San Francisco's Vision for Gen AI in Civic Life

San Francisco, a global hub for Artificial Intelligence, is not just witnessing the AI revolution—it's actively shaping its responsible integration into civic life. With comprehensive guidelines issued on July 7th, 2025, the City and County of San Francisco is embarking on a forward-thinking journey to harness Generative AI (GenAI) for public good, aiming to significantly enhance the effectiveness, efficiency, and responsiveness of City services for all San Franciscans. This progressive approach, rooted in a "trust but verify" philosophy, balances the immense potential of GenAI with a strong commitment to public trust, data integrity, and ethical governance.

San Francisco Leads the Way: Charting a Responsible Course for Generative AI in City Services

The City recognizes that while GenAI tools can create new content like text, images, and code rapidly, offering unprecedented opportunities, they also pose unique risks such as inaccuracy, bias, misinformation, privacy breaches, and cybersecurity threats. Over-reliance on GenAI for decisions affecting public rights or safety could further erode transparency and accountability. To navigate these complexities, San Francisco has established a clear framework.

The Foundation: San Francisco's Top 5 Guiding Principles

At the heart of San Francisco's strategy are five foundational principles designed to ensure that innovation goes hand-in-hand with accountability:

• Ultimate Responsibility Rests with You: Whether content is created by AI or a human, the individual using or sharing it bears full responsibility.

• Secure Tools are Paramount: The City strongly encourages the use of secure, City-vetted enterprise tools like Copilot Chat, specifically discouraging public or consumer GenAI tools. Crucially, sensitive, confidential, or City data should never be entered into unvetted public tools, as these tools may store or reuse the information.

• Verify, Verify, Verify: All AI-generated content must be meticulously reviewed, edited, fact-checked, validated, and/or tested, as it is not always accurate or unbiased.

• Transparency Through Disclosure: When GenAI substantially contributes to public-facing or sensitive work, its use must always be disclosed. This involves proper recording in the City’s 22J inventory and providing direct notice to impacted individuals.

• Deepfakes Are Strictly Prohibited: GenAI must never be used to create misleading "deepfakes"—fake images or recordings—or any content that could be mistakenly interpreted as real, even with disclaimers.

A Layered Approach: Risk-Based Guidelines for Responsible Use

San Francisco's guidelines adopt a practical, risk-tiered approach to GenAI implementation, ensuring appropriate safeguards for every scenario.

Low-Risk Use: Boosting Internal Efficiency

For internal tasks utilizing City-procured Enterprise Generative AI Tools, the risk is considered low, focusing on improving efficiency and reducing workload. Examples include drafting internal emails or memos, summarizing documents or meetings, or writing/debugging code. While helpful, vigilance remains key: users must be familiar with the material, always double-check factual claims and references, and critically review outputs before use or sharing. For these internal support tasks, formal disclosure is not required, but individual responsibility for content remains absolute.

Medium- to High-Risk Use: Navigating Public-Facing and Sensitive Work

When GenAI assists with tasks impacting public communication, services, or decisions, extra caution and additional steps are mandated. This includes drafting public-facing content, supporting hiring processes, or aiding policy-related data summaries. Critically, AI should never make final decisions that affect individuals or public outcomes. For these scenarios, users must possess deep subject-matter expertise to critically review outputs for errors, bias, and harmful implications. Outputs must reflect the City’s values, promote equity, and adhere to ethical and digital accessibility standards, including the use of inclusive language.

Transparency is paramount here: Public-facing or sensitive GenAI use requires formal documentation via the City’s 22J process and a direct, plain-language, multilingual notice to impacted individuals. This notice, as per a sample template provided, must include the tool's name and version, confirm City staff review, and provide contact information for questions or appeals. Furthermore, any AI-generated content (text, image, data) incorporated into work must be cited appropriately, similar to any external source.

Strictly Prohibited Uses

To safeguard public trust and safety, certain GenAI uses are outright forbidden. These include relying on AI to create City official documents or make decisions without expert human review, generating images, audio, or video that could be mistaken for real people (specifically creating "deepfakes" or impersonations, even with disclaimers), fabricating fictional survey respondents or public input, or relying on AI to review legal or regulatory issues.

Navigating the Digital Landscape: Public vs. Enterprise Tools

The City strongly advocates for its secure, City-approved Enterprise Generative AI Tools, such as Microsoft Copilot, ChatGPT Enterprise, and Snowflake Cortex, which are procured, configured, and vetted for City use, with strict privacy, security, and data retention standards. In stark contrast, the use of free or Public/Consumer Generative AI Tools is strongly discouraged and requires prior departmental approval for any experimental use. The primary concern is their inadequate privacy protections, as information shared often trains underlying models, posing significant confidentiality risks. Sensitive or protected data, including personal information, health information, City data, and/or financial information, must never be entered into these unapproved tools, as it can become publicly accessible and part of the public record.

Data at the Core: Protection and Governance

Robust data protection and governance are fundamental to San Francisco's GenAI strategy. Enterprise tools like Copilot Chat and Snowflake can handle higher levels of sensitive data (Level 4 and below), provided Business Associate Agreements (BAAs) are in place for Protected Health Information (PHI). All content—inputted or generated—may constitute public records subject to the California Public Records Act and City retention requirements. The City emphasizes that Generative AI tools depend on accurate, clean, and well-organized data. Therefore, departments are tasked with keeping records updated, using standardized formats, and managing data meticulously to ensure reliable and fair AI usage.

Leadership and Future Development: A Collaborative Path Forward

Departmental IT leaders play a pivotal role in guiding GenAI adoption, focusing on solving public problems rather than just deploying technology. They are responsible for ensuring compliance with transparency requirements under Chapter 22J of the Administrative Code and coordinating with the City's Emerging Technology Team. This involves diligent vendor information collection for new tools, designating 22J Leads for compliance, and encouraging staff training to understand AI risks and functionalities.

San Francisco's GenAI guidelines are not static; they are a living document. Developed by the Emerging Technology Team in coordination with the City’s AI Advisory Committee, these guidelines will be regularly updated to reflect new laws, regulations, lessons learned, and advancements in GenAI technology. This iterative approach underscores San Francisco’s commitment to leading with responsible and accountable governance in the civic use of AI, striving to deliver more effective, efficient, and responsive services to all San Franciscans.