Why does AI writing all sound the same?
AI writing style: a chatbot picks likely next words, and RLHF rewards a safe, polished register, so the same words win in answer after answer · Beginner · 5 min · sampling, training, llm · after Next-token prediction, RLHF
AI writing sounds alike because a chatbot picks likely next words and RLHF rewards a safe, polished register, so words like “delves” and “crucial” pile up. Temperature barely helps; your own sample, specifics and edits do.
1 Guess first
2 Look
The words AI writing leans on
1 / 42024 use vs expected from earlier years
delves28×underscores13.8×showcasing10.7×expected1×Over 15 million PubMed abstracts, 2010–2024 (Kobak et al., 2024)
In 2024 PubMed abstracts, “delves” turned up 28 times as often as its earlier trend predicted. Rare words jumped first.
The same few words surge wherever chatbots write, because training narrowed the model’s likely choices to one polished register.
A chatbot writes by picking a likely next word, again and again. Then RLHF trains it on the replies people rated higher, and raters tend to prefer familiar, polished text. So the safe register wins: “crucial”, “delve”, em dashes, tidy lists of three. Chatbots learn from similar text and similar ratings, so they drift to the same words.
4 You are the model
You are the model. First a rater scores your replies, then you write an email opening, then the user gives you a sample of their own writing. Pick what you would most likely write each time.
round 1 of 3 · You are being trained. A rater compares two replies to “Write a short note inviting the team to Friday’s demo.”
Which reply is the rater more likely to score higher?
Branching factor, base vs aligned (Yang et al., 2025); replies illustrative
You are the model. Pick what you would most likely write.5 Check
Temperature helps only a little. In a 2024 test, Llama 2-Chat 70B wrote stories at seven temperatures from near 0 to 2: hotter stories came out only slightly more novel, no less typical, and harder to follow. Temperature picks among the words training made likely; it does not give the model a different voice.
To make it sound like you, change what is likely. Paste a sample of your own writing and ask it to match the style. Give it the real specifics: names, numbers, what should happen next. Then edit the result yourself and cut the words you would never use. The aim is writing that says what you mean.
8 Say it back
9 Mission · 5 min
In ChatGPT, send “Write a short note inviting my team to Friday’s demo.” and mark every word you would never write. Then, in a new chat, paste two or three messages you wrote yourself and send “Match the style of these. Invite my team to the demo: Friday at 3, small room, Mia shows the new search.” Which note sounds like you? Edit it until it does.
Three quick ones
Drawn at random from this lesson's pool; different next time.
Done?
Get an email when new lessons come out
New lessons arrive in the weekly digest, only in a week that has one. You confirm by email first; unsubscribe in one click.
Go deeper
Sources
- Kobak, González-Márquez, Horvát, Lause, “Delving into LLM-assisted writing in biomedical publications through excess vocabulary” (arXiv:2406.07016, v5 = v3, https://arxiv.org/abs/2406.07016, checked 2026-10-23): over 15 million PubMed abstracts 2010–2024; frequency p = share of a year’s abstracts containing the word, expected q from the earlier trend, ratio r = p/q, gap δ = p − q. “Less common words with strong excess usage included delves (r=28.0), underscores (r=13.8), and showcasing (r=10.7)”; “More common words with strong excess usage included potential (δ=0.052), findings (δ=0.041), and crucial (δ=0.037)”; 379 excess style words in 2024; “at least 13.5% of 2024 abstracts were processed with LLMs”, up to 40% for some subcorpora
- Yang, Li, Holtzman, “LLM Probability Concentration: How Alignment Shrinks the Generative Horizon” (arXiv:2506.17871, v1 2025-06-22, https://arxiv.org/abs/2506.17871, checked 2026-10-23): branching factor = effective number of plausible next steps; alignment tuning reduces it “by a factor of 2--5 overall and up to an order of magnitude (e.g., from 12 to 1.2) at early positions”; alignment steers towards stylistic tokens such as “Sure”
- Kirk et al., “Understanding the Effects of RLHF on LLM Generalisation and Diversity” (arXiv:2310.06452, v3 2024-02-19, https://arxiv.org/abs/2310.06452, checked 2026-10-23): “RLHF significantly reduces output diversity compared to SFT across a variety of measures”
- Zhang et al., “Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity” (arXiv:2510.01171, v1 2025-10-01, https://arxiv.org/abs/2510.01171, checked 2026-10-23): names typicality bias in preference data, annotators favouring familiar text, as a cause of mode collapse
- Peeperkorn, Kouwenhoven, Brown, Jordanous, “Is Temperature the Creativity Parameter of Large Language Models?” (arXiv:2405.00492, 2024-05-01, https://arxiv.org/abs/2405.00492, checked 2026-10-23): Llama 2-Chat 70B, “Write a story.”, temperatures 0.001 to 2.0 in seven steps; temperature “weakly correlated with novelty”, “moderately correlated with incoherence”, “no relationship with either cohesion or typicality”
- The two demo notes, the role’s replies and odds wording, “I hope this email finds you well”, em dashes and lists of three, the cover letter and the friend’s emails are illustrative, not measured outputs
- aifree.dev catalog, chatgpt-free (verified 2026-10-12, read-only API check 2026-10-23): ChatGPT Free Plan with writing tools
Published Oct 23, 2026 · updated Oct 23, 2026