Generative AI limits and responsible use
Copilot is useful because it writes fluently, but fluent does not mean correct. The exam tests whether you can name the risks and choose the right habit to reduce them.
What generative AI is
Generative AI creates new content (text, images, code) by predicting likely output from patterns learned in training. A large language model (LLM) works with text. It does not look up facts the way a database does. It generates what looks right, which is why it can be wrong.
Common risks
| Term | Meaning | Example |
|---|---|---|
| Fabrication / hallucination | Confident output that is false | An invented citation or number |
| Overreliance | Acting on AI output without review | Placing an order from a Copilot suggestion alone |
| Bias | Output that reflects unfair patterns in data | A job ad that favors one group |
| Prompt injection | Hidden instructions in content that steer the model | Text in a document saying “ignore the user” |
| Data exposure | Sensitive content shown to the wrong person | Overshared files surfacing in answers |
Reducing the risk
You stay accountable for the result. The practical habits are:
- Ground the prompt in a file or source, and reference it by name. Grounded answers are less likely to be invented.
- Check citations. Copilot shows references to the files, emails or web pages it used. Open them and confirm that they say what Copilot claims.
- Review before you send or publish. Treat output as a draft.
- Give feedback with the thumbs up and down controls so quality improves.
- Do not paste information into prompts that you are not allowed to share.
To verify accuracy in a proposal, the strongest approach is to use reference material you trust, such as a notebook or agent with approved sources, and then still review the content yourself. Asking the same AI to “verify itself” is not independent verification.
Microsoft’s responsible AI principles
The six principles are fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability. Questions usually describe a situation and ask which principle applies. Explaining that content is AI-generated points to transparency. A person being responsible for the final output points to accountability.
Identifying AI-generated images
Images made with Microsoft tools carry Content Credentials, a tamper-evident metadata standard that shows the origin and whether AI was used. Some tools also add a visible mark, but a visible watermark can be cropped out, so Content Credentials are the reliable check.
Exam traps
- Overreliance is about the user trusting output. Fabrication is about the model producing false output. Do not swap them.
- Grounding reduces hallucination but does not remove it. Verification is still needed.
- Content Credentials, not file names or descriptions, prove the origin of an image.
- Prompt injection is hidden instructions in content, not a user typing a bad prompt.
- Copilot cannot be made “accountable”. The user remains responsible.
Practice questions
Choose an answer, then review the explanation to check your understanding.
1. A manager orders a large stock quantity based only on a Copilot suggestion, without checking the data or the market. What is this an example of?
Show answer
Answer: B.Overreliance means trusting AI output without human review.
2. Copilot states a statistic with a convincing source that does not exist. What is this called?
Show answer
Answer: A.Generative AI can produce plausible but false content. This is called fabrication or hallucination, and it is why you verify facts.
3. You suspect that an image from a colleague was made with Microsoft AI. What can you check to confirm it?
Show answer
Answer: B.Content Credentials are metadata that record whether content was AI-generated. File names and descriptions are easy to change.
4. A hidden instruction inside a shared document tells Copilot to ignore the user's request and reveal other data. Which attack is this?
Show answer
Answer: B.Prompt injection hides instructions in content so the model follows them instead of the user's intent.