LLM Voice

Oct 29, 2026•ibobev•View Original

TL;DR Highlight

Martin Fowler: LLM-drafted prose keeps an uncanny tone even after human editing.

Who Should Read

Developers who use LLMs to write READMEs, blog posts, or PR descriptions

Core Mechanics

LLM Voice is a human reaction
- Fowler says a growing sense of reading LLM output erodes his desire to read on.
- He says detection tools come too late because the reaction has already happened.
Don't let an LLM draft your prose
- He says it is very hard to remove LLM Voice even after human editing.
- He also considers accepting rephrasing suggestions dangerous.
Outlines only, at most
- He says using an LLM for an outline may be okay.
- Authors should fill in the prose themselves.
Writing is a thinking tool
- He says writing is how he gets his thoughts coherent.
- Delegating writing means delegating thinking and comprehension.
He allows exceptions
- Jim Highsmith, who has Parkinson's disease, finds an LLM essential to his writing.
- Translation, form letters, and reports nobody will read can reasonably use LLMs.
His site policy is disclosure
- Authors who use AI to write should declare how, usually in the acknowledgments.
- Fowler uses AI only as a search tool and verifies information at the source.

Community Reactions

Even an AI fan finds reading harder
- One commenter says samey, rambling text makes them abandon posts midway.
- They said they now retreat to pre-2023 content.
Model and settings mattered more than prompts
- A builder of an AI nonfiction book product reported this from tests.
- Sonnet 5 on high effort wrote the weakest prose, ran slower and costlier, and some chapters ran out of tokens mid-sentence.
Use LLMs as critics only
- One commenter says to ask for objective critique with no word suggestions.
- The better the suggested words look, the more they contaminate your voice.
Pushback on code versus prose
- One commenter questions why LLM code is accepted while LLM prose is not.
- Another says AI helps people with language barriers and blanket dismissal reeks of privilege.

How to Apply

Audit your last three documents for LLM tells
- Read three READMEs or PR descriptions aloud today and mark awkward sentences.
- Rewrite the marked sentences yourself.
Switch your review prompt to critique-only
- Forbid word and sentence suggestions when asking for feedback.
- Make the fixes yourself.
Add an AI-use disclosure rule
- Have team docs and blog posts state where AI was used in one line.
- Separate outline, translation, and search usage.

Try It

copy & run

# Paste into Claude or ChatGPT (simple version of the critique-only approach from the comments)

Evaluate the text below. Never suggest alternative words or sentences.

{{text}}

Tell me only:
1. Weak or missing logic
2. Repetitive or unnecessary parts
3. Parts readers may find hard to follow

Next: rewrite it yourself based on the critique, then paste it again to compare.

Terminology

LLM VoiceThe distinctive feel of text that seems LLM-written. It is smooth but lacks human character.

uncanny valleyThe discomfort of something almost human but subtly off. Fowler applies it to prose.

StylometryThe measurement of writing style, like a fingerprint for an author.

Related Resources

LLM Voice

Oct 29, 2026•ibobev•View Original

TL;DR Highlight

Martin Fowler: LLM-drafted prose keeps an uncanny tone even after human editing.

Who Should Read

Developers who use LLMs to write READMEs, blog posts, or PR descriptions

Core Mechanics

LLM Voice is a human reaction
- Fowler says a growing sense of reading LLM output erodes his desire to read on.
- He says detection tools come too late because the reaction has already happened.
Don't let an LLM draft your prose
- He says it is very hard to remove LLM Voice even after human editing.
- He also considers accepting rephrasing suggestions dangerous.
Outlines only, at most
- He says using an LLM for an outline may be okay.
- Authors should fill in the prose themselves.
Writing is a thinking tool
- He says writing is how he gets his thoughts coherent.
- Delegating writing means delegating thinking and comprehension.
He allows exceptions
- Jim Highsmith, who has Parkinson's disease, finds an LLM essential to his writing.
- Translation, form letters, and reports nobody will read can reasonably use LLMs.
His site policy is disclosure
- Authors who use AI to write should declare how, usually in the acknowledgments.
- Fowler uses AI only as a search tool and verifies information at the source.

Community Reactions

Even an AI fan finds reading harder
- One commenter says samey, rambling text makes them abandon posts midway.
- They said they now retreat to pre-2023 content.
Model and settings mattered more than prompts
- A builder of an AI nonfiction book product reported this from tests.
- Sonnet 5 on high effort wrote the weakest prose, ran slower and costlier, and some chapters ran out of tokens mid-sentence.
Use LLMs as critics only
- One commenter says to ask for objective critique with no word suggestions.
- The better the suggested words look, the more they contaminate your voice.
Pushback on code versus prose
- One commenter questions why LLM code is accepted while LLM prose is not.
- Another says AI helps people with language barriers and blanket dismissal reeks of privilege.

How to Apply

Audit your last three documents for LLM tells
- Read three READMEs or PR descriptions aloud today and mark awkward sentences.
- Rewrite the marked sentences yourself.
Switch your review prompt to critique-only
- Forbid word and sentence suggestions when asking for feedback.
- Make the fixes yourself.
Add an AI-use disclosure rule
- Have team docs and blog posts state where AI was used in one line.
- Separate outline, translation, and search usage.

Try It

copy & run

# Paste into Claude or ChatGPT (simple version of the critique-only approach from the comments)

Evaluate the text below. Never suggest alternative words or sentences.

{{text}}

Tell me only:
1. Weak or missing logic
2. Repetitive or unnecessary parts
3. Parts readers may find hard to follow

Next: rewrite it yourself based on the critique, then paste it again to compare.

Terminology

LLM VoiceThe distinctive feel of text that seems LLM-written. It is smooth but lacks human character.

uncanny valleyThe discomfort of something almost human but subtly off. Fowler applies it to prose.

StylometryThe measurement of writing style, like a fingerprint for an author.

LLM Voice

TL;DR Highlight

Who Should Read

Core Mechanics

Community Reactions

How to Apply

Try It

Terminology

Related Papers

Related Resources

LLM Voice

TL;DR Highlight

Who Should Read

Core Mechanics

Community Reactions

How to Apply

Try It

Terminology

Related Papers

Related Resources