Thomas Ptacek, a security researcher and writer, advises treating large language models (LLMs) primarily as editorial tools rather than as sources of phrasing to be copied verbatim. His core recommendation is a strict rule: do not use any specific turn of phrase or word suggested by an LLM.
The intent of this rule is to keep any particular phrasing generated by the model off-limits. Ptacek argues that LLM-generated wording carries a distinctive "smell," and avoiding direct reuse of those turns of phrase helps maintain authorial discipline and originality.
At the same time, Ptacek does not reject LLMs outright. He says he will not let LLMs write content for his blog, but he uses them for fact-checking, spelling and grammar checks, and occasionally as a thesaurus.
In his discussion he includes a screenshot of his personal LLM-based copyediting tool and shares a prompt intended to help others get started building their own workflows. He also points readers to a related Twitter thread where the topic has been discussed further.
This guidance offers a practical middle path for writers who want to leverage LLMs for accuracy and style polishing while retaining their own voice and avoiding direct reuse of model-generated expressions.
Practical boundaries
- Do not allow an LLM to generate whole pieces of content in place of the author.
- Use LLMs for fact-checking, spelling, and grammar.
- Use LLMs as an occasional thesaurus, but do not adopt any suggested phrasing verbatim.
Ptacek’s approach is conservative about authorship yet pragmatic about utility: it acknowledges the usefulness of LLMs while clearly distinguishing human authorship from machine-suggested text.
