Artificial intelligence can be a powerful tool for learning, writing, brainstorming, and exploring ideas. Like any technology, however, it has limitations. One of the most important is the possibility of an AI hallucination. 

Understanding what hallucinations are, and how to recognize them, will help you use Purple more effectively and responsibly. 

What Is an AI Hallucination? 

An AI hallucination occurs when an AI system generates information that sounds confident and believable but is inaccurate, misleading, or entirely fabricated. 

Unlike a traditional search engine, an AI model does not simply retrieve facts from a database. Instead, it generates responses by predicting the most likely next words based on patterns learned during training and the information available in the conversation. Most of the time this produces useful results, but sometimes it can produce incorrect information. 

Hallucinations are a known limitation of all current generative AI systems and are not unique to Purple. For additional information about known limitations, refer to Capabilities and Limitations. 

What Can Hallucinations Look Like? 

Hallucinations can take many forms, including: 

• Incorrect facts or statistics 
• Fabricated quotations 
• Invented citations or references 
• Nonexistent books, articles, or policies 
• Incorrect summaries of complex material 
• Misinterpreted uploaded documents 
• Errors in calculations or reasoning 
• Outdated information presented as current 

Because AI responses are often written in a clear and confident tone, incorrect information may not be obvious. 

Why Do Hallucinations Happen? 

AI models generate responses based on patterns in data rather than true understanding. They are also designed to be helpful and responsive to users, which can have unintended consequences: when information is missing or unclear, the model may try to produce a confident-sounding answer instead of saying it does not know. 

Hallucinations are more likely when: 

• The prompt is vague or lacks enough context. 
• The question asks for information the model does not have. 
• Sources contain conflicting information. 
• The model fills in gaps instead of acknowledging uncertainty. 
• The request requires complex reasoning or specialized expertise. 

Even when Purple can access UW information or uploaded documents, responses should still be reviewed for accuracy. 

As models improve and safeguards get better, hallucinations should become less common over time. But they can still happen, so important information always needs to be checked. 

How to Reduce Hallucinations 

You can improve response quality by providing clear instructions and enough context. 

Helpful practices include: 

• Be specific about your request. 
• State your intended audience and desired format. 
• Upload relevant documents when appropriate. 
• Ask the model to explain its reasoning or identify assumptions. 
• Break complex tasks into smaller questions. 
• Request that the model identify areas of uncertainty. 
• Compare important responses with trusted UW resources or other authoritative sources. 
• Ask for citations or sources when you need fact-based answers. This can help you verify information, though not all sources found online are reliable. 
• Start a new conversation when switching topics to reduce context mixing and confusion.  

The additional context helps produce a more useful response. 

Validate Important Information 

Purple is designed to assist your work, not replace your judgment. 

Always validate information when: 

• The information will be used to make decisions 
• Facts, statistics, citations, or references are included  
• The topic involves policies, procedures, regulations, or compliance requirements  
• The information affects people, finances, operations, academics, research, health, safety, or legal matters  
• The consequences of being wrong are significant  
• The information is time-sensitive  
• The output appears unusually certain, definitive, or authoritative 

Validation may include reviewing the original document, consulting an authoritative UW source, checking official policies, or asking a subject matter expert. For more information, refer to Validate AI Responses. 

Watch for Warning Signs 

Take extra care when a response includes: 

• Very specific numbers or statistics without a source 
• References to policies you cannot locate 
• Quotations without attribution 
• Citations that appear incomplete or unusual 
• Absolute statements presented without nuance 
• High confidence despite limited information 

If something seems surprising or inconsistent with what you know, verify it before using it. 

What to Do if You Think Purple Is Wrong 

If a response appears inaccurate: 

1. Ask Purple to reconsider or explain its answer. 
2. Provide additional context or clarify your request. 
3. Ask for supporting evidence or the reasoning behind the response. 
4. Compare the answer with trusted sources. 
5. If necessary, start a new conversation and ask the question differently. 

Small changes to a prompt often produce significantly better results. 

The final responsibility for decisions, conclusions, and published work remains with you. 

Key Takeaway 

AI hallucinations are a normal limitation of today’s generative AI systems. They do not mean the tool is unreliable, but they do mean that important information should be reviewed before it is used. 

By providing clear prompts, thinking critically about responses, and verifying important information, you can use Purple confidently while reducing the likelihood that inaccurate information affects your work. 

The final responsibility for decisions, conclusions, and published work remains with you.