', 'generally' unless the uncertainty is real and load-bearing."
- Then edit it yourself. Add one thing only you know — a specific memory, an internal number, an opinion the model can't have. This is the only step that actually works, because the model has no experiences to draw on.
- If this is for school or work, check the AI policy first. Some allow AI for grammar review only; misrepresenting AI output as your own is defined as fraud under many policies.
Key facts
- Em dash frequency in LLM output ranges from 0.0 per 1,000 words (Llama) to 9.1 (GPT-4.1) even under explicit suppression instructions, per a 2026 arXiv study.
- That study concludes explicit em dash prohibition fails to eliminate the artifact in some models, reframing frequency as a diagnostic of fine-tuning methodology rather than a prompt-fixable style choice.
- RLHF human evaluators reward outputs that are clear, well-structured and thorough, systematically reinforcing the tidy register that reads as "AI writing."
- In a study of human post-editing of LLM drafts, participants removed only 23% of the 254 em dashes present — most AI fingerprints survive human editing.
- Sam Altman publicly acknowledged that em dash frequency in ChatGPT output was adjusted in response to user preference, confirming punctuation-level features are targeted during fine-tuning.
- Common App's fraud policy treats intentionally misrepresenting the substantive content or output of an AI platform as one's own original work as fraud.
▶ The 60-second explainer (script)
Here's what to tell AI to humanize text — and why it only partly works. AI text sounds like AI because the model picks high-probability words. After "in today's fast-paced," the likeliest next word is "world." Real writers are noisier. We pick odd words, run one sentence long and clip the next, drop details nobody asked for. Averaged over a paragraph, the model is smoother than any actual person. That smoothness is the tell. So humanizing isn't adding personality — it's adding back the roughness the model optimized away. So don't say "make it more human." You'll get the same essay with contractions. Name the artifacts. Say: no em dashes. No delve, moreover, furthermore, it's important to note. Vary sentence length — follow a long sentence with one under six words. Replace every abstract claim with a number, a name, a date. Cut the intro that restates the prompt and the conclusion that summarizes. Best single move: paste 300 words of your own writing and say match this voice. A sample beats any adjective. Now the honest part. A 2026 study measured em dash rates when models were explicitly told to stop. Llama dropped to zero. GPT-4.1 stayed at 9.1 per thousand words — under suppression. The authors call it a fingerprint of the fine-tuning, not a style you can prompt away. The only step that really works: add something only you know. The model has no experiences. That's the gap it can't close. And if this is for school or an application, read the AI policy first. Misrepresenting AI output as your own is defined as fraud in many of them.
What authoritative sources say
People also ask
What's the single best prompt to humanize AI text?
Paste 300–500 words of your own real writing and ask the model to match its voice, rhythm and vocabulary. A concrete sample outperforms any list of adjectives, because "human" isn't a style — your writing is.
Will humanizing prompts beat AI detectors?
Sometimes, unreliably, and it's the wrong goal. Detectors produce both false positives and false negatives, so you can't verify success. More importantly, if you're under an academic or workplace integrity policy, evading detection is the violation — not being detected.
Why can't AI stop using em dashes when I tell it to?
Research measured exactly this. Suppression works in some models and fails in others — GPT-4.1 still produced 9.1 em dashes per 1,000 words while instructed not to. The behavior comes from fine-tuning, deeper than the prompt can reach.
Are "AI humanizer" tools worth paying for?
They mostly do word swaps and syntax shuffling, which degrades clarity while leaving deeper fingerprints — rhythm, structure, absence of specifics. Rewriting a few sentences yourself and adding one thing only you know does more, for free.
What actually makes writing sound human?
Specifics the model can't have: a number from your own records, a name, a date, a memory, an opinion with a reason behind it. AI can imitate rhythm. It cannot imitate having been there.