Appearance
Understanding AI-Generated Content at Work — What You Can and Can't Trust
Applies to: Microsoft Copilot, ChatGPT, Google Gemini, Claude
Article Type: Informational
Last Updated: 2026-07-10
Summary
Artificial Intelligence (AI) tools like Copilot and ChatGPT are incredibly powerful assistants, but they are not infallible search engines. They can, and do, make mistakes with extreme confidence. This guide explains what AI "hallucinations" are, when to trust AI output, and when you must verify the facts yourself.
Prerequisites
- No prerequisites — this is a reference guide for using AI tools safely in the workplace
Instructions
1. Understand How AI Actually Works
Modern AI tools (Large Language Models, or LLMs) do not "think" or look up facts in a database like Google does. Instead, they are essentially highly advanced autocomplete engines.
They predict the most logical next word based on the massive amount of text they were trained on. Because of this, they prioritize sounding plausible over being strictly accurate.
2. Watch Out for "Hallucinations"
When an AI doesn't know the answer, it rarely says "I don't know." Instead, it confidently invents a plausible-sounding answer. This is called an AI hallucination.
Common examples of hallucinations:
- Inventing fake court cases or legal precedents
- Providing incorrect formulas in Excel that look perfectly structured
- Attributing a quote to the wrong executive in a meeting summary
- Making up software features that don't actually exist
3. When to Trust AI (Low-Risk Tasks)
AI is fantastic for tasks where you are providing the context and the cost of a mistake is low, or where you can immediately spot an error.
High Trust / Safe Uses:
- Summarizing text you provide: "Summarize this 10-page PDF I uploaded."
- Changing tone: "Rewrite this email to sound more professional and less angry."
- Brainstorming: "Give me 10 ideas for an all-hands meeting icebreaker."
- Drafting boilerplate: "Write a standard out-of-office message."
4. When to Verify AI (High-Risk Tasks)
If you ask the AI to retrieve facts from outside your immediate document, or if the output will be used for critical business decisions, you must verify the output.
Low Trust / Verify Everything:
- Financial data: "What was our Q3 revenue?" (Verify against actual reports).
- Legal/Compliance: "Does our contract with Vendor X allow early termination?"
- Technical instructions: "Write a PowerShell script to delete old user accounts."
- Performance reviews: Do not let AI write performance reviews unchecked, as it may hallucinate achievements or negative feedback based on generic templates.
5. Follow the "Trust, but Verify" Rule
You are ultimately responsible for the work you produce, even if an AI wrote it.
- Check the citations: If you use Copilot, it provides small numbered links (citations) next to its claims. Always click these to ensure the AI interpreted the source document correctly.
- Do the math yourself: AI models struggle with complex arithmetic. If it summarizes a spreadsheet, spot-check the totals.
- Never upload sensitive data to public AI: If you use a public tool like the free version of ChatGPT, anything you type can be used to train the model. Only use company-approved tools (like Copilot or Enterprise ChatGPT) for confidential company data.
Troubleshooting
WARNING
If you catch an AI tool making a dangerous or highly inaccurate claim about company policy, report it to IT so they can adjust the internal indexing or provide feedback to the vendor.
| Symptom / Error | Potential Cause | Solution |
|---|---|---|
| AI invents a meeting that never happened | Hallucination / Context window error | Provide specific dates when asking for meeting recaps |
| Copilot gives the wrong answer about an internal policy | Outdated HR document in SharePoint | Copilot reads all files it has access to. IT needs to archive outdated files |
| Public AI leaks company data | User uploaded confidential files to a free AI | Never use free public AI tools for company business. See Understanding Which AI Tools Are Safe to Use at Work |
| AI code or formula causes an error | AI hallucinated a function | Always test AI-generated code in a safe environment first |