Responsible AI principles for Copilot

Copilot is a probabilistic tool, not an authority. The exam checks that you know the responsible AI principles and, more importantly, that you treat every suggestion as a draft that a human owns.

The six Microsoft principles

Principle	Meaning
Fairness	The system treats all people equitably and does not favour a group
Reliability and safety	It performs consistently and safely, also in unexpected conditions
Privacy and security	It protects personal data and resists misuse
Inclusiveness	It works for people of different abilities and backgrounds
Transparency	People can understand how it works and its limits
Accountability	People, not the AI, are responsible for outcomes

Match the keyword in the question to the principle. “Understandable” is transparency. “Humans answer for the result” is accountability. “Works for everyone” is inclusiveness.

Limitations of Copilot

Copilot is trained on large amounts of public code and text, so it carries the weaknesses of that data.

• It can suggest code that is wrong, insecure or does not compile.
• It reflects common patterns, so suggestions lean toward the most popular practices and may be homogeneous.
• It can suggest deprecated syntax or APIs. It does not reliably filter them.
• It may not understand your whole codebase. It only sees limited context.
• It supports many languages, but quality is better for languages that are well represented in training data.
• It can reflect bias present in training data.

Reducing bias and risk

Bias is handled during development and during use. Developers of models use diverse data, fairness metrics and human oversight. Users apply review, tests, code scanning and security tooling to what Copilot produces.

Copilot Chat can help explain and fix security issues, for example by pointing out a likely vulnerability and proposing a fix. It does not guarantee that the result is secure.

Human in the loop

Accountability stays with the developer. Good practice is:

1. Read the suggestion before accepting it.
2. Run tests and linters.
3. Run code scanning and dependency checks in the pipeline.
4. Refactor suggested code to match team standards.
5. Never put secrets or personal data in prompts or comments.

Exam traps

• “More training data” is not the answer to bias. Diversity, fairness metrics and oversight are.
• Copilot does not guarantee code that compiles, is secure, or is free of deprecated features.
• Accountability belongs to people, not to the AI or to GitHub.
• Transparency is about understanding how the system works, not about open-sourcing the model.
• Copilot Chat explanations still need validation. They can be confidently wrong.

Practice questions

Choose an answer, then review the explanation to check your understanding.

1. A team wants AI systems that are understandable and give clear information about how they work and what their limits are. Which principle is this?

1. A.Fairness
2. B.Transparency
3. C.Accountability
4. D.Inclusiveness

Show answer

Answer: B.Transparency means people can understand how the system works and what it can and cannot do.

2. Which approach best reduces bias when an AI model is developed?

1. A.Train on the largest possible volume of data
2. B.Optimise only for speed
3. C.Use diverse data, measure fairness, and keep human oversight
4. D.Remove all human review to avoid human bias

Show answer

Answer: C.Bias is reduced with diverse data, fairness metrics and human oversight. More data alone can repeat the same bias at larger scale.

3. A developer accepts a Copilot suggestion that builds a SQL query from user input and ships it without review. What is the main risk?

1. A.Copilot will refuse to run the code
2. B.The code may contain a security flaw such as SQL injection
3. C.The suggestion is always licensed under GPL
4. D.Copilot will delete the original file

Show answer

Answer: B.Suggestions can contain security vulnerabilities because the model reflects patterns from its training data. Review and test every suggestion.

4. Why might Copilot suggest outdated syntax for a framework?

1. A.It reads your private repositories only
2. B.Its training data contains older code and it has a knowledge cut-off
3. C.Deprecated features are always filtered out
4. D.The user disabled content exclusions

Show answer

Answer: B.The model learns from existing code, so it may suggest deprecated syntax. It does not automatically filter or update it.