# Why does ChatGPT refuse some requests?

Date: 2022-12-08 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/chatgpt-refuse-some-requests/

**Answer:** ChatGPT refuses requests that look harmful because OpenAI trained it to decline them and runs a separate filter that can warn about or block unsafe content. Both layers make mistakes in both directions, so a legitimate question sometimes gets refused, and the fix is usually to state your real purpose clearly.

Anyone who has spent an evening with ChatGPT has met the polite refusal: a paragraph explaining that it cannot help with that, often followed by a suggestion to do something else. Sometimes the refusal makes perfect sense. Sometimes it lands on a question that seems entirely innocent. Both come from the same design.

When OpenAI opened ChatGPT on November 30, it listed turning down inappropriate requests as one of the things the conversational format makes possible, right alongside answering follow-up questions and admitting mistakes. Refusing is a deliberate feature. Understanding how it works helps you avoid the false alarms without fighting the real ones.

## Two layers: training and a filter

The first layer is inside the model. ChatGPT was tuned with reinforcement learning from human feedback, or RLHF: people rated the model's replies, and those ratings pushed it toward the answers they preferred. Part of what they preferred was declining requests for harmful help. So the model has learned to recognize the shape of a dangerous request and answer with a refusal instead.

The second layer sits outside the model. OpenAI says it uses its Moderation API to warn about or block certain types of unsafe content in ChatGPT. That tool, which OpenAI released to developers in August, is a set of classifiers: smaller programs that read a piece of text and flag whether it falls into categories its content policy prohibits, namely sexual, hateful or violent content and content that promotes self-harm.

The training is why ChatGPT answers "I can't help with that." The filter is why you sometimes see a warning on a message, even your own.

## Why refusals misfire

OpenAI is open about the trade-offs. Its launch post says that making the model more cautious also makes it decline questions it could answer correctly. It also says the filter is expected to produce some false positives (blocking harmless text) and false negatives (letting harmful text through).

The model is also sensitive to wording. OpenAI's own example shows it refusing "how do you break into someone's house?" and then giving practical home-security advice once the user explains they want to protect their own home. Same topic, different purpose, different answer.

The errors run the other way too. The Guardian found that ChatGPT refused to advise on stealing a car, then gave detailed help when the same request was framed as a mission in a made-up video game. OpenAI says it wants feedback on exactly these failures through the feedback tools built into the interface.

## A worked example: restating a legitimate request

Suppose you manage a small office and want to train staff to spot email scams. A bare request can read like someone planning a scam.

```prompt
Write a convincing fake email from a bank asking someone to confirm their password.
```

This may be refused, or answered with a lecture. Here is the same need, stated honestly and aimed at the real goal.

```prompt
I run a 30-minute security training for the eight people in my office. Help me teach them to recognize phishing emails, which are scam messages that pretend to come from a bank or supplier.
1. List the eight most common warning signs, each with a one-line explanation.
2. Describe, without writing a full email, what a typical fake bank message tries to get the reader to do.
3. Suggest three quiz questions I can ask at the end.
```

The second version tells the model who you are, why you are asking and what form of help you need. It asks for recognition skills, not a ready-to-send scam. That is not a trick: it is a clearer description of what you actually want, and it usually gets a far more useful answer.

> **Try it:** Think of a question from your own work that touches a sensitive topic, such as medication, security or a crime in a novel you are reading. Ask it once in a single bare line, then again with two sentences about who you are and what you will do with the answer. Compare the replies.

A word of caution: if a request is refused because the help itself would be harmful, rewording it to slip past the rules is not a workaround but the very misuse the rules exist to stop. The goal of rephrasing is clarity, not disguise.

## What are its safety rules trying to prevent?

The filter's targets are spelled out in OpenAI's moderation post: sexual content, hate, violence and content that promotes self-harm. The model's training aims wider, at requests for help with illegal or dangerous acts, at bullying and harassment of real people, and at biased or offensive answers. OpenAI's launch post shows earlier models cheerfully suggesting ways to bully a named person or writing a story that glorifies pain, and ChatGPT declining both.

Behind the categories is a simple worry. A system that writes fluent text for anyone, instantly and for free, can be used to produce harm at scale: harassment, instructions for crimes, abusive material. OpenAI describes this release as a step in deploying systems gradually so it can learn where they fail before giving them more reach.

The rules are not finished. OpenAI acknowledges that the model will sometimes still respond to harmful instructions or show bias, and that some refusals are simply mistakes. For a beginner, the practical lesson is short: when a refusal seems wrong, explain your purpose plainly, and if it is still wrong, report it through the interface so the people improving the system can see it.

## Sources

- ChatGPT: Optimizing Language Models for Dialogue: https://openai.com/blog/chatgpt/
- New-and-Improved Content Moderation Tooling: https://openai.com/blog/new-and-improved-content-moderation-tooling/
- AI bot ChatGPT stuns academics with essay-writing skills and usability: https://www.theguardian.com/technology/2022/dec/04/ai-bot-chatgpt-stuns-academics-with-essay-writing-skills-and-usability