OpenAI Published a 'Slop Words' Blocklist
OpenAI just did something rare for an AI company: it admitted its own model sounds like AI, and it published the exact words it wants developers to ban.
The documentation, released alongside GPT-6 Astra's general rollout, spells out where the company's most capable model tends toward unwanted behavior and how developers can work around it. The centerpiece is a blocklist of AI "slop words" -- phrases that instantly flag a piece of text as machine-generated.
The Slop Words Blocklist
OpenAI calls them "slop words" directly in the documentation. The list includes:
"Delve into," "promote," "use/leverage," "it's worth noting," "what's important is," "foster," "harness," "in summary," and using "Conclusion:" as a section header. Made-up hyphenated compounds like "exact-head checks" or "editorial-row layouts" should also be avoided.
The model should "write like a person, not like an AI language model trying to sound impressive," the guidance says. For technical writing, OpenAI recommends keeping jargon to cases where it actually helps: "Use plain language over jargon, and reference technical details only to the degree that it helps illustrate a concept."
The blocklist is notably self-aware. Every phrase on it is one that OpenAI's own models have used in production, producing text users immediately recognize as AI-generated.
Teaching GPT-6 Astra to Take Initiative
GPT-6 Astra asks clarifying questions more often than GPT-5.6 Sol, according to OpenAI, making it a "more effective collaborator." But the trade-off is that the model can become overly cautious, pausing for approval on routine decisions.
To push it toward more initiative, OpenAI recommends a system prompt that tells the model to infer the user's "intent" from context and show a "bias towards action." The sample prompt reads:
"You should infer the user's intent and the scope of the task from the instructions and the conversation context so far. Your task is to demonstrate a tendency to act and to follow through on the user's goal."
The model should wait to ask for approval only after it has already prepared "a concrete, reviewable result." Unsolicited offers by the model to execute a plan are encouraged.
The Debugging Prompt
GPT-6 Astra follows longer instructions better than its predecessors, but it is also more sensitive to context. Unclear or contradictory instructions in skill files can cause the model to block work or veer off unexpectedly.
OpenAI recommends a specific debugging prompt: force the model to name the exact skill file and quote the specific instruction that caused it to pause. The prompt reads:
"If a skill causes you to ask for permission or confirmation, pause, leave requested work unfinished, or diverge from the user's intent, name and link to the exact SKILL.md file you read, quote the relevant instruction, and explain how that instruction led to your behavior."
This gives developers a traceable path to the root cause rather than guessing why the model stopped working.
The Coding Test Overshoot
On coding tasks, GPT-6 Astra runs thorough tests before wrapping up. For small changes, that can mean executing test suites wildly out of proportion to the actual work.
OpenAI recommends telling the model to report back on what kind of testing is appropriate for each change, rather than running the full suite every time. Messages between agents can also contain grammar or spacing errors, requiring explicit formatting instructions.
What This Says About the State of AI
The slop words blocklist is the most telling detail in the entire guidance. OpenAI is effectively saying: we trained a model that, when left to its own devices, produces text that sounds like an AI parody of itself. Developers need a banned-word list to make it write naturally.
This is remarkable candor from the company that runs ChatGPT, the most widely deployed AI writing tool in the world. It confirms what users have long suspected: the unmistakable "AI voice" -- the relentless bullet points, the cautious hedging, the empty connector phrases -- is baked into the models, not an accident of prompting.
Whether the blocklist actually improves output quality, or merely trades one set of verbal tics for another, remains to be seen. But the fact that OpenAI felt the need to publish one at all says more about the current state of AI-generated text than any blog post about capabilities ever could.