Copilot makes things up, confidently, plausibly, convincingly. That is a structural property of language models, not a bug. Most of the time results are useful; sometimes something reads perfectly and is wrong.
You are still the boss.
The more specific and verifiable a claim, the more carefully you should check it. General summaries are usually fine; numbers, dates, names, citations, and statistics are highest risk.
Verification is not extra work, it is what makes the output yours. Review before you send or present.
Core principles
- High trust, draft emails you read before sending, brainstorm options, reformat content you know is accurate, summarize meetings you attended.
- Medium trust, summarize documents you have not read, Excel analysis, drafts that represent a commitment, anything sent externally.
- Lower trust, specific statistics or external sources, legal/compliance/HR content, financial projections, content about real people, leadership presentations without deep familiarity.
- For summaries: read source material alongside the summary for high-stakes decisions; ask what key points were omitted.
- For data: spot-check against raw data; verify formulas; ask Copilot how it calculated a result.
- MillerKnoll: work Copilot runs in your M365 tenant; your data is not used to train public Copilot models; colleagues cannot see your Copilot chats. That is not the same as "Copilot cannot see too much."
- Oversharing: if you can open a file, licensed Copilot can use it when Work IQ is on. Before you send a draft, ask whether you would forward those sources to this audience. If a citation points to a document you did not attach, open it or treat the claim as unverified.
- Consumer ChatGPT and consumer Microsoft Copilot stay red for MillerKnoll and customer work. Questions → IT and the AI program.
Go deeper. Getting Started: verification
Check yourself
Which types of Copilot output carry the highest verification risk?
General summaries can be checked at a glance against what you know. Specific claims, numbers, dates, names, citations, can be confidently wrong. Those are the highest-risk outputs. Verify them before you send or present.
Do this in Copilot
How do you decide how much to trust someone else's summary? Apply that standard to Copilot.
Use a real document. Then try:
Flag uncertainty in a summary
Summarize this document. If there are areas where the content is ambiguous or where you are uncertain, flag those specifically rather than guessing.
- Review this email draft. Are there any claims that should be verified before I send it? Any places where the tone could be misread?
- I will use this analysis in a presentation to leadership. What are the weakest points in the argument, and what questions might I get that this does not answer?
Did you run this in Copilot? Mark complete when you have tried it.
RecordedNext lesson: Building Your Copilot Practice →
Navigate: press j for next lesson,k for previous.