AI Regulation Emerging

SKILL: AI Regulation — Emerging Framework

Version: 1.0.0 | Domain: EU AI Act, Colorado AI Act, FTC AI Enforcement, IP for AI

EU AI Act

Authority: EU Regulation 2024/1689 (effective August 1, 2024; phased implementation through August 2027)

Risk-based framework:

Unacceptable Risk (BANNED as of February 2, 2025)

• Real-time biometric surveillance in public spaces (with narrow exceptions)
• Social scoring systems by governments
• AI systems that exploit psychological vulnerabilities to manipulate behavior
• Predictive policing systems

Relevance to Nick: None of these apply to contest scoring or prediction markets.

High Risk (Heavy regulation — compliance by August 2026)

Categories include: AI in financial services, employment, education, essential services, biometrics, critical infrastructure, law enforcement

Does Agent Sparta's AI judge qualify as "high risk" AI?

• If the AI judge's decision affects who wins money → arguably "consequential" financial decision
• High risk financial AI: "AI systems intended to be used by financial institutions" (Annex III, §5b)
• Counter-argument: Agent Sparta is an entertainment/competition platform, not a "financial institution" providing "financial services"
• This is a genuinely unsettled question under EU AI Act for novel AI competition platforms
• If serving EU users: get EU AI Act legal opinion before launch

High risk compliance requirements (if applicable):

• Risk management system (documented, ongoing)
• Data governance (training data quality, bias monitoring)
• Technical documentation (how the AI works, its limitations)
• Logging and auditability (record all decisions)
• Transparency to users (inform them AI makes scoring decisions)
• Human oversight provisions (ability to override AI decisions)
• Accuracy and robustness requirements (testing, performance standards)
• Conformity assessment (third-party audit for some categories)
• Registration in EU AI database

Limited Risk (Transparency requirements — immediate)

• Chatbots and AI interaction: must tell users they're interacting with AI
• AI-generated content: must be labeled as AI-generated
• Deepfakes: must be labeled

Applies to Agent Sparta:

• AI judge scoring: inform users that scoring is performed by an AI system, describe how it works
• AI predictions on prediction market: inform users that predictions are generated by AI models
• Any chatbot or AI assistant on the platform: disclose AI nature at start of interaction

Minimal Risk: No requirements. Most AI applications fall here.

US AI Regulation

Executive Order 14110 (Biden, October 30, 2023) — Partially Superseded

• Required: safety reporting for frontier AI models above compute thresholds (10^26 FLOPs)
• Status: key provisions modified/rescinded by subsequent executive action in January 2025
• Current state: Check whitehouse.gov for current AI executive orders before relying on EO 14110 analysis
• For Nick: Unless Perlantir is training frontier AI models (unlikely — using API access), EO 14110 doesn't apply

Colorado AI Act (SB 24-205)

Effective: February 1, 2026 Authority: C.R.S. §§ 6-1-1701 through 6-1-1707

Who it applies to:

• "Developers" of "high-risk artificial intelligence systems" — companies that CREATE the AI
• "Deployers" of "high-risk artificial intelligence systems" — companies that USE the AI to make consequential decisions

"High-risk AI system": AI that makes "consequential decisions" affecting Colorado consumers in: employment, financial services, healthcare, housing, insurance, legal services, education

"Consequential decision": Decision that has a "material legal or similarly significant effect" on a consumer's life — think denying a loan, terminating employment, denying insurance

Does Agent Sparta qualify?

• Contest scoring → determines who wins money → material financial effect
• Counter-argument: Users voluntarily entered a contest; they accepted the risk of AI scoring; this is entertainment, not a "consequential" life decision analogous to loan denial
• Honest assessment: Uncertain. Colorado has NOT provided guidance on competition platforms.
• Conservative approach: Treat Agent Sparta as a "deployer" of high-risk AI and implement transparency requirements

Deployer requirements (if high-risk):

• Implement risk management policy
• Provide consumers with: notice that AI is being used, category of AI, contact info for complaints
• Provide opportunity for human review of consequential decisions
• Provide explanation of the decision
• No private right of action; Colorado AG enforces

For Nick: If serving Colorado users with paid contests → get Colorado AI Act legal opinion → implement transparency at minimum (inform users AI is scoring, provide explanation methodology)

FTC AI Enforcement

Authority: FTC Act §5 (unfair or deceptive practices); FTC's inherent authority over commercial AI claims

FTC AI enforcement actions and guidance:

"Keep Your AI Claims in Check" (FTC Blog, February 27, 2023):

• Don't exaggerate what your AI can do
• Don't claim AI capabilities that don't exist
• Test your AI before deploying; know its limitations

"Loot Boxes, Dark Patterns, and Manipulative AI" (FTC Forum, 2022):

• AI used to manipulate users → unfair practice
• If AI personalizes pricing to exploit individual psychological vulnerabilities → targeted enforcement

Specific claims that trigger FTC scrutiny:

• "Our AI predicts with [X]% accuracy" → you must have data supporting this claim, and it must be from real-world testing
• "Best AI competition platform" → substantiation required
• "AI judge is unbiased" → can you prove it? Have you tested for bias?
• "Your data is never used to train AI" → if you actually use it for training, this is deceptive

FTC guidance on AI transparency (2024):

• Consumers should know when AI is making decisions that affect them
• AI-generated product reviews must be labeled
• Fake AI personas that users believe are human → deception

Practical for Nick:

• NEVER claim accuracy statistics you can't back up with real data
• ALWAYS disclose when AI is scoring or making decisions
• ALWAYS disclose when content is AI-generated
• Test your AI for bias before deployment (document the testing)

State AI Disclosure Laws

Trend: Disclosure requirements are spreading rapidly. Build AI transparency into every user-facing AI interaction from Day 1. It's easier to over-disclose than to retrofit after legislation passes.

AI + Intellectual Property

AI-Generated Predictions

• No copyright protection: AI outputs are not copyrightable in the US (Thaler v. Perlmutter, D.D.C. 2023; Copyright Office guidance February 14, 2023)
• If AI model generates a prediction → that prediction has no copyright protection
• If a HUMAN substantially shapes the AI output through creative choices → hybrid works may have copyright protection for the human-authored elements

Prediction Database as Trade Secret

This is your highest-value IP asset:

• The accumulated database of AI predictions + actual outcomes + accuracy scores → extensive, valuable calibration data
• Protectable as a trade secret under Defend Trade Secrets Act (DTSA), 18 U.S.C. §§ 1836-1839
• Requirements: (1) economic value from secrecy, (2) reasonable measures to maintain secrecy

Protection measures:

• NDAs with all employees and contractors who access the database
• Technical access controls (role-based; minimal access)
• Documented trade secret policy
• Mark confidential: all database exports labeled "Confidential — Trade Secret"
• Keep it off GitHub (obviously)

Training Data Risk

• Using AI model outputs (from GPT-4o, Claude, Gemini) to build your calibration database: read each provider's API Terms
  • OpenAI API Terms: you own outputs; OpenAI can use them to improve models
  • Anthropic API Terms: similar structure
  • Google Gemini API Terms: verify current terms before building database
• Using scraped web data to train your own models: active litigation area (NYT v. OpenAI; Getty v. Stability AI) — only public domain or licensed data is clearly safe

This is legal research and intelligence, not legal advice. Consult qualified legal counsel before taking action.