# Responsible AI, fairness and verification

> Describe who uses the system, who may be affected, and who can correct mistakes. Automation does not remove human responsibility.

Course: [AI foundations and responsible use](https://studybrix.com/en/courses/ai-foundations/index.md)
Level: school
Prerequisites: Basic AI concepts and critical reading.

## Notes

### Accountability

Describe who uses the system, who may be affected, and who can correct mistakes. Automation does not remove human responsibility.

### Bias

Data coverage and measurement choices can cause uneven performance. Examine relevant groups and error types without assuming one aggregate metric settles fairness.

### Verification

Check important factual outputs against reliable sources. Do not invent citations or treat a generated explanation as proof of the model’s internal process.

### Boundaries

Protect private data, respect rights and require appropriate human review for consequential uses. These are design principles, not a guarantee of legal compliance.

## Practice questions

### 1. Does equal overall accuracy guarantee fair outcomes for every relevant group?

A. Yes, by definition
B. Yes, with enough parameters
C. No
D. Only if the output is text

Answer: No

Explanation: Aggregate metrics can hide different error patterns and contextual harms.

### 2. What should happen when an AI output cites a source?

A. Treat the title as the full evidence
B. Assume the citation proves everything
C. Remove every uncertainty statement
D. Verify that the source exists and supports the claim

Answer: Verify that the source exists and supports the claim

Explanation: Citations can be incorrect or fail to support a claim.

### 3. In a toy audit, group A has 8 errors in 40 cases. Find error rate.

Answer: 0.2

Explanation: 8/40=0.2.

### 4. Group B has 3 errors in 30 cases. Find error rate.

Answer: 0.1

Explanation: 3/30=0.1; comparison still needs context and uncertainty.

### 5. Why should a system offer a route for reviewing or correcting mistakes?

Sample answer / marking guidance: Affected people and operators need a way to challenge errors, fix data and prevent repeated harm. Responsibility must be operational, not merely a slogan.

Explanation: Credit a concrete correction and accountability pathway.

### 6. Which is an appropriate privacy practice for a classroom exercise?

A. Assume anonymization is always perfect
B. Use authorized or synthetic data without unnecessary personal details
C. Publish all raw identifiers
D. Upload private records without permission

Answer: Use authorized or synthetic data without unnecessary personal details

Explanation: Minimize exposure and use data appropriately.

### 7. Explain why a generated rationale does not necessarily reveal the actual internal cause of a model's output.

Sample answer / marking guidance: It is itself generated text and may be a plausible explanation rather than a faithful trace of computation. Independent analysis is needed for causal claims.

Explanation: Credit explanation versus verified mechanism.

### 8. Which claim should be avoided without evidence?

A. This test used a defined dataset
B. This metric has limitations
C. This output requires checking
D. This AI system is universally unbiased

Answer: This AI system is universally unbiased

Explanation: Universal fairness claims exceed what finite evaluations establish.

## Further reading

- https://developers.google.com/machine-learning/crash-course
- https://cs229.stanford.edu/

Original AI-assisted educational draft; not independently teacher-reviewed. References support further study, not copied exercises, affiliation or curriculum accreditation.