Why is GPT-4 refusing requests it used to allow
GPT-4 suddenly says no to prompts it handled fine last month. You're not imagining it—OpenAI continuously updates model behaviour, tightening safety filters without advance notice.
Why is GPT-4 refusing requests it used to allow
GPT-4 suddenly says no to prompts it handled fine last month. You're not imagining it—OpenAI continuously updates model behaviour, tightening safety filters without advance notice.
What's actually happening
OpenAI deploys what they call "system-level mitigations" between model versions. You're likely using `gpt-4-turbo` or the default GPT-4 endpoint, which silently switches to newer snapshots like `gpt-4-turbo-2024-04-09` or later. Each snapshot includes modified content policies that catch different patterns.
The refusal you see—"I can't help with that" or "This request violates our usage policies"—doesn't mean you broke rules. It means the threshold shifted. OpenAI's classifiers now flag content categories they previously allowed: detailed medical advice, certain creative writing scenarios, code that could theoretically harm systems, or business strategy that touches regulated industries.
This affects paying users most. The $20/month ChatGPT Plus subscription uses whatever GPT-4 version OpenAI designates as current. API users on `gpt-4` or `gpt-4-turbo` endpoints get automatic updates. You can't opt out unless you pin to a specific dated model like `gpt-4-0613`, but those versions enter deprecation cycles and eventually stop working entirely.
How to fix it
1. Rephrase your prompt without trigger words
Remove terms the classifier targets: "hack", "manipulate", "exploit", "bypass", medical diagnoses, drug names, or anything resembling jailbreak attempts. Ask for "alternative approaches" instead of "workarounds." Request "creative fiction" rather than "realistic scenarios."
If GPT-4 still refuses, break your request into smaller questions. Instead of "write a complete vulnerability assessment," ask "explain how authentication systems validate tokens" first, then build from there.
2. Switch to a dated snapshot
In the API, change your model parameter from `gpt-4-turbo` to `gpt-4-0613`. This locks you to the June 2023 version with older filters. Check OpenAI's model deprecation timeline before committing—`gpt-4-0613` remains available through mid-2025.
ChatGPT Plus users can't select dated models directly, but you can switch to GPT-4o in the model dropdown. Different model families use different moderation systems. GPT-4o sometimes allows requests GPT-4 refuses, though it's not consistent.
3. Check if your account has additional restrictions
Go to platform.openai.com/account/limits. If you see "Usage tier 1" or warnings about policy violations, your account carries extra moderation. Previous flags—even if resolved—make future refusals more aggressive. This compounds across requests: three refusals in one conversation increases the likelihood the fourth similar prompt gets blocked.
Clear your chat history and start a new conversation. The system maintains context across messages, so earlier refusals bias later responses. In the API, pass `messages` with only your current request, not the full thread.
4. Use Custom Instructions to set appropriate context
In ChatGPT Plus, click your profile icon → Custom Instructions. Add context like "I'm a security researcher working in a controlled environment" or "This is for educational fiction writing." The model sees this with every request and adjusts its interpretation. Don't claim credentials you lack—false context can trigger trust and safety reviews.
If that doesn't work
You're hitting a hard filter that won't budge through rephrasing. Submit feedback through the thumbs-down button in ChatGPT or contact OpenAI support via help.openai.com. Include:
- Exact prompt text that got refused
- The model version (check Settings → Data controls → Model)
- What similar requests worked previously
- Your use case—academic research, creative writing, software development
Support responds in 2-4 business days for Plus users, longer for free accounts. They won't override model behaviour, but patterns in feedback influence future policy adjustments. If multiple users report the same legitimate use case getting blocked, OpenAI sometimes loosens that specific filter in subsequent updates.
For API users hitting this repeatedly, consider Claude (Anthropic) or Gemini (Google) as alternatives. They maintain different content policies and may allow your specific requests.
Questions people actually ask
Q: Can I permanently use an older GPT-4 version that was less restrictive?
A: Only through the API with dated model names like `gpt-4-0613`, and only until OpenAI deprecates that version. ChatGPT Plus always uses the current default model—you can't lock to older snapshots.
Q: Does GPT-4 remember that I got flagged before?
A: Within a conversation, yes—context affects moderation. Across different chats, no, unless your account has formal policy violations logged. Start fresh conversations for unrelated requests.
Q: Will paying for API access give me looser content filters than ChatGPT Plus?
A: No. The same model version uses identical moderation whether accessed through chat or API. Rate limits differ, but content policies don't.
What to remember
- OpenAI updates GPT-4 behaviour continuously—what worked last month may not work today
- Dated models like `gpt-4-0613` preserve older behaviour until deprecation
- Rephrasing removes trigger words the classifier targets
- Multiple refusals in one chat make subsequent blocks more likely
- Model behaviour changes affect all users simultaneously, not just your account
---
*openai-support.com is an independent resource, not affiliated with OpenAI Inc.*