GH-300

Microsoft's Fairness principle states that AI systems should treat everyone fairly and avoid affecting similarly situated groups of people in different ways. Inclusiveness is about empowering and engaging everyone, including people with disabilities, rather than equal treatment in outcomes. Privacy and Security covers protecting data, and Reliability and Safety covers consistent, safe operation.

Bias is reduced by training on diverse, representative data, measuring outcomes with fairness metrics across groups, and keeping humans in the loop to review results. Simply collecting more data can amplify existing bias if the data is skewed, and accuracy alone does not guarantee that different groups are treated fairly. Computational efficiency and speed have no effect on bias.

Secure code protects the AI tool and the data it handles from unauthorized access, preventing breaches that could expose sensitive code, prompts or personal information. Maintaining system integrity is a related benefit, but the primary reason for securing code is to stop unauthorized access and data breaches. Handling larger datasets and building more advanced features are not outcomes of code security.

Transparency means people can understand how an AI system makes decisions that affect them, so explaining what content is filtered and how the system reaches its conclusions directly promotes it. Relying on a reputable vendor, updating the algorithm, or measuring user satisfaction may improve trust or quality, but none of them explains how the system works.

Copilot Chat combines the foundation model's knowledge from its training data with context gathered from your code and repositories, and, when enabled, Bing web search results for recent or highly specific topics. Relying only on training data ignores the repository context it gathers, Bing is an optional supplement rather than the primary source, and user-provided documentation is never the exclusive source.

Transparency requires that people understand how AI systems make decisions, with interpretability providing useful explanations of system behavior. Accountability is about the people who design and deploy AI being responsible for how it operates, Fairness concerns equal treatment, and Inclusiveness concerns empowering everyone.

A recognized risk of AI is that its decisions can be hard to interpret, which is why transparency and interpretability are core responsible AI concerns. AI does not eliminate the need for privacy regulation and can in fact perpetuate existing biases, so B and C are false statements. Claiming AI replaces developers in most fields is an overstatement, not an established risk.

Copilot's content filtering detects and blocks harmful content, including hate and unfairness (such as slurs and offensive stereotypes) and sexual content. Code with logical errors is a quality problem, not toxic content, so the filter does not flag it. Comments expressing strong opinions or criticism are not harmful content unless they cross into hateful or offensive language.

Fairness in responsible AI means the system treats all people equitably, which is achieved by identifying and removing bias from the training data and the algorithms that learn from it. Accessibility is a separate responsible AI principle (inclusiveness), so it does not by itself address unfair outcomes. Monitoring performance is good practice but measures quality rather than removing the bias that causes unfair outputs. A larger dataset can carry the same or more bias, so size alone does not make a model fair.

Copilot learns from large amounts of public code, so its suggestions can reproduce insecure patterns and must be reviewed and tested like any other code. For the same reason, suggestions can reflect outdated libraries, deprecated APIs or practices that do not match current standards. Irrelevant suggestions or prompt effort are minor productivity costs that a developer can simply ignore, not the core risks of relying on generated code.

A generative model predicts output from patterns in its training data, so biased, outdated or inconsistent data leads directly to inaccurate or skewed results. Downtime and heavy load affect availability and latency, not the correctness of what the model produces. Models are not deliberately programmed to prefer creativity over accuracy; inaccuracy comes from how they learn from data.

Copilot generates suggestions from patterns learned from existing code, so its statistical code may not suit the assumptions or shape of a particular dataset and must be validated by the developer. Copilot does not automatically find and correct statistical errors, cannot run or verify the significance of your results, and does not invent new, undocumented statistical methods.

In Copilot Chat in Visual Studio, /optimize analyzes and improves the running time of the selected code. It does not translate code to another language, apply formatting rules or rewrite documentation; documentation comments are handled by /doc.

Organizations that keep their code in Azure DevOps can buy Copilot seats for their developers without adopting GitHub Enterprise. GitHub lets you set up an enterprise account that manages only Copilot Business licenses, so members are billed only for Copilot and no GitHub Enterprise Cloud licenses are consumed. Copilot Enterprise requires GitHub Enterprise Cloud, Copilot Individual is a personal plan rather than one an organization manages, and there is no plan called Copilot Teams.

Copilot Business policies are managed by organization owners in the organization's Settings, under Copilot, then Policies (and related settings such as content exclusion). There is no copilot.policy file, in individual repositories or in the .github repository, and Copilot policies are not applied through GitHub Actions workflows.

The Copilot user management REST API includes an endpoint that lists all Copilot seat assignments for an organization, and the Copilot usage metrics endpoints return usage reports for the organization's members. The API never exposes the actual code suggestions a user received, and there is no endpoint that lists members who are not subscribed to Copilot.

Copilot Chat collects feedback directly in the conversation: each response has thumbs up and thumbs down reactions you tap to rate it, which ties the feedback to the exact response. A general feedback page or a social media post loses that context, and the GitHub Mobile Settings menu has no Copilot Chat feedback option.

The /fix slash command asks Copilot Chat to propose a fix for problems in the selected code, such as errors, bugs and programming mistakes. It does not convert pseudocode, it is not a general code generator, and it does not run static analysis tools or start a code review.

For Copilot Business and Copilot Enterprise, GitHub does not use customers' prompts, suggestions or usage data to train its models; this is part of the data protection terms for organizations and enterprises. Individual plans (Copilot Pro, formerly Copilot Individual) leave this to a personal setting rather than excluding it by plan. Copilot Codespace is not a Copilot plan.

In supported IDEs, Copilot Chat can explain selected code and suggest improvements (for example with /explain) and generate unit tests (for example with /tests). Asking about a repository's releases and commits is a Copilot Chat on GitHub.com use case that relies on GitHub's repository data. Planning coding tasks is not one of the core IDE Chat features this question targets.