GPT-4 responses getting shorter and refusing more
You're getting clipped answers where GPT-4 used to write detailed explanations, or you're hitting refusals on prompts that worked fine last month. You're not imagining it — and you're not alone.
GPT-4 responses getting shorter and refusing more
You're getting clipped answers where GPT-4 used to write detailed explanations, or you're hitting refusals on prompts that worked fine last month. You're not imagining it — and you're not alone.
What's actually happening
OpenAI regularly updates GPT-4 and GPT-4o models with new versions that change response behaviour. These aren't bugs — they're deliberate adjustments to model training, safety filters, or response generation patterns. When OpenAI deploys a new snapshot (like `gpt-4-turbo-2024-04-09` replacing an older version), you'll see immediate changes in how the model responds.
The shorter responses usually come from two places. First, OpenAI trains newer versions to be more concise because most users want faster, cheaper answers. The model learns that shorter responses score better in internal evaluations. Second, you might be hitting the response length limit sooner. GPT-4 has a maximum output token limit (usually 4,096 tokens for standard responses), and if you're running complex conversations with long context, the model prioritises staying within that budget by cutting responses short.
The increased refusals are typically from updated safety guidelines. OpenAI adjusts content policies every few months based on abuse patterns they observe across millions of conversations. What used to pass as a legitimate creative writing prompt might now trigger refusal logic if it resembles harmful patterns the new model version was trained to reject. You'll see messages like "I can't assist with that" or "I need to decline this request" more often — especially for prompts involving public figures, medical advice, legal guidance, or content generation that could violate copyrights.
How to fix it
1. Check which model version you're actually using
Go to your ChatGPT settings (gear icon, bottom-left) and look under "Model" or check the dropdown at the top of a new conversation. If you're on ChatGPT Plus, you should see "GPT-4" or "GPT-4o" as options. The web interface doesn't always show the exact snapshot version (like `-0613` or `-turbo-2024-04-09`), but API users can check their request logs to see which specific version responded.
2. Tell the model exactly what length you need
Start your prompt with explicit length instructions: "Write a 500-word explanation" or "Give me a detailed response of at least three full paragraphs." This doesn't override hard limits, but it signals to the model that you want comprehensive output, not a summary.
3. Break requests into smaller chunks
Instead of asking for a complete 2,000-word article in one go, ask for "Part 1: Introduction and first two sections" then follow up with "Continue with Part 2." This avoids hitting the response length cap and often produces better quality output anyway. For comprehensive troubleshooting steps when ChatGPT isn't working as expected, this chunking approach helps isolate where behaviour changed.
4. Rephrase refused prompts
If you get a refusal, reframe your request to be more obviously legitimate. Instead of "Write a speech for [politician]," try "Explain the main policy arguments a politician might use when discussing [topic]." Add context showing your use case is educational, analytical, or clearly fictional. Remove names of real people if the task doesn't actually require them.
5. Try GPT-4o instead of standard GPT-4
The GPT-4o model (the "o" means "omni") sometimes has different response patterns than base GPT-4. Switch models using the dropdown at conversation start. Some users find GPT-4o more willing to engage with borderline prompts, while others report it's more conservative — it varies by prompt type.
If that doesn't work
Document exactly what changed. Save screenshots of the old responses (if you still have them) and the new refusals or shortened outputs. Record the exact prompt you're using and which model version is responding.
Then contact OpenAI support through help.openai.com with this information. Include your account email, the date you first noticed the change, and specific examples. Response times average 3-5 business days. OpenAI won't roll back model changes for individual users, but if your use case is legitimate and you're hitting false positive refusals, they can sometimes whitelist your account or provide guidance on prompt adjustments.
For more context on why GPT model behaviour changes over time, check the OpenAI API documentation for model deprecation schedules and snapshot version notes.
Questions people actually ask
Q: Did OpenAI make GPT-4 dumber to save money?
A: No. Newer versions are trained differently, often optimising for speed and conciseness rather than verbosity. That feels like reduced capability, but it's an intentional trade-off based on aggregate user preferences.
Q: Can I force GPT-4 to use the old version?
A: Not through ChatGPT directly. API users can specify exact model snapshots like `gpt-4-0613` until OpenAI deprecates them, but those eventually stop working. The web interface always uses the current default version.
Q: Why does it refuse things that aren't against the rules?
A: Safety filters cast a wide net and catch false positives. The model can't distinguish between you researching historical events and someone trying to generate harmful content about the same topic — it just sees pattern matches.
Q: Will responses get longer again?
A: Only if OpenAI releases a new version optimised for longer outputs. User feedback influences this, so reporting the issue through official channels matters.
What to remember
- Model updates change behaviour permanently — old versions eventually disappear
- Specify exact length requirements in your prompts to get fuller responses
- Chunk large requests into multiple smaller prompts to avoid length caps
- Rephrase refused prompts with more context showing legitimate use
- Document changes and report false positive refusals to OpenAI support
---
*openai-support.com is an independent resource, not affiliated with OpenAI Inc.*