How to make AI sound more human: what held up for us after 1,000+ hours on our writing skill and hundreds of blind A/B tests
retained text
**TL;DR**: Word-level fixes help a little. The biggest change for us came from giving Claude a real person's point of view, from a transcript of them talking about that topic. That last part comes from my own testing, since we found no published study on it. I run an SEO and GEO agency where we write our articles with Claude. I've put well over 1,000 hours into our writing skill over the last three years and graded hundreds of A/B tests, two versions of the same article without knowing which was which. A lot of you have your own setups, so take this as one data point. I also had Claude spend four or five hours finding every study and test on this. Some of what I believed didn't hold up. **Word lists and rules help less than I thought**. At one point we had about a hundred writing rules. Claude would apply those rules very diligently, but it still is missing that personality or that soul that comes from human writing. More rules made it worse. A few client accounts where we could never get natural drafts each had about 80 rules in their guidelines, mostly Claude overcorrecting after a client hated one sentence. When we cut the rules down and required every article to take a position, the drafts started sounding more like a real person at the brand. Several things changed at once, so I can't give the rule cut all the credit. **"Sound more human" barely does anything**. A 2026 study of 11 models found the main differences between AI and human text held up even under prompts written to get human-like text. In my tests, "be more conversational" and "write more clearly" haven't done anything either. **Examples plateau fast**. In Wang et al. (Findings of EMNLP 2025, 100 everyday writers), an authorship check accepted AI blog posts as the author's own 8-10% of the time with no examples and 17-21% with five of their posts. Going from 2 examples to 10 changed little. Examples on the same topic as the new post did worse than random ones, 10% against 19%. The first time we gave it examples that were all paragraphs, it stopped using bullets and tables. **A transcript of the person talking has done the most for us**. Every post on my blog starts from a transcript of one of my videos, and those posts kept sounding better than anything else we made. I assumed Claude was copying my voice, and I spent a lot of time trying to capture each client's voice the same way, which never worked very well. In October I looked closer and saw it rarely uses my wording. What it picks up is my opinions, the fact that I hedge, the fact that I go back and forth between two ideas. **What we do now:** 1. Get 2-5 minutes of the person talking with no script. Six studies of people sharing opinions found speakers used more emotional language than writers, and the gap disappeared when speakers were told to plan first. Written Q&As don't work nearly as well for us. 2. Take out false starts and leave everything else, including the "I think"s. 3. Have Claude quote the 5-10 lines that carry their opinions, doubts and stories, then draft around those. 4. Check the draft for certainty it added. If they said "usually", it shouldn't say "always". Don't use an AI judge to pick the better draft. Three models matched professional copy-editors at about chance on what counted as slop. Frequent AI users are the better test: five of them voting together caught 99.3% of AI articles with no false flags, while people who rarely use AI were close to guessing (Russell et al., ACL 2025, on 2024 models). What I still can't fix is theory of mind. Claude over-explains the obvious and then breezes past something complex in one short sentence, and no skill or guidance has solved that for us. I expect smarter models to do most of that work. I wrote up all of it with the sources, the two prompts we use and a table of every method we graded. Happy to put the link in a comment if people want it.
document fetched 2026-11-04 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xmea8t/how_to_make_ai_sound_more_human_what_held_up_for