How to Reduce AI Slop With Technical English

TL;DR

Give an AI a complete, mechanically testable writing system instead of banning isolated words. In tests across six developer writing tasks and two models, ASD-STE100 reduced measured slop by at least half and was best or tied for best, although the linter measured only checkable features and could not certify full compliance.

Transcript

The same prompt, same model, two answers. Answer one says, "Designed to slide into your existing stack with minimal friction and no window lock in." And the answer two says, "A normal cache matches requests by exact text. A small change inverting then causes a cash miss." And one of these is slop. And two million people just watched a thread claime... Read More

Key Insights

• AI slop is a collection of identifiable writing habits, including synonym rotation, stacked hedging, nominalization, marketing adjectives, run-on sentences, and chatty phrasal verbs. Defining these habits mechanically makes them easier to prevent and measure than relying on a general feeling that prose sounds artificial.
• ASD-STE100 is a controlled English specification first released in 1986 for aircraft maintenance documentation. It was designed to reduce ambiguity for readers, especially mechanics who were not native English speakers, because misunderstanding a maintenance instruction could have severe consequences.
• The ASD-STE100 dictionary contains roughly 900 approved words, with each word assigned one meaning and one grammatical role. It permits a preferred term such as "start" while excluding alternatives such as "begin," "commence," "initiate," and "originate," which prevents synonym rotation by design.
• The standard suppresses weak prose through explicit rules. It requires one name for each thing, rejects complicated verb constructions, favors verbs over nouns for actions, excludes unapproved marketing adjectives, bans semicolons, and replaces phrasal verbs with direct alternatives such as "remove the panel."
• Sentence length is constrained according to purpose. An instruction can contain no more than 20 words, while a descriptive sentence can contain no more than 25 words. These limits discourage multiple ideas from being joined into long sentences that readers must decode.
• Human-reader evidence supports controlled English, with important limits. A 1996 study of 175 aircraft technicians reported comprehension rising from 76% to 86%, while non-native speakers improved from 69% to 87%. Other findings suggest oversimplification can sometimes slow readers, and improved memory was not established.
• The Claude experiment found that the baseline scored 4.36 violations per 100 words, a banned-word list scored 4.21, Orwell's rules scored 2.48, and ASD-STE100 scored 1.12. The controlled writing system therefore reduced measured slop by 74% compared with the baseline.
• Model behavior is not uniform. On GPT-5.5, ASD-STE100 reduced measured slop by 50%, the banned-word list reduced it by 40%, and Orwell's rules tied ASD-STE100. The result that held across both models was that a complete writing system cut violations by at least half.

Questions & Answers

Q: How can I reduce AI slop in technical writing?

Give the model a complete writing system with explicit vocabulary, grammar, sentence-length, and terminology rules. The tested ASD-STE100 approach requires one name for each thing, direct verbs for actions, short sentences, approved vocabulary, and limited phrasal verbs. Across six developer writing tasks and two models, this system reduced measured slop by at least half and was best or tied for best.

Q: What is ASD-STE100 and why was it created?

ASD-STE100 is a controlled version of English created for aircraft maintenance documentation and first released in 1986. It was designed to reduce ambiguity for mechanics around the world, including readers who were not native English speakers. The current issue discussed in the source was released in January 2025, and the complete standard contains 434 pages of procedures, rules, and controlled vocabulary.

Q: What writing habits make AI-generated text sound like slop?

Six recurring habits define the problem mechanically: using several names for the same thing, stacking hedging verbs, turning actions into abstract nouns, inserting marketing adjectives, joining too many ideas into run-on sentences, and using soft phrasal verbs. Examples include calling one person a user, customer, and client, or writing "perform an analysis" instead of the direct verb "analyze."

Q: Why does banning words fail to fix AI writing reliably?

A banned-word list targets visible symptoms without supplying a coherent method for constructing clear prose. In the Claude test, it reduced the measured score only from 4.36 to 4.21, a 3% improvement, even though it reduced em dashes from six to one. On GPT-5.5 it performed better, reducing violations by 40%, which shows that its effectiveness varies by model.

Q: Which ASD-STE100 rules directly prevent AI slop?

The standard requires one name for each thing and assigns approved words one meaning, which prevents synonym rotation. It rejects complicated verb constructions, requires verbs rather than nouns for actions, excludes unapproved marketing adjectives, bans semicolons, limits instructions to 20 words and descriptions to 25 words, and replaces phrasal verbs with direct terms such as "remove" instead of "take off."

Q: Does simplified technical English improve comprehension?

The strongest study cited tested 175 aircraft technicians using maintenance work cards. Comprehension rose from 76% with regular English to 86% with simplified English. Among non-native speakers, it rose from 69% to 87%. A separate Microsoft Research study tested 520 sentences translated into Chinese, French, Dutch, and Arabic, and the controlled versions performed better in every language.

Q: How was ASD-STE100 tested on AI-generated writing?

The experiment used six practical developer writing jobs: a README, pull request description, API documentation, error message, getting-started guide, and deprecation notice. Each job was generated under four conditions: a plain baseline, a banned-word list, Orwell's six rules, and an ASD-STE100 skill. A linter then counted sentence length, semicolons, passive voice, phrasal verbs, and marketing adjectives per 100 words.

Q: Can a linter certify full ASD-STE100 compliance?

No. The linter in the experiment measured only mechanical features associated with slop, including sentence length, semicolons, passive voice, phrasal verbs, and marketing adjectives. The standard's maintainers state that software cannot certify complete compliance because some rules require human judgment about whether a sentence makes sense. The claim is narrower: the mechanical slop patterns can be checked by a script.

Summary & Key Takeaways

• AI slop can be described through recurring mechanical habits: synonym rotation, hedging, actions converted into nouns, marketing adjectives, run-on sentences, and chatty phrasal verbs. Naming these specific patterns makes them measurable. It also enables writers to replace subjective requests for better prose with explicit constraints that models and scripts can apply consistently.

• ASD-STE100 is a controlled form of English first released in 1986 for aircraft maintenance documentation. Its procedures, restricted dictionary, fixed word meanings, sentence-length limits, and grammatical rules directly target ambiguity. These constraints also suppress many traits associated with AI-generated prose, even though the standard was created decades before modern language models.

• The experiment compared a baseline, a banned-word list, Orwell's six rules, and an ASD-STE100 skill across six developer writing tasks and two models. ASD-STE100 reduced measured violations by 74% on Claude and 50% on GPT-5.5. The broader conclusion is that coherent writing systems are more reliable than isolated stylistic bans.