AI Usage Policy

PsyData Labs L.L.C. (PDL) - AI Usage Policy

Effective Date: July 30, 2026

1. Purpose and Scope

This AI Usage Policy establishes the comprehensive governance framework for the development, deployment, and use of Artificial Intelligence (AI) and Machine Learning (ML) systems at PsyData Labs L.L.C. (PDL). It ensures that all AI systems, particularly those processing psychological, behavioral, and telemetry data, are developed and utilized in a secure, ethical, and compliant manner. This policy applies to all PDL employees, contractors, vendors, and third-party partners who interact with PDL's AI systems or data.

2. AI Risk Tiering

All AI systems and models at PDL must be classified into a risk tier before deployment, aligning with our AI Risk Tiering Policy (AIG-002) and international frameworks (e.g., EU AI Act, NIST AI RMF). The tier determines the level of scrutiny, governance, and oversight required.

• Tier 1 (Low Risk): Internal tools, non-sensitive administrative tasks, and low-impact automation. Requires standard security review.
• Tier 2 (Medium Risk): Systems supporting internal decision-making, aggregate analytics, and non-sensitive behavioral feature extraction. Requires model cards, prompt logging, and continuous evaluation metrics.
• Tier 3 (High Risk): Systems generating psychological profiles, automated inferences on Restricted (S3+) data subjects, and models affecting employment, credit, or insurance. Requires strict Human-in-the-Loop review, bias testing, active drift monitoring, and incident learning protocols.
• Tier 4 (Unacceptable Risk): Systems engaging in subliminal manipulation, unconsented biometric categorization, or deepfake generation for deception. Strictly prohibited at PDL.

3. Human-in-the-Loop (HITL) Requirements

In accordance with PDL's Human-in-the-Loop Policy (AIG-004), GDPR Art. 22, and EU AI Act Art. 14, automated systems must not independently execute high-impact decisions.

• Mandatory Oversight: Any high-risk model output affecting Restricted (S3+) data subjects must undergo a mandatory human review gate prior to final execution, deployment, or external communication.
• Prohibition on Sole Automation: Fully automated decisions concerning employment, credit, insurance, or significant psychological assessments based solely on AI inferences are strictly prohibited without documented human intervention.
• Subject Rights: Data subjects retain the right to request human intervention for decisions based on solely automated processing. PDL must respond to and process such requests within 30 days.

4. Transparency and Hallucination Disclaimers

Generative AI systems and Large Language Models (LLMs) used in external-facing applications, analytics, or research must transparently communicate their nature to end-users.

• AI Disclosure: All AI-generated content, psychological insights, or behavioral summaries must clearly state that they are generated or augmented by an AI system.
• Hallucination Disclaimer: Applications outputting generative text or predictive inferences must display a prominent disclaimer indicating that AI models may occasionally produce inaccurate, misleading, or "hallucinated" information. End-users must be advised to independently verify critical information.
• Explainability: AI models should provide explainable logic, evaluation gates, or confidence scores where technically feasible, enabling end-users to understand the basis of an inference.

5. Data Privacy, Security, and Governance

The processing of data within AI systems must strictly adhere to PDL’s data governance and security standards.

• Model Documentation: A current model card must be maintained for each production model, detailing its purpose, risk tier, limitations, and evaluation metrics.
• Logging and Retention: Prompts, inferences, and high-risk evaluation gate outputs must be securely logged with retention periods aligned to EU AI Act accountability expectations.
• Synthetic Data Utilization: Where possible, synthetic data should be utilized for model training and incident learning to protect the privacy and confidentiality of actual data subjects.

6. Model Monitoring, Drift, and Incident Management

Production AI systems are subject to continuous lifecycle management and proactive monitoring.

• Drift and Bias Monitoring: Models must be continuously monitored for data drift, concept drift, and emergent bias. Thresholds for automated rollback or alerts must be defined for all High-Risk (Tier 3) systems.
• Incident Response: Any unexpected behavior, safety violation, or data exposure caused by an AI system must trigger an immediate AI incident postmortem and corrective actions, overseen by the AI Governance Lead.

7. Enforcement and Compliance

The AI Governance Lead, Security Lead, and the Compliance Team are responsible for the overarching enforcement of this policy. Compliance teams will sample a minimum of 5% of system changes monthly for adherence to the AI Risk Tiering and Human-in-the-Loop policies. Violations of this policy may result in immediate access revocation, disciplinary action, and project suspension.