Responsible AI (RAI) while using Copilot

In the Privacy and Security Guidance for using Microsoft 365 Copilot article, we introduced the S.A.F.E-A.I. principles — a way to help you engage with AI technologies responsibly and confidently. In this article, we’ll take a closer look at each principle, explore what they mean in practice and highlight how they’re being applied at SFU. By understanding and applying these principles, you’ll build the confidence and critical thinking skills needed to evaluate AI tools thoughtfully, ask the right questions and make responsible choices in your work.

What's Included

• Secure (Privacy & Security)
• Accountable
• Fair
• Explainable
• Auditable (Safety & Reliability)
• Inclusive

Secure(Privacy & Security)

One of the most pressing aspects of responsible AI is addressing privacy and security concerns. As AI systems, particularly generative AI, become more powerful, they introduce new risks to data security, personal privacy and even broader cybersecurity. There is a lot to consider in this section, so we'll outline some of the key concerns and challenges.

Data Privacy and Unintended Exposure

AI systems often train on large data sets that include personal or proprietary information. A major concern is unintended data leakage – models might inadvertently expose sensitive information that was in the training data. For instance, if employees input confidential text into a generative AI service, that data might get used to further train the model and later emerge in another user’s output. 

Generative AI models can “remember” fragments of their training data, raising the risk that private information (like a person's contact info or proprietary code) could be regurgitated in responses. As a precaution, many organizations limit what data can be fed into public AI tools and AI providers are working on techniques to better filter or forget sensitive training data.

Intellectual Property and Copyright

Generative AI raises questions about intellectual property (IP). Models trained on vast internet data might output content that unintentionally copycats copyrighted material. For example, a generative model might produce a passage of text almost identical to something in its training set (e.g., from a book or article), or an image model might generate something that resembles a specific artist’s work. This blurring of originality complicates who owns AI-generated content and whether using certain training data infringes copyrights. Organizations deploying generative AI must be careful to avoid legal issues from AI output. Responsible AI use entails respecting IP rights, perhaps by curating training data or allowing authors to opt out and by guiding users to use AI outputs ethically (not passing off AI-generated art as human-made if that violates terms, for instance).

Regulatory Compliance

Privacy laws like FIPPA (in B.C., Canada) or GDPR (in Europe) and various data protection regulations worldwide impose requirements on automated decision-making and data handling. Organizations using AI need to ensure compliance, such as providing explanations for decisions, obtaining consent for personal data processing and enabling individuals to opt out of purely AI-based decisions. Future AI-specific regulations (like the proposed Canadian Artificial Intelligence Data Act) will likely enforce certain privacy and transparency standards. Staying ahead of these governance requirements is a part of responsible AI strategy for any institution.

Accountable

Fair

AI systems should treat individuals and groups fairly, avoiding discrimination. This means the AI’s decisions or outputs should not be prejudiced against people on the basis of characteristics like race, gender, or other protected attributes. A known challenge is that AI models can inherit biases present in their training data. Mitigating bias requires careful data set curation and ongoing evaluation. Techniques include using diverse training data, testing outcomes for disparate impact and correcting any biased decision rules.

Explainable

There is a need for transparency in how AI systems make decisions or produce outputs. For users and organizations to trust AI, they should be able to get some insight into what data was used and how the model works. This is difficult with complex models (the “black box” problem), but efforts like Explainable AI (XAI) aim to provide human-understandable explanations for AI decisions.

Transparency also involves being clear when content is generated or consumed by AI. Being transparent helps ensure accountability and allows individuals to contest or question an AI outcome.

Auditable(Safety & Reliability)

AI should be reliable and safe, meaning it performs as intended under expected conditions and fails gracefully under unexpected conditions. For instance, if an AI system controls a vehicle or provides medical advice, it must undergo rigorous testing and auditing to handle edge cases and avoid dangerous failures.

Safety also involves the AI not causing harm through its output. In generative AI, safety can mean preventing the AI from generating violent or illegal instructions, hate speech, or disinformation.

Inclusive

AI should be designed to empower as many people as possible and avoid marginalizing any group. This principle emphasizes user-centric design – making AI accessible and useful to people with diverse backgrounds and abilities. For example, inclusive AI might involve ensuring that a voice assistant works for different accents and speech patterns, or that generative AI tools are available in many languages. In the context of responsible AI, inclusivity overlaps with fairness, but also encourages broad stakeholder input when designing AI (so that many perspectives are considered).