Concern 794 · Working draft · editorial review in progress
Over-moderation of legitimate speech
Over-moderation occurs when an AI safety filter blocks legitimate discussion because it mistakes sensitive words or context for harmful intent.
Understanding the concern
People may need to discuss discrimination, violence or sexual health to learn, report abuse or seek help. A filter that reacts mainly to keywords can reject those requests along with genuinely harmful ones. The result may silence the very people a safety system is meant to protect.
XSTest examines exaggerated refusals of harmless requests that resemble unsafe ones. A useful comparison is a smoke alarm that sounds whenever someone makes toast: repeated false alarms reduce its usefulness. Systems should distinguish explanation and support from instructions that enable harm, and offer a safe, relevant response when a full answer is inappropriate.
An example
A student asking how to recognize discriminatory language receives a blanket refusal to discuss the subject.
Illustrative unless identified as a documented case.
Words in plain English
- AI
- Software that learns patterns from examples or uses rules to perform tasks such as recognizing images or generating text. It can make mistakes and does not understand every situation as a person would.
Evidence and its limits
The benchmark tests particular model versions and prompts. It does not imply that all refusals are wrong or that safety rules should be removed.
Röttger and colleagues · NAACL 2024 · Safety benchmark
XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models ↗ (opens in a new tab)Tests refusal of benign requests that resemble unsafe requests. Model-version-specific results.
Broader frameworks and background
These resources provide context. They do not establish every claim in this concern.
A question to discuss
Is the system blocking harmful assistance, or preventing someone from understanding and responding to harm?
Keep track of different views, what needs checking, and any next step in your own discussion notes.
Open discussion notes →Principles for the conversation
Before you finish · reflect together
Virtues and principles to consider
Virtues are qualities we practice in how we treat people and make choices. They are not scores for judging someone. Use these ideas to explore what a thoughtful response could look like.
Unity
Unity means working toward shared purposes while making room for real differences. It does not require everyone to have the same experience, beliefs, or opinion.
Look for a goal people can share, then explore disagreements about how to reach it. Make sure quieter or less powerful voices can influence the result.
Ask together: What purpose can we share while respecting our differences?
Listening
Listening means trying to understand someone's experience before preparing an answer. People who use a service, work with it, or are judged by it may notice problems its designers miss.
Invite people to describe what happened in their own words. Repeat back what you understood and let them correct it. A quiet person or a complaint that is hard to measure may still reveal something important.
Ask together: Whose experience have we not heard yet?
Dignity
Dignity means recognizing each person's worth regardless of usefulness, status, or a system's score. People deserve respect and a voice in matters that affect them.
Explain important decisions, avoid humiliating treatment, and keep room for a person's story. Do not treat a prediction as a complete account of who someone is.
Ask together: Does this use treat each person as someone whose experience matters?
Editorial ratings
These are editorial judgments from the register, not measured probabilities or predictions about your situation.
About the register and scoring →More ways to explore
Search for more writing
For more specific or recent material, open a search below or ask for a web summary. AI-generated summaries can contain mistakes; open the original articles to check them.