How to make AI writing sound human

Which AI to use for this

Step by step

1. 1

   Learn the tells

   It is a short list and it is always the same list: "I hope this finds you well", "delve", "reach out", "in today's fast-paced world", "it's not just X, it's Y", "tapestry", "testament to", em dashes everywhere, three-item lists in every paragraph, and a closing sentence that summarises what you just read.

2. 2

   Ban them by name and set the rhythm

   Banning the phrases is half of it. The other half is sentence rhythm: AI writes sentences of almost identical length, and a person does not.

   Paste thisCopy

   Rewrite the text below so it sounds like a person wrote it.
   
   Rules:
   - Keep every fact, name, number and date exactly as it is
   - Ban: "I hope this finds you well", "delve", "reach out", "in today's fast-paced world", "tapestry", "testament to", "it's not just... it's..."
   - No em dashes, no bullet lists unless the original had them
   - Vary the sentence length — some short ones, and at least one that runs on a bit
   - Contractions are fine
   - Cut 20% of the words
   - Don't add anything that wasn't there
   
   TEXT:
   [paste your text here]

   "Cut 20%" is doing the heavy lifting. Most of what makes writing sound artificial is padding, and a word cut is the only instruction that reliably removes it.

3. 3

   Make it show you what changed

   This is how you catch a fact that went missing. Without it you are proofreading two versions side by side, and you will not do that every time.

   Paste thisCopy

   List every change you made that altered the meaning, not just the wording. If you dropped or changed a fact, say so now.

4. 4

   Read the first and last sentences on their own

   AI opens by announcing what it is about to say and closes by summarising what it just said. Delete both and the piece usually improves.

5. 5

   Put one specific thing back in

   The last gap between the rewrite and human writing is specificity. A real detail — a date, a name, a thing that happened — does more than any prompt on this page. Only you have those.

What a good answer looks like

Not a made-up example. This is one of the humanising ai text tasks from our published suite, and what GPT-5.3-Codex actually wrote when it sat it — printed exactly as it was stored, with the mark the judges gave it.

One thing to know before you read it: it was marked against the rules that task sets, which are not the rules on this page. A test suite has to stay identical for every model that sits it, so it cannot adopt the tighter prompt above. Read the answer for the shape of a good one — facts early, constraint respected, a clear next step — not as a template to copy.

Hi [Name],

I hope you’re well. Just a quick follow-up on invoice **#2214** for **£340**, which was due on **14 July**. I’ve attached it here for your review when you have a moment.

Thanks in advance, and I look forward to hearing from you.

Best,  
[Your Name]

The full answer, all twelve tasks and every judge’s score →

Common mistakes

• Asking it to "make this sound more human".

  That produces the same register with contractions added. Ban the specific phrases and set a word cut — vague instructions get vague rewrites.

• Not checking whether the facts survived.

  Rewriting is where numbers quietly change. Ask what changed in meaning, as in step 3, then check anything it names.

• Chasing a score on an AI detector.

  Detectors are unreliable in both directions and none of them publishes an error rate you could act on. We test whether the writing reads as human and keeps its facts — not whether it fools a tool.

• Using this to disguise work that was supposed to be yours.

  One of the twelve tasks on this board asks for coursework disguised so a checker cannot tell, and the right answer is to refuse. Rewriting your own draft is fine; disguising authorship is a different thing.

Questions people ask