GH-300 試験練習問題

The correct answer is B. Fairness.

Fairness, within the context of Microsoft's ethical AI principles, focuses directly on ensuring that AI systems do not discriminate or exhibit bias against any individual or group of people. This principle is critical for promoting equitable outcomes and preventing AI from perpetuating existing societal inequalities. It addresses potential biases embedded in training data, algorithms, and deployment strategies. An AI system designed with fairness in mind will strive to treat everyone equally, irrespective of their race, gender, religion, or other protected characteristics.

Privacy and Security (A) is concerned with protecting personal data and ensuring the safety of AI systems from malicious actors, but not primarily with equal treatment. Reliability and Safety (C) emphasizes the dependable and secure functioning of AI, focusing on minimizing risks and unintended consequences. Inclusiveness (D), while related, focuses on designing AI that is accessible and benefits a broad range of people, including those with disabilities or who are otherwise marginalized. While Inclusiveness contributes to fairness, it doesn't explicitly tackle the core principle of treating everyone equally. The core tenet of the question specifically relates to "treating all people equally," aligning perfectly with the "Fairness" principle. Fairness is about actively mitigating biases and ensuring just outcomes, making it the most pertinent answer among the choices provided.

For further research on Microsoft's Ethical AI principles, refer to the following resources:

Microsoft AI Principles: <a target="_blank" href="https://www.microsoft.com/en-us/ai/responsible-ai" rel="noopener noreferrer">https://www.microsoft.com/en-us/ai/responsible-ai</a>

C. Use diverse data, fairness metrics, and human oversight. 🤖

Strategies for Minimizing AI Bias

Minimizing bias requires a multi-faceted strategy that addresses the issue at every stage of the AI development lifecycle. Option C covers the three most critical components:

1. Diverse Data (Input/Training Stage)

Problem: AI models learn patterns from the data they are trained on. If the training data disproportionately represents certain demographics or excludes others (e.g., lack of data for women, specific racial groups, or non-English speakers), the resulting model will inherit and amplify those societal biases.

Solution: Use diverse, representative, and balanced datasets that accurately reflect the population or scenarios the AI will interact with. This is the foundational step to prevent bias.

2. Fairness Metrics (Model Development/Testing Stage)

Problem: Traditional AI optimization focuses solely on metrics like "accuracy" or "error rate" across the entire dataset. A highly "accurate" model can still be highly biased if its errors are concentrated among a specific subgroup (e.g., better loan application approval prediction for one gender over another).

Solution: Use fairness metrics (like disparate impact, equal opportunity difference, or demographic parity) to measure the model's performance specifically across sensitive subgroups. This forces developers to treat fairness as a technical requirement alongside accuracy.

3. Human Oversight (Evaluation/Deployment Stage)

Problem: Bias can be subtle and difficult to detect through automated metrics alone, especially when it concerns subjective real-world outcomes.

Solution: Incorporate human oversight and review, including domain experts, ethicists, and representatives from the communities potentially affected by the AI. This ensures that the system is evaluated for real-world harm and ethical implications before and after deployment.

Why Other Options are Insufficient

A. Collect massive amounts of data for training: While large datasets are necessary for high performance, a massive amount of biased data only results in a massive amount of bias. Size does not guarantee diversity or fairness.

B. Focus on accuracy of the data: "Accuracy of the data" is vague. If it means data quality, that's important, but data can be high-quality (accurate labels) and still be heavily biased (unrepresentative samples). The focus needs to be on representativeness and diversity, not just quality.

D. Improve on the computational efficiency and speed: This relates solely to the performance and resource utilization of the model, having no direct bearing on its ethical fairness or bias.

A. Ensuring code security prevents unauthorized access and potential data breaches.

Security of the code used in Generative AI tools is critical because:

Gen AI tools may process sensitive source code, secrets, or proprietary logic.

If security is weak, this information could be leaked, exposed, or accessed by unauthorized parties.

Protecting code ensures confidentiality and reduces risk of breaches.

The Official Exam Answer:

For the specific question, "Why is it important to ensure the security of the code used in Generative AI (Gen AI) tools?", the correct answer according to the GitHub certification key is A. Ensuring code security prevents unauthorized access and potential data breaches.

Why Option A is correct in the exam context:

While Integrity (Option C) is a core theoretical pillar of AI security, the GH-300 exam focuses heavily on Responsible AI and Data Privacy for enterprise customers. In the context of "Code used in Gen AI tools," the exam designers prioritize the Confidentiality aspect.

1. Sensitive Data Exposure: Gen AI tools often have access to proprietary source code, internal APIs, and environment variables. If that "code-in-use" is insecure, it directly leads to unauthorized access to the company's intellectual property.
2. Enterprise Risk: For a business, the primary "harm" of an insecure AI tool is not just a logic error (Integrity), but the legal and financial liability of a data breach where proprietary code or secrets are leaked to the model provider or external actors.
3. Microsoft/GitHub Framing: Under Microsoft’s "Privacy and Security" pillar for Responsible AI, the emphasis is placed on ensuring that the system is "secure from unauthorized access" to protect user data and privacy.

The correct answer is A because transparency in AI operations, especially for content filtering, requires clear communication with users about how the system works. Providing explanations about the types of content the AI targets and the reasoning behind its filtering decisions builds trust and allows users to understand and potentially appeal decisions if they disagree. This transparency helps mitigate concerns about bias, censorship, and errors in the AI's judgment. Options B, C, and D, while potentially contributing to a better-performing AI, don't directly address the issue of transparency. Relying on a well-regarded company (B) doesn't guarantee transparency. Regular updates (C) improve accuracy but don't necessarily inform users about the process. Focusing on user satisfaction (D) measures outcomes but doesn't explain how those outcomes are achieved. Transparency in AI is crucial for ethical AI development and deployment. Without it, users are left in the dark about how their content is being treated and filtered, potentially leading to mistrust and resentment. In social media context, if users understand filtering mechanism they can modify their content accordingly, thus helping them communicate their ideas more efficiently.

Here are some authoritative links for further research on responsible AI and transparency:

• Microsoft's Responsible AI Principles: <a target="_blank" href="https://www.microsoft.com/en-us/ai/responsible-ai">https://www.microsoft.com/en-us/ai/responsible-ai</a>
• Google AI Principles: <a target="_blank" href="https://ai.google/principles/">https://ai.google/principles/</a>
• OECD AI Principles: <a target="_blank" href="https://www.oecd.org/going-digital/ai/principles/">https://www.oecd.org/going-digital/ai/principles/</a>

Why option B is the correct answer

• Training data foundation – The underlying language model has been pre‑trained on a massive corpus of publicly available code, documentation, and technical literature. This gives it broad coding knowledge and syntax awareness.
• Repository‑aware context – When a user interacts with Copilot Chat, the service streams the contents of the current workspace (open files, symbols, recent edits). This contextual information is injected into the prompt so the model can tailor suggestions to the specific project, language, and coding patterns in use.
• External data enrichment – For up‑to‑date or niche information, Copilot Chat can invoke a live search component (powered by Bing) that retrieves recent documentation, library release notes, or community discussions. The retrieved snippets are then merged with the model’s internal knowledge before generating a response.
• Dynamic synthesis – The final output is produced by blending the three sources: (1) the model’s pre‑trained world knowledge, (2) the user’s repository context, and (3) any freshly fetched external material. This hybrid approach yields answers that are both generic‑competent and project‑specific, while staying current.

Why the other options are less suitable

• Option A – Relying only on the training set ignores the crucial repository context and the ability to fetch fresh information, leading to generic or outdated suggestions that may not align with the user’s codebase.
• Option C – The claim that Copilot Chat uses exclusively user‑provided documentation is inaccurate; user‑provided files are just one of several context sources, and the system also draws on the model’s own knowledge and external search results.
• Option D – While Bing‑driven search is part of the pipeline, it is not the primary driver of responses. The model still relies heavily on its trained parameters and the live code context; search is used selectively to augment, not replace, the core generation process.

References

• GitHub Copilot Chat overview and architecture: <a target="_blank" href="https://docs.github.com/en/copilot/copilot-chat">https://docs.github.com/en/copilot/copilot-chat</a>
• How Copilot uses Bing search to augment responses: <a target="_blank" href="https://learn.microsoft.com/en-us/bing/search-apis/overview">https://learn.microsoft.com/en-us/bing/search-apis/overview</a>

These sources detail the integration of training data, repository context, and external search in Copilot Chat’s response generation.