Make it sound more human: what 400 human texts show about AI writing

Key findings

• Sentence length: in our sample, the median Claude Sonnet 5.5 email averaged 9.00 words per sentence against 12.93 for human emails, about 30% shorter. Only 52% landed in the human range, against 81% of GPT-5.2 emails.
• Rhythm: in our sample, Claude Sonnet 5.5 blog posts had a sentence length standard deviation of 6.88 against 9.13 for human blog posts. Only 65% were in the human range.
• Contractions: in our sample, Claude Sonnet 5.5 blog posts had 1.35 contractions per 100 words against 0.57 in human blog posts, more than twice as many.
• Leftovers: 18 of 50 Claude Sonnet 5.5 emails kept a placeholder in square brackets, against 3 of 48 GPT-5.2 emails.
• Rewrites: on the Claude drafts, Light and Medium changed only 3% to 8% of the words and barely moved the structure numbers.

What we compared

The human texts predate AI writing tools: Enron work emails from 2000–2002 and Common Crawl blog posts from 2019. Each model got the points of one human text and wrote a new text of about the same length, so every AI text has a human counterpart. We quote none of the human text.

What we measured

For human texts we report p10, median and p90. For each AI group we report its median and the share of its texts in the human range, meaning between the human p10 and p90.

Today’s AI text rarely uses the stock AI words

A common first step is to hunt for overused words. In this sample that would have found little. We ran the Less AI rules, a pattern check and not a detector, over every AI text. The median number of flagged sentences was 0 for both the Claude Sonnet 5.5 and the GPT-5.2 blog posts.

The difference is structure

The Claude Sonnet 5.5 emails stand out. Their sentences were about 30% shorter than the human median, and their paragraphs were 15.83 words against 24.00. In blog posts, mean sentence length was mostly in range, but the AI posts were more even from one sentence to the next.

GPT-5.2 emails were close to the human numbers on both counts. Blog paragraph length is not measured, because the blog source drops paragraph breaks.

Contractions are an AI tell, not a human one

The usual advice to make it sound more human is to use contractions. On this data it runs the other way. The human email median was 0.00 per 100 words and the Claude email median 0.60. In blog posts the human median was 0.57 and the Claude median 1.35. This holds for these 400 texts, from one company’s email and from web blogs, and is not a claim about all human writing.

Dashes did not separate the groups: the median was 0.00 everywhere. Our own tool shows how easy it is to overshoot, because an earlier version of its instructions told the model to use contractions where natural.

The old line took the median from 0.60 to 2.00. With it removed, a full rewrite sat at 0.91, against 0.00 in human emails.

Try it on your own text

Placeholders and leftovers

Another trace was the subject line: 47 of 50 Claude emails began with a Subject: line, against 1 of 48 GPT-5.2 emails. That follows from our instruction to write an email. It says nothing about drafts a real user pastes in.

Claude Sonnet 5.5 against GPT-5.2

The two models differed. In our sample, GPT-5.2 emails were close to the human range for sentence length and paragraph length (81% and 88%) and used almost no contractions. Claude Sonnet 5.5 wrote shorter email sentences and more contractions in blog posts. The table lists words each model used more than the human blogs, as a log2 ratio: each step up of one means double the frequency. The Claude words are mostly contractions, and the GPT-5.2 words are plain ones such as meaningful and particularly.

What rewriting changed

Less AI rewrites at three levels, and we measured whether each helps make it sound more human. Light changes only the flagged sentences, Medium reworks paragraphs that have AI patterns, and Strong rewrites every paragraph. Fact drift is an automatic check that includes false alarms, such as a reworded subject line. Claude Sonnet 5.5 drafts first (99 texts).

Light and Medium barely moved the structure. Claude email sentence length stayed at 52% in the human range and reached only 54% after Strong. Strong did help blog posts, lifting the spread from 65% to 78%.

Now the GPT-5.2 drafts (98 texts).

Strong changed far more of the GPT-5.2 texts (25% of the words in emails, 31% in blog posts), with more fact drift and meaning alerts: 7 and 8 of 50 blog posts. The more a rewrite changes, the more you need to check it. Strong also added contractions, from 0.33 to 0.71 in blog posts.

Rule false positives, before and after tuning

We also used the human texts to test our own rules.

Human blog posts dropped from 58 of 200 (29%) to 20 of 200 (10%). The biggest source of early false alarms was dashes: 146 hits, 105 of them en dashes, many inside number ranges. We tuned how en dashes and number ranges count, raised the cluster threshold from 3 to 4, and weakened or removed words human writers also use often. Of the 20 posts still flagged, the categories were AI vocabulary 11, a two-part contrast formula 7, inflated significance 4, empty -ing tail 2, fake depth opener 1 and dashes 1. Across all 400 human texts, none had an AI index above 20, before or after tuning.

How to make it sound more human: a checklist

Each step comes from a result above. None of them guarantees a human-sounding result.

1. Count your sentence length. Claude emails ran about 30% under the human median.
2. Check the spread. The AI blog posts were more even, 6.88 and 6.99 against 9.13.
3. Do not add contractions to sound casual. Human emails had a median of 0.00.
4. Check paragraph length. Claude emails had 15.83 words per paragraph, human emails 24.00.
5. Fill or delete every placeholder. 18 of 50 Claude emails kept one.
6. Strip leftover Markdown and any Subject: line before pasting into an email.
7. Do not expect a light pass to change structure: Light and Medium changed 3% to 8% of Claude words.
8. Check facts after any big rewrite. Strong on GPT-5.2 blog posts showed fact drift in 7 of 50.

Limitations

• The sample is small: 400 human texts and 197 AI texts.
• The Enron emails come from one company in 2000–2002, so they reflect that era and industry.
• Some C4 blog pages come from companies, not individuals.
• Each AI text was generated once. We did not measure variation between runs.
• Paragraph length is reported for emails only, because the blog source drops paragraph breaks.
• The AI texts were written from human points, not from real user drafts.
• Only two models were tested, and we did not test any AI detector.

How to reproduce this

Sample corbt/enron-emails and the English validation split of allenai/c4 with the counts and filters in the first table, then keep 200 texts from each. Give each model the points of one human text, ask for a new text of about the same length, use no system instructions, and run each source once per model. We do not publish the full prompt. Compute the metrics above, take p10, median and p90 from the human texts, and count how many texts in each AI group fall between the human p10 and p90. For rewrites, run each AI text at all three levels and apply the last three metrics in the definitions table.

Questions about AI writing and human writing