Safe AI: Practical Foundations for Responsible AI Use

Artificial intelligence is increasingly used to create content, analyze information, support decisions, automate workflows, write software, and interact with organizational data and systems. As AI capabilities expand, particularly with AI agents that can use tools and take actions, safe and responsible use becomes an important responsibility for both technical and non-technical users.

This half-day course introduces the practical foundations of Safe AI: using, developing, deploying, and overseeing AI systems in ways that reduce avoidable harm while preserving their usefulness. Participants will explore common AI risks, including inaccurate outputs, inappropriate disclosure of sensitive information, bias, security threats, unsafe automation, excessive reliance on AI, and poorly controlled agentic behavior.

The course uses accessible workplace examples rather than requiring expertise in machine learning or cybersecurity. Technical concepts are introduced at an accessible level, with optional deeper considerations for participants involved in building, integrating, or administering AI systems.

Learners will have the opportunity to:
1. Explore what "Safe AI" means in an organizational context.
2. Consider how AI risk depends on the task, information, capabilities, users, and consequences, not simply on which model is being used.
3. Examine common risks associated with generative and agentic AI.
4. Explore how hallucinations, misleading outputs, and inappropriate levels of confidence appear in AI responses.
5. Practice techniques for verifying AI-generated information.
6. Discuss approaches for handling confidential, personal, proprietary, and otherwise sensitive information when using AI.
7. Explore best practices for working with Stanford data in AI tools, including Stanford's data risk classifications and approved AI services.
8. Examine common security concerns involving AI systems and AI-enabled workflows.
9. Discuss bias, fairness, and other risks associated with using AI-supported recommendations or decisions.
10. Compare low-risk AI assistance with higher-risk AI actions or decisions.
11. Consider approaches to human oversight of AI-assisted work.
12. Explore the additional risks introduced when AI systems can access tools, files, applications, and external systems.
13. Work with a simple risk-based framework for considering safeguards before using or deploying AI.

Topic Outline:

1. What Does "Safe AI" Mean?
- What Safe AI means in practical terms
- AI safety versus AI security versus responsible AI
- Why AI safety concerns have increased
- Generative AI, predictive AI, and agentic AI
- The difference between AI producing information and AI taking actions
- Safety as a shared responsibility
- Different responsibilities for:
  - Everyday AI users
  - Managers and process owners
  - Developers and technical teams
  - Security, privacy, legal, compliance, and governance functions

2. Understanding AI Risk
- Rather than treating all AI use as equally risky, participants explore how to evaluate the context:
  a. What is the AI being asked to do?
  b. What information does it have access to?
  c. What can it do?
  d. Who will rely on the result?
  e. What happens if it is wrong?

3. Hallucinations, Accuracy, and Reliability
- What an AI hallucination is
- Why fluent responses are not necessarily correct
- Fabricated facts, citations, sources, or calculations
- Missing context and incorrect assumptions
- Out-of-date information
- Confidence versus correctness
- Why AI can produce different answers to similar questions
- Risks of summarizing incomplete source material
- Risks of relying on AI-generated analysis without verification

4. Data Privacy and Confidential Information
- Public versus internal versus confidential information
- Personal and sensitive personal information
- Customer and employee information
- Intellectual property and trade secrets
- Credentials, passwords, and security information
- Contracts, legal material, and regulated information
- Uploaded files and attachments
- Connected applications and repositories
- Organizational policies governing approved AI systems

5. Best Practices for Working with Stanford Data
- Stanford's data risk classifications: Low, Moderate, and High Risk
- Matching data risk levels to Stanford-approved AI tools and services
- Why personal or consumer AI accounts differ from Stanford-licensed tools
- Protected health information (PHI) and HIPAA considerations
- Student records and FERPA considerations
- Research data, data use agreements, and sponsor requirements
- Sharing the minimum necessary: de-identifying, redacting, or summarizing before using AI
- Chat history, model training, and data retention settings
- Where to find current Stanford AI guidance and who to contact when unsure

6. AI Security and Adversarial Risks
- AI systems as a new organizational attack surface
- Malicious or manipulated inputs
- Prompt injection
- Indirect prompt injection from documents, websites, and other sources
- Malicious instructions embedded in content
- Social engineering using AI
- AI-generated phishing and impersonation
- Unsafe code or commands generated by AI
- Excessive permissions
- Connecting AI systems to organizational tools

7. Bias, Fairness, and AI-Supported Decisions
- Where bias can enter AI systems
- Bias in data
- Bias in prompts and task definitions
- Bias in model outputs
- Automation bias in human reviewers
- Why AI recommendations may appear objective even when they are not
- Representativeness and missing populations
- High-impact decisions
- Appropriate human review
- Documenting decision criteria
- Testing AI systems across relevant scenarios

8. Safe Use of AI Agents and Automation
- How the risks change when an AI system can act, not merely answer
- Agentic risk factors
- Access to files
- Access to email and messaging
- Database access
- Ability to modify records
- Ability to execute code
- Ability to make purchases
- External communications
- Repeated or automated execution

9. Human Oversight and Avoiding Over-Reliance
- AI as an assistant rather than an unquestioned authority
- Confirmation bias
- Loss of human expertise through excessive delegation
- When human judgment matters most
- Reviewing outputs versus merely approving them
- Appropriate escalation
- When subject-matter expertise is required
- Maintaining accountability

Pre-requisite:
No technical background is required. General familiarity with generative AI tools such as ChatGPT, Claude, Gemini, Microsoft Copilot, or similar systems is helpful but not necessary.