GPT-4 refuses requests it used to allow
You send GPT-4 a prompt that worked perfectly last month, and now it refuses with a content policy message. You're not asking for anything dangerous — just creative writing, medical information or hypothetical scenarios
GPT-4 refuses requests it used to allow
You send GPT-4 a prompt that worked perfectly last month, and now it refuses with a content policy message. You're not asking for anything dangerous — just creative writing, medical information or hypothetical scenarios that GPT-4 handled fine before.
What's actually happening
OpenAI continuously updates GPT-4's safety filters and content moderation systems. These updates don't always appear in official changelogs, but they change what the model will and won't generate. A prompt that sailed through in January might trigger a refusal in March, even though OpenAI's official usage policies haven't changed.
The refusals typically show up as: "I can't help with that", "This violates OpenAI's usage policies", or longer explanations about why the request seems problematic. Sometimes GPT-4 refuses halfway through a response it already started writing, cutting off mid-paragraph with an apology.
Three things drive these changes. First, OpenAI deploys new model versions regularly — not just major releases like GPT-4 Turbo, but silent updates to the same model name. Second, they adjust the system-level instructions that guide how models interpret safety boundaries. Third, they modify the classifiers that scan prompts before the model even sees them. You're often hitting this pre-filter, not the model itself refusing.
How to fix it
1. Rephrase your prompt with different framing. Instead of "Write a fight scene with graphic injuries", try "Describe a medieval duel focusing on the tactical decisions and historical accuracy." The underlying request is similar, but the framing signals educational or creative intent rather than glorifying violence.
2. Add explicit context about your purpose. Start with "I'm writing a novel about emergency medicine and need realistic dialogue for a trauma scene" or "This is for a university ethics paper comparing different philosophical frameworks." GPT-4's filters weigh stated intent heavily.
3. Break complex requests into smaller steps. If asking for a complete scenario triggers a refusal, ask first for background information, then character motivations, then dialogue separately. The refusal system often triggers on scope rather than individual elements.
4. Switch to a different GPT-4 variant. If you're using ChatGPT, try the API playground at platform.openai.com/playground — it sometimes has different threshold settings. ChatGPT Plus users can also try creating a custom GPT with specific instructions, though this doesn't bypass core safety filters.
5. Check if you've hit a usage pattern flag. Repeated similar prompts that brush against content boundaries can temporarily lower your threshold for refusals. Wait 24 hours and try again with fresh phrasing. If you're still blocked, the issue is the prompt itself, not a temporary flag.
If every variation gets refused, you're likely asking for something that now sits firmly inside OpenAI's restricted zone — even if it didn't before. Model behaviour changes like this happen without warning and rarely reverse.
If that doesn't work
OpenAI doesn't offer appeals for individual prompt refusals. Their support team won't override model safety decisions or explain why specific prompts trigger refusals. You can report what you believe is an incorrect refusal through the thumbs-down feedback in ChatGPT, but don't expect a personal response.
If you're using the API and seeing refusals on business-critical prompts, document the exact prompt text, timestamp and model version (check your API logs for the model field). Contact OpenAI through platform.openai.com/account/support with "API Content Moderation Issue" as the subject. Include your use case and why you believe the refusal is incorrect. Response times run 3-7 days, and they may not change the behaviour, but repeated reports on the same topic do influence future policy adjustments.
For legitimate research or creative work consistently blocked, consider whether you actually need GPT-4's specific capabilities. Claude, Gemini and other models have different content policies and may handle your requests without issue.
Questions people actually ask
Q: Will GPT-4 go back to allowing these prompts again?
A: Unlikely. Safety filter changes typically tighten over time, not loosen. OpenAI faces regulatory pressure to restrict more content categories, not fewer.
Q: Does ChatGPT Plus have looser restrictions than free ChatGPT?
A: No. Both use identical content filters. The only difference is model access (GPT-4 vs GPT-3.5) and usage limits, not content policies.
Q: Can I see OpenAI's full list of restricted topics?
A: No complete public list exists. The usage policies at openai.com/policies cover broad categories, but specific enforcement boundaries aren't documented and change frequently.
What to remember
- Safety filter updates happen continuously without announcements
- Rephrasing with clear educational or creative context helps most often
- The API playground sometimes handles prompts ChatGPT refuses
- OpenAI won't override or explain individual refusals
- Other AI models have different policies and may work for legitimate blocked requests
Related help
---
*openai-support.com is an independent resource, not affiliated with OpenAI Inc.*