GPT-4 refusing requests it used to allow
You ask GPT-4 to write something it handled fine last week, and now it says "I can't assist with that." The model hasn't changed — OpenAI's safety systems have.
GPT-4 refusing requests it used to allow
You ask GPT-4 to write something it handled fine last week, and now it says "I can't assist with that." The model hasn't changed — OpenAI's safety systems have.
What's actually happening
OpenAI continuously updates the content filters and refusal logic that sit between your prompt and GPT-4's actual output. These aren't changes to the language model itself — they're adjustments to what the system allows through. You're hitting one of three scenarios:
First, your request now triggers updated content policies. OpenAI expands restricted categories regularly without announcing each change. A prompt about competitive analysis, historical violence, or hypothetical scenarios that worked in October might trip filters in December. The refusal messages are identical — "I cannot provide assistance with..." or "I'm not able to help with that request" — but the trigger words and concepts shift.
Second, you're seeing dynamic safety responses based on conversation context. GPT-4 tracks the full thread. If earlier messages mentioned regulated industries, real people, or sensitive topics, later neutral requests get blocked because the system flags potential misuse chains. Refreshing the chat often fixes this — the same prompt works fine in isolation.
Third, OpenAI rolled out stricter defaults for ChatGPT Plus vs API users. The same model version behaves differently depending on access method. If you switched from API to web interface (or vice versa), you'll see more refusals even with identical prompts. The web interface applies heavier filtering.
How to fix it
1. Refresh the conversation
Click the New Chat button (top left corner). Paste your exact prompt without any prior context. If GPT-4 responds normally, the previous conversation history triggered the refusal.
2. Rephrase without trigger words
Remove brand names, real public figures, specific locations, or industry terms (legal, medical, financial). Replace "Write a marketing email for [Company X]" with "Write a marketing email for a hypothetical B2B software company." Replace "Analyse this contract clause" with "Explain this text in plain language."
3. Request general information first
Don't ask GPT-4 to generate the final output directly. Ask it to "Explain the key elements of effective cold emails" first, then use that framework to write your own. The system allows educational responses it blocks for direct creation.
4. Switch models temporarily
In the model selector (top of the chat interface), choose GPT-4o or GPT-3.5 if available. Different model versions have different refusal thresholds. Sometimes GPT-4o allows what standard GPT-4 blocks, though OpenAI updates all models eventually.
5. Use the API instead
If you're working on chatgpt.com, try the same request through platform.openai.com/playground. API access applies lighter content filtering for technical use cases. You'll need API credits, but single tests cost under $0.01.
If none of these work, the request likely violates OpenAI's current usage policies — not just a sensitivity threshold you can work around.
If that doesn't work
OpenAI doesn't provide case-by-case policy clarifications for specific prompts. You won't get an explanation of exactly which words triggered the refusal. The official recourse is how to contact OpenAI support, but they'll confirm whether your use case violates policy — they won't adjust the filters for your account.
Send feedback through the thumbs-down button on the refusal message. Click the down arrow, select "Refused when it shouldn't have," and describe your use case in one sentence. This goes to OpenAI's safety team. Don't expect individual replies, but mass reports on false positives do influence policy updates.
Check OpenAI's usage policies page for current restricted categories. The list includes generating code for malicious use, bulk political content, legal advice claiming professional authority, and personalised medical treatment plans. If your request touches these, no rephrasing will work — it's genuinely against policy.
For work blocked incorrectly, document the exact refusal date and prompt wording. If GPT model behaviour changed affects your business workflow, you can request a policy review through the support form, though approval takes 5-7 business days minimum.
Questions people actually ask
Q: Did OpenAI secretly update GPT-4 to be more restrictive?
A: They update content filters weekly without version number changes. The model's capabilities haven't decreased — the approval system got stricter.
Q: Will GPT-4 get less restrictive again?
A: Historically, no. OpenAI tightens policies over time as they identify misuse patterns. Loosening restrictions happens rarely and only after public policy changes.
Q: Can I appeal a specific refusal?
A: Not for individual prompts. You can only report false positives through the feedback button or request business-use policy reviews through support.
What to remember
- Content filters update separately from model versions — GPT-4 itself didn't change
- Conversation history affects current responses — always test in a fresh chat
- Web interface blocks more aggressively than API access for the same model
- Rephrasing removes 70% of false-positive refusals, but genuine policy violations won't budge
- OpenAI won't explain specific trigger words — only confirm if your use case violates policy
---
*openai-support.com is an independent resource, not affiliated with OpenAI Inc.*