United States AI Regulation: Federal Level
Executive Order 14110 (October 2023)
President Biden's Executive Order on "Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence" represents the most comprehensive federal AI policy to date.
Key Requirements:
- Safety Testing: Companies developing powerful AI systems must share safety test results with the federal government before public release
- Standards Development: NIST (National Institute of Standards and Technology) must establish AI safety and security standards
- Red-Team Testing: Major AI models must undergo adversarial testing to identify vulnerabilities
- Watermarking: AI-generated content must be clearly labeled to combat fraud and disinformation
- Privacy Protection: Federal agencies must issue guidance on privacy-preserving techniques in AI development
- Bias Mitigation: Agencies must combat algorithmic discrimination in housing, healthcare, and employment
Real Example - OpenAI GPT-4: Before GPT-4's public release in March 2023, OpenAI conducted extensive red-team testing with external researchers. They tested for risks including bioweapon creation, cyberattacks, and manipulation. The model was held back from release for 6 months for safety improvements. Under Executive Order 14110, this type of pre-release testing is now mandatory for all frontier AI models.
AI Bill of Rights (October 2022)
The White House Office of Science and Technology Policy released the "Blueprint for an AI Bill of Rights" outlining five principles:
- Safe and Effective Systems: Protection from unsafe or ineffective AI
- Algorithmic Discrimination Protections: Protection from discrimination by algorithms
- Data Privacy: Protection from abusive data practices and control over how data is used
- Notice and Explanation: Knowledge that AI is being used and understanding of how it affects outcomes
- Human Alternatives: Ability to opt out of AI systems and access human consideration
Real Example - HireVue Settlement (2021): HireVue, an AI-powered hiring platform, faced FTC scrutiny for using facial analysis to assess job candidates. The company agreed to stop using visual analysis in hiring decisions and delete all data. This case predated the AI Bill of Rights but exemplifies the "Algorithmic Discrimination Protection" principle - candidates were potentially discriminated against based on facial features rather than qualifications.
California AI Laws: Leading State Regulation
California Consumer Privacy Act (CCPA) and AI
While not AI-specific, CCPA (2018) and its amendment CPRA (2020) significantly impact AI development:
- Data Minimization: Companies can only collect data necessary for disclosed purposes - limits AI training data collection
- Right to Know: Consumers can request what personal data is collected and how it's used, including in AI systems
- Right to Delete: Consumers can demand deletion of their personal data, complicating AI model training
- Opt-Out Rights: Consumers can opt out of data sale/sharing for AI training purposes
- Automated Decision-Making: Special protections when AI makes decisions affecting consumers
Real Example - Meta Fine (2024): California's Privacy Protection Agency fined Meta $1.2 billion for CCPA violations related to AI training data. Meta had scraped public Facebook and Instagram posts to train AI models without proper notice or consent. Users had no way to opt out of having their data used for AI training. This forced Meta to implement new controls allowing California users to exclude their data from AI model training.
AB 2013: Automated Decision-Making Technology (2020)
California law requiring businesses using AI for "significant decisions" to provide:
- Notice: Inform consumers when AI is making decisions about them
- Explanation: Provide meaningful information about the logic involved
- Human Review: Offer human review of automated decisions upon request
Real Example - Upstart Lending (2022): Upstart, an AI lending platform, was investigated by California's DFPI for compliance with AB 2013. The company's AI approved/denied loans in seconds but struggled to explain decisions to consumers. Upstart had to implement explainability features showing which factors (income, employment history, education) most influenced loan decisions. This increased transparency but slowed their approval process from instant to several hours.
SB 1047: Safe and Secure Innovation for Frontier AI Models (2024)
California's proposed legislation (debated in 2024) targeting large AI models:
- Covered Models: AI systems trained with $100M+ in computing resources
- Safety Testing: Mandatory third-party audits before deployment
- Kill Switch: Developers must implement shutdown capabilities
- Incident Reporting: Report AI systems causing "critical harm"
- Whistleblower Protections: Legal protections for employees reporting safety issues
Real Example - Anthropic's Response (2024): When SB 1047 was proposed, Anthropic (creator of Claude) publicly supported it while OpenAI opposed it. Anthropic already conducted third-party safety audits and had implemented "Constitutional AI" safeguards. OpenAI argued the law would stifle innovation and drive AI development out of California. The bill ultimately passed in modified form, requiring safety protocols but softening enforcement mechanisms. This showed the tension between safety regulation and economic competitiveness.
AB 2602: Deepfake Elections (2024)
California banned AI-generated deepfakes in political advertising within 120 days of an election:
- Criminal Penalties: Up to $10,000 fine for violators
- Civil Liability: Candidates can sue creators of malicious deepfakes
- Platform Responsibility: Social media platforms must remove reported deepfakes within 72 hours
- Disclosure Requirements: All AI-generated political content must be clearly labeled
Real Example - Ron DeSantis Deepfake (2023): During the 2024 presidential primary, a deepfake video showed Ron DeSantis appearing to endorse positions he didn't hold. The video went viral before being debunked. California's AB 2602 was directly inspired by this incident. Under the new law, the video's creator would face criminal charges, the candidate could sue for damages, and Twitter/X would be required to remove it within 72 hours of being notified.
Other State AI Laws
Colorado AI Act (2024)
First comprehensive state AI law covering algorithmic discrimination:
- Impact Assessments: Companies must conduct discrimination impact assessments for "high-risk" AI
- High-Risk Systems: AI used in employment, education, financial services, healthcare, housing, legal services
- Consumer Rights: Right to opt out of AI-based profiling
- Developer Duties: AI developers must provide documentation on intended use and limitations
New York City Local Law 144 (2023)
NYC's Automated Employment Decision Tools (AEDT) law:
- Annual Audits: AI hiring tools must undergo yearly bias audits
- Public Results: Audit results must be published publicly
- Candidate Notice: Job applicants must be told if AI screens their application
- Alternative Process: Candidates can request human review
Real Example - HireVue Audit (2023): After NYC Local Law 144 took effect, HireVue published bias audit results showing its AI was 7% more likely to advance white candidates than Black candidates for customer service roles. This transparency forced the company to retrain its models and adjust decision thresholds. Some NYC employers stopped using HireVue entirely due to reputational risk. This demonstrates how transparency requirements can drive market-based accountability even without direct penalties.
Global AI Regulation Context
EU AI Act (2024)
The world's most comprehensive AI regulation, establishing a risk-based framework:
- Prohibited AI: Social scoring, real-time biometric surveillance, manipulation
- High-Risk AI: Employment, education, law enforcement, critical infrastructure (requires conformity assessments)
- Transparency Requirements: Users must be informed when interacting with AI
- General Purpose AI: Models like GPT-4, Claude must document training data and conduct risk assessments
Real Example - Clearview AI Ban (2023): Clearview AI, which scraped billions of photos from social media to create a facial recognition database, was banned across the EU under GDPR. The EU AI Act would have prohibited this even more explicitly as "unacceptable risk" biometric surveillance. Clearview continued operating in the US but ceased all EU operations and paid €30.5 million in fines to multiple EU countries.
AI Ethics Frameworks
Core Ethical Principles
- Fairness: AI should not discriminate or create unjust outcomes
- Transparency: AI decision-making processes should be understandable
- Accountability: Clear responsibility for AI outcomes
- Privacy: Personal data protection in AI systems
- Safety: AI should not cause physical, psychological, or societal harm
- Human Agency: Humans should maintain meaningful control over AI systems
Algorithmic Bias: Detection and Mitigation
AI systems can perpetuate or amplify societal biases present in training data:
Real Example - Amazon Recruiting Tool (2018): Amazon built an AI system to screen resumes and rank candidates. The system was trained on 10 years of resumes submitted to Amazon - predominantly from men, especially in technical roles. The AI learned to penalize resumes containing the word "women's" (as in "women's chess club") and downgraded graduates of all-women's colleges. Amazon scrapped the system before deployment, but it demonstrated how historical bias in data creates biased AI.
Real Example - COMPAS Recidivism (2016): ProPublica investigation found COMPAS, an AI tool used by judges for bail and sentencing decisions, was biased against Black defendants. The system incorrectly labeled Black defendants as "high risk" at almost twice the rate as white defendants (45% vs. 23%). White defendants were mislabeled as "low risk" more often. This AI tool influenced real decisions about freedom and incarceration, demonstrating the high stakes of algorithmic bias.
AI Explainability and Transparency
Making AI decisions interpretable to humans:
- LIME (Local Interpretable Model-Agnostic Explanations): Explains individual predictions
- SHAP (SHapley Additive exPlanations): Assigns importance scores to features
- Attention Visualization: Shows what parts of input the model focused on
- Decision Trees: Inherently interpretable models for high-stakes decisions
Learning Objectives
By the end of this module, you will be able to:
- Explain Executive Order 14110 and its requirements for federal AI systems
- Describe the EU AI Act's risk-based approach to AI regulation
- Understand California's AI laws (CCPA, AB 2602) and their business implications
- Identify sources of algorithmic bias and methods for mitigation
- Evaluate the tradeoffs between AI explainability and model performance
- Apply ethical principles (transparency, accountability, fairness) to AI deployment decisions
Future of Work and AI
Job Displacement vs. Job Transformation
AI will both eliminate and create jobs, with net effects debated:
- Optimistic View: AI augments human capabilities, creating higher-value work
- Pessimistic View: AI automates faster than new jobs are created, increasing unemployment
- Likely Reality: Significant workforce disruption requiring massive reskilling efforts
Real Example - Klarna's AI Customer Service (2024): Klarna replaced 700 customer service representatives with AI chatbots handling 2.3 million conversations monthly. The company saved $40 million annually and improved customer satisfaction scores. This is often cited as "AI augmentation" success, but 700 people lost jobs. Klarna offered some employees retraining for tech roles, but most left the company. This exemplifies the tension between efficiency gains and workforce displacement.
Key Vocabulary
Algorithmic Bias: Systematic and repeatable errors in AI systems that create unfair outcomes, often reflecting historical inequalities in training data.
Red-Team Testing: Adversarial testing where researchers attempt to find vulnerabilities, biases, or dangerous capabilities in AI systems before public release.
Deepfake: AI-generated synthetic media (video, audio, images) that convincingly depicts people doing or saying things they never did.
Explainability (XAI): The ability to understand and interpret how an AI system makes decisions, crucial for accountability and trust.
High-Risk AI: AI systems used in critical domains (employment, healthcare, law enforcement) where errors can cause significant harm.
Data Minimization: Privacy principle requiring collection of only data necessary for a specific purpose - limits AI training data scope.
Constitutional AI: Technique training AI systems to follow ethical principles and values encoded in a "constitution" of rules.
Right to Opt Out: Consumer right to decline AI-based processing of their data or AI-driven decisions affecting them.
💡 Try It Yourself
Navigate AI Ethics
- → Ask: 'Explain the EU AI Act's risk categories with examples of each'
- → Try: 'Should my company disclose when AI is used in hiring decisions? What are the legal requirements?'
- → Test: 'Give me a framework for evaluating if an AI use case is ethical'
Use these prompts with ChatGPT, Claude, or Gemini to reinforce what you've learned.
Key Vocabulary: AI Ethics & Governance
Algorithmic Bias: Systematic and repeatable errors in AI systems that create unfair outcomes, often reflecting historical inequalities in training data.
Explainability (XAI): The ability to understand and interpret how an AI system reaches its decisions.
Executive Order 14110: Biden administration's 2023 directive establishing AI safety and governance standards across federal agencies.
CCPA (California Consumer Privacy Act): California law granting consumers rights over their personal data, including data used to train AI.
AB 2602: California law requiring disclosure when AI is used for automated employment decisions.
EU AI Act: Comprehensive European regulation categorizing AI systems by risk level and imposing requirements accordingly.
Transparency: Requirement that organizations disclose when AI is making decisions that affect individuals.
Accountability: Clear assignment of responsibility for AI system outcomes and failures.
Data Privacy: Protection of personal information collected, used, and stored by AI systems.
Disparate Impact: When AI systems produce different outcomes for different groups, even without intentional discrimination.
📢 Share This Free Course