AIF-C01 guidelines for responsible AI
Guidelines for Responsible AI is worth 14% of AIF-C01’s scored content — around 7 of the 50 scored questions. It is one of the two smallest domains, and it is the cheapest to learn: the concepts are few, they recur predictably, and the expected answers follow a pattern.
The core concerns
| Concern | The question it answers |
|---|---|
| Fairness and bias | Does the system treat comparable people comparably? |
| Transparency | Can people tell how a decision was reached, and that AI was involved? |
| Explainability | Can the reasoning be described in terms a person understands? |
| Robustness | Does it behave safely under unexpected or hostile input? |
| Privacy and security | Is personal data protected, including in outputs? |
| Governance and accountability | Who is answerable for it? |
| Veracity | Is the output actually true, and how would you know? |
Exam framing is almost always: a scenario describes a problem, and you identify which concern it primarily raises. Unequal outcomes between groups → fairness. Nobody can explain a decision → transparency or explainability. Confident falsehood → veracity, addressed by grounding.
Where bias comes from
Worth understanding rather than memorising, because it drives the right answers:
- Training data that under-represents or misrepresents a group
- Historical bias baked into past decisions the model learned from
- Labelling decisions made by people
- Deployment context — a model used on a population it was not built for
The consequence: bias is mostly a data problem, so the mitigations are mostly data and testing mitigations — more representative data, bias testing across groups before and after deployment, and monitoring outcomes in production.
Guardrails
Controls applied around a model rather than inside it: filtering harmful content, blocking disallowed topics, and constraining what the model will discuss. The reason they exist as a separate layer is that you cannot fully control a foundation model’s behaviour by prompting alone.
The human role
The most reliably correct instinct in this domain. When a scenario involves a consequential decision — employment, credit, health, safety, anything affecting rights — the expected answer involves a human with the authority to review and override.
Two related ideas:
- Over-reliance is a named risk. A system that is right most of the time trains people to stop checking.
- Transparency includes telling people they are interacting with AI.
Trade-offs
AIF-C01 expects awareness that responsible AI involves real trade-offs rather than free wins:
- A more accurate model may be less explainable.
- Stricter guardrails reduce harm and also reduce usefulness.
- Collecting more data can improve fairness and worsen privacy.
Questions sometimes present a trade-off and ask which choice fits a stated priority. Read the priority.
Sample questions
Question 1. A loan model approves applicants from one region at a much lower rate than equally qualified applicants elsewhere. Which responsible AI concern does this primarily raise?
- A. Fairness and bias
- B. Explainability
- C. Robustness
- D. Privacy
Show answer
Answer: A
An unequal outcome between comparable groups is a fairness and bias problem. Explainability concerns whether the reasoning can be described, robustness concerns behaviour under unexpected input, and privacy concerns data protection. All may be relevant, but the described harm is unequal treatment.
Question 2. A company deploys an AI system that recommends which employees are selected for promotion. Which safeguard should be in place?
- A. A disclaimer stating the recommendation is generated by AI
- B. A requirement that the model achieve at least 95% accuracy
- C. Human review of recommendations, with authority to override
- D. Restricting access to the output to senior managers only
Show answer
Answer: C
A consequential decision affecting individuals requires human review with genuine authority to override the recommendation. A disclaimer does not change the decision process, higher accuracy does not remove the need for oversight, and restricting access to the output does not address how the decision is made.
Question 3. An assistant repeatedly produces confident answers that are not supported by any company source. Which responsible AI concern is this, and what most directly addresses it?
- A. Fairness; rebalance the training data
- B. Veracity; ground responses in retrieved company content
- C. Robustness; test with adversarial inputs
- D. Privacy; encrypt the model’s outputs
Show answer
Answer: B
Confident but unsupported output is a veracity problem, and grounding responses in retrieved company content is the direct mitigation. Fairness concerns unequal outcomes, robustness concerns hostile input, and privacy concerns data protection — none of which describes stating things that are not true.
What to practise
Write the seven concerns on one card with a one-line scenario beside each, in your own words. Then, for any scenario you meet in practice questions, ask what the described harm actually is before looking at the options. That single habit answers most of this domain, and it takes an evening to build.