GH-300 試験練習問題

Justification

• Transparency directly refers to the requirement that AI systems be understandable, with clear documentation and explanations of their operation, data sources, and decision logic.
• Option C (Transparency) precisely matches this definition and is the principle most often cited in AI ethics frameworks (e.g., EU AI Act, OECD AI Principles).
• Fairness (A) focuses on avoiding bias and ensuring equitable outcomes, which is unrelated to explainability.
• Accountability (B) concerns responsibility for outcomes and the ability to attribute decisions to individuals or organizations, not the clarity of internal workings.
• Inclusiveness (D) emphasizes broad access and representation for diverse user groups, again distinct from the need for clear internal explanations.

Therefore, option C is the only principle that explicitly emphasizes understandability and clear informational disclosure about system functioning.

References

1. European Commission, Proposal for a Regulation on Artificial Intelligence (EU AI Act), 2021. <a target="_blank" href="https://digital-strategy.ec.europa.eu/en/policies/ai-act">https://digital-strategy.ec.europa.eu/en/policies/ai-act</a>
2. OECD, Principles on Artificial Intelligence, 2019. <a target="_blank" href="https://www.oecd.org/ai/principles/">https://www.oecd.org/ai/principles/</a>

Justification

• Option C is the best choice because it highlights a critical misconception that can lead to unmanaged bias in AI deployments. When stakeholders assume that AI algorithms are incapable of perpetuating existing biases, they may forgo bias‑assessment practices, allowing hidden discriminatory outcomes to persist unchecked. This complacency is a well‑documented risk in responsible‑AI frameworks, as it can cause inequitable results and legal exposure.
• Option A overstates AI’s impact on employment. While AI automates certain tasks, it also creates new roles and augments existing ones; the claim that it “replaces the need for developer opportunities in most fields” is an exaggeration and not a primary risk associated with AI use.
• Option B misrepresents regulatory obligations. Data‑privacy regulations (e.g., GDPR, CCPA) remain applicable regardless of AI adoption; AI does not eliminate the need for such compliance, so this is not a valid risk description.
• Option D describes a genuine technical limitation (lack of interpretability), but the question asks for a risk associated with using AI, and the most salient risk emphasized in responsible‑AI guidance is the potential for unchecked bias due to false assumptions of neutrality. Hence, Option C aligns best with the intended answer.

References

• Microsoft Responsible AI resources: <a target="_blank" href="https://learn.microsoft.com/en-us/azure/machine-learning/responsible-ai-overview" rel="noopener noreferrer">https://learn.microsoft.com/en-us/azure/machine-learning/responsible-ai-overview</a>
• European Commission guidance on AI bias and risk mitigation: <a target="_blank" href="https://ec.europa.eu/info/strategy/priorities-2019-2024/digital-transformation/ai-committee_en" rel="noopener noreferrer">https://ec.europa.eu/info/strategy/priorities-2019-2024/digital-transformation/ai-committee_en</a>

Technical Justification

• Option A – Hate speech or discriminatory language
  The Copilot toxicity filter scans user‑provided prompts for language that promotes hate, discrimination, or harassment (e.g., racial slurs, ethnic stereotypes). Such content violates Microsoft’s responsible AI policies and is explicitly blocked or flagged to prevent the generation of harmful outputs.

• Option B – Sexually suggestive or explicit content
  Prompts containing sexual innuendo, pornographic descriptors, or explicit language are also filtered out. The model is designed to refuse or safe‑complete any request that could produce sexually explicit material, aligning with content‑safety standards.

• Option C – Code that contains logical errors or produces unexpected results
  Logical errors or unexpected behavior are not a toxicity concern; they are addressed by standard debugging, testing, or code‑review practices. The toxicity filter does not target syntactic or algorithmic correctness, only potentially harmful language.

• Option D – Code comments containing strong opinions or criticisms
  Expressing personal opinions or criticisms in comments is permissible unless it includes hateful or harassing language. Strong opinions alone do not trigger the toxicity filter; only language that meets the criteria of A or B would be flagged.

Conclusion
The filter’s purpose is to detect and block harmful language—specifically hate speech/discrimination (A) and sexually explicit content (B). Options C and D represent legitimate coding issues or stylistic choices that are handled outside the toxicity‑filter scope.

References

• GitHub Copilot Safety & Moderation Documentation: <a target="_blank" href="https://learn.microsoft.com/en-us/github-copilot/safety">https://learn.microsoft.com/en-us/github-copilot/safety</a>
• Responsible AI principles for Azure AI services: <a target="_blank" href="https://learn.microsoft.com/en-us/azure/ai-services/content-moderation/overview">https://learn.microsoft.com/en-us/azure/ai-services/content-moderation/overview</a>

Why option A is the most appropriate

• Accessibility aligns with fairness – Fairness in AI operation is often expressed as equitable access and usability for all user groups, regardless of ability, language, or context. By designing the tool to be accessible (e.g., supporting screen readers, multilingual interfaces, adjustable input methods), you ensure that the benefits of the AI are not limited to a subset of users, which directly embodies the principle of fairness.
• Compliance with inclusive design standards – Accessibility requirements are codified in standards such as WCAG 2.2 and Section 508. Incorporating these from the operational phase makes the AI output usable by diverse populations, thereby mitigating systematic exclusion that would otherwise constitute unfair treatment.
• Holistic view of fairness – Rather than focusing solely on data bias or performance metrics, accessibility addresses the end‑to‑end user experience, guaranteeing that all stakeholders can interact with and benefit from the AI system on equal footing.

Why the other options are less suitable

• Option B – While reducing bias in training data and algorithms is essential for ethical AI, it does not guarantee fairness in the operational context. An unbiased model can still be delivered through an inaccessible interface, thereby excluding certain users and producing unfair outcomes.
• Option C – Continuous performance monitoring helps detect drift and degradation, yet it primarily serves to maintain accuracy and reliability. Fairness requires more than correct outputs; it demands that those outputs are available and meaningful to all user groups, which monitoring alone does not assure.
• Option D – Larger datasets can improve model coverage, but size alone does not eliminate skewed representations or guarantee equitable access. Over‑reliance on dataset size may propagate existing biases and does not address usability barriers that prevent certain users from receiving the AI’s benefits.

Conclusion

Integrating fairness into AI operation most directly involves ensuring the tool is accessible to all intended users, making option A the best answer. This approach embeds fairness into the very way the AI is experienced, rather than treating it as a secondary technical concern.

References

• Microsoft Responsible AI Dashboard – Accessibility considerations: <a target="_blank" href="https://learn.microsoft.com/en-us/azure/machine-learning/responsible-ai-dashboard" rel="noopener noreferrer">https://learn.microsoft.com/en-us/azure/machine-learning/responsible-ai-dashboard</a>
• WCAG 2.2 Standard (Web Content Accessibility Guidelines): <a target="_blank" href="https://www.w3.org/TR/WCAG22/" rel="noopener noreferrer">https://www.w3.org/TR/WCAG22/</a>

Correct options: A and B

• A – Security vulnerabilities: Copilot draws from publicly available repositories that may contain code with known defects or exploits. When its completions are accepted without thorough review, these insecure patterns can be introduced into production code, increasing the attack surface of the application.
• B – Misalignment with best practices / latest standards: The model is trained on a historical code base and therefore may suggest outdated APIs, anti‑patterns, or coding conventions that have been superseded by newer language releases or industry standards. Accepting such suggestions uncritically can lead to maintainability issues and non‑compliant code.

Why C and D are less suitable

• C – Increase development lead time: While irrelevant suggestions can cause extra iteration, Copilot’s primary design goal is to accelerate coding by providing context‑aware snippets. The occasional mismatch does not fundamentally extend project timelines; the risk is more about quality than schedule.
• D – Decrease velocity via prompt engineering: Effective Copilot usage typically requires minimal prompt effort—developers simply accept or reject suggestions. The learning curve is low, and the tool is integrated into the IDE, so the overhead is generally negligible compared with the productivity gains.

Key Takeaway
The primary concerns when relying on Copilot are security‑related suggestions and deviation from up‑to‑date coding standards, both of which necessitate rigorous code review and validation.

References

• GitHub Copilot Security Overview: <a target="_blank" href="https://docs.github.com/en/copilot/security">https://docs.github.com/en/copilot/security</a>
• Secure coding best practices for Copilot: <a target="_blank" href="https://docs.github.com/en/copilot/security-best-practices">https://docs.github.com/en/copilot/security-best-practices</a>