What is algorithmic bias?

D. Systematically unfair outcomes for certain groups of people

What is a common source of bias in AI systems?

A. Training data that does not represent all groups well

What does explainability mean?

B. Understanding why a model produced a particular output

For which decision is human review most important?

C. Approving or rejecting a loan application

Before sending customer messages to an external LLM, what is good practice?

D. Remove or mask personal data that the task does not need

What is red teaming in AI?

A. Deliberately trying to make a system fail or behave unsafely to find weaknesses

How can you check a model for fairness?

B. Compare error rates and outcomes across different groups

What is a model card?

C. A document describing a model's intended use, limitations and evaluation results

D. Synthetic media that makes it look as if a real person said or did something

What does data minimisation mean?

D. Collecting and keeping only the data you actually need

Why should users be told when they are talking to an AI system?

A. So they can judge the information appropriately and choose to reach a human

An AI tool gives a confident medical suggestion. What is the right response?

B. Verify it with a qualified professional, because AI can be wrong

Which approach does the EU AI Act take to regulating AI?

C. A risk-based approach, with stricter rules for higher-risk uses

What is a copyright concern with generative AI?

D. Outputs may closely reproduce protected material

Which is a strong defence against prompt injection in an AI agent?

A. Treat external content as data and require approval for risky actions

Responsible AI

Responsible AI and Safety Practice Test

Test your knowledge of fairness, privacy, transparency, safety testing and the risks every AI practitioner and user should understand.

beginner

AI Skills

Before you start

Questions: 15
Time: 15 minutes
Format: Single section

The timer starts when you press Start. When it reaches zero, your answers are submitted automatically.
Use the numbered navigator to move between questions, and flag any you want to come back to.
For typed answers, keep within the word limit. Answers over the limit are marked wrong, as in the real test.
You can submit early. Your score, the correct answers and explanations appear straight away.

Responsible AI and Safety Practice Test: all 15 questions with answers

Prefer to work through the test on paper? Answer the questions, then open “Show answer” to check.

Choose the correct answer, A, B, C or D.

Q1.What is algorithmic bias?
- A.A model that runs slowly
- B.A model with too many layers
- C.A coding style preference
- D.Systematically unfair outcomes for certain groups of people
Show answer
Answer: D. Systematically unfair outcomes for certain groups of people
Bias means some groups are treated worse than others by the system's outputs.
Q2.What is a common source of bias in AI systems?
- A.Training data that does not represent all groups well
- B.Using a fast computer
- C.Writing comments in code
- D.Using English variable names
Show answer
Answer: A. Training data that does not represent all groups well
If some groups are under-represented or mislabelled in the data, the model learns those gaps.
Q3.What does explainability mean?
- A.Making the model larger
- B.Understanding why a model produced a particular output
- C.Translating the model into another language
- D.Hiding how the model works
Show answer
Answer: B. Understanding why a model produced a particular output
Explanations help users trust decisions and help developers find errors.
Q4.For which decision is human review most important?
- A.Suggesting a song
- B.Choosing a font
- C.Approving or rejecting a loan application
- D.Sorting emails by date
Show answer
Answer: C. Approving or rejecting a loan application
High-stakes decisions that affect people's lives need human oversight.
Q5.Before sending customer messages to an external LLM, what is good practice?
- A.Add more personal data for context
- B.Post the messages publicly
- C.Skip any privacy review
- D.Remove or mask personal data that the task does not need
Show answer
Answer: D. Remove or mask personal data that the task does not need
Data minimisation and redaction reduce privacy risk.
Q6.What is red teaming in AI?
- A.Deliberately trying to make a system fail or behave unsafely to find weaknesses
- B.Painting servers red
- C.Training two models together
- D.Choosing a team leader
Show answer
Answer: A. Deliberately trying to make a system fail or behave unsafely to find weaknesses
Red teams probe for harmful outputs, jailbreaks and security holes before attackers do.
Q7.How can you check a model for fairness?
- A.Measure only overall accuracy
- B.Compare error rates and outcomes across different groups
- C.Count the parameters
- D.Check the file size
Show answer
Answer: B. Compare error rates and outcomes across different groups
A model can look accurate overall while performing much worse for one group.
Q8.What is a model card?
- A.A payment card for API usage
- B.A graphics card
- C.A document describing a model's intended use, limitations and evaluation results
- D.A business card for the developer
Show answer
Answer: C. A document describing a model's intended use, limitations and evaluation results
Model cards make limitations visible to the people who use or deploy the model.
Q9.What is a deepfake?
- A.A model with many layers
- B.A fake training dataset
- C.A type of encryption
- D.Synthetic media that makes it look as if a real person said or did something
Show answer
Answer: D. Synthetic media that makes it look as if a real person said or did something
Deepfakes can be used for fraud and misinformation, so detection and disclosure matter.
Q10.What does data minimisation mean?
- A.Compressing data files
- B.Deleting the model after use
- C.Using the smallest model
- D.Collecting and keeping only the data you actually need
Show answer
Answer: D. Collecting and keeping only the data you actually need
Less data collected means less to leak or misuse.
Q11.Why should users be told when they are talking to an AI system?
- A.So they can judge the information appropriately and choose to reach a human
- B.It is never necessary
- C.To make the system slower
- D.To collect more data
Show answer
Answer: A. So they can judge the information appropriately and choose to reach a human
Transparency is a basic expectation, and in some places a legal requirement.
Q12.An AI tool gives a confident medical suggestion. What is the right response?
- A.Follow it without question
- B.Verify it with a qualified professional, because AI can be wrong
- C.Share it widely
- D.Ignore all medical advice
Show answer
Answer: B. Verify it with a qualified professional, because AI can be wrong
Models can hallucinate, so high-stakes advice needs expert checking.
Q13.Which approach does the EU AI Act take to regulating AI?
- A.It bans all AI
- B.No rules at all
- C.A risk-based approach, with stricter rules for higher-risk uses
- D.Rules only for video games
Show answer
Answer: C. A risk-based approach, with stricter rules for higher-risk uses
The Act sorts AI uses into risk levels, with the strictest obligations for high-risk systems.
Q14.What is a copyright concern with generative AI?
- A.Generated text is always original
- B.Images cannot be generated
- C.Copyright only applies to software
- D.Outputs may closely reproduce protected material
Show answer
Answer: D. Outputs may closely reproduce protected material
Check licences and review outputs before publishing them commercially.
Q15.Which is a strong defence against prompt injection in an AI agent?
- A.Treat external content as data and require approval for risky actions
- B.Give the agent admin access to everything
- C.Let web pages give the agent new instructions
- D.Turn off all logging
Show answer
Answer: A. Treat external content as data and require approval for risky actions
Separating instructions from untrusted data and limiting permissions contain the damage.

Responsible AI and Safety Practice Test

Before you start

Responsible AI and Safety Practice Test: all 15 questions with answers

Q1.What is algorithmic bias?

Q2.What is a common source of bias in AI systems?

Q3.What does explainability mean?

Q4.For which decision is human review most important?

Q5.Before sending customer messages to an external LLM, what is good practice?

Q6.What is red teaming in AI?

Q7.How can you check a model for fairness?

Q8.What is a model card?

Q9.What is a deepfake?

Q10.What does data minimisation mean?

Q11.Why should users be told when they are talking to an AI system?

Q12.An AI tool gives a confident medical suggestion. What is the right response?

Q13.Which approach does the EU AI Act take to regulating AI?

Q14.What is a copyright concern with generative AI?

Q15.Which is a strong defence against prompt injection in an AI agent?