- Author
- Eryk Salvaggio (Cybernetic Forests)
- Published
- 31 May 2026
- Kind
- Newsletter essay, argued without data
- Original
- mail.cyberneticforests.com/its-not-just-data-its-post-training/
- Subject
- Writing with models
What it is
Eryk Salvaggio writes the Cybernetic Forests newsletter. This eight-minute essay covers where the tic may come from. It also covers what detection tools do to writers.
What it says
- Salvaggio calls negative parallelism a rhetorical device, as lazy or inspired as its content.
- Salvaggio suspects the tic comes from RLVR. That training boosts wording used before a verified right answer.
- Salvaggio illustrates this with two friends working out which day they saw a weird dog.
- Grammarly flagged 27 passages in one section of the essay. Salvaggio paid Pangram $20 to confirm a journal article was not AI-written.
- Salvaggio cites a UK AI essay grader that rewarded length, vocabulary range and sentence complexity.
- Salvaggio applies Goodhart's law to language. Detector pressure, Salvaggio warns, removes useful tools of argument.
- Salvaggio says a 99.8% accuracy figure compounds per paper, citing Arvind Narayanan. Up to 10% of students could then be falsely accused.
What evidence it gives
The essay offers argument, analogy and one personal anecdote about Grammarly and Pangram. It cites a UK grader test and Arvind Narayanan on error rates. The RLVR explanation is the author's own suspicion.
Interests and limits
The author states the RLVR link as a suspicion, not a finding. No model was tested. The essay is opinion from an artist and critic of AI. Its figures come from other sources.