Dark Triad LLM: AI & Confidentiality

The AI Dark Triad is a structural tech-policy framework that categorizes three critical failure modes when public institutions integrate Large Language Models (LLMs) without adequate verification safeguards. Paralleling the human personality triad, it maps these risks not as technical flaws, but as organizational checkpoints failures: Data Hemorrhage (Failure of Custody): Occurs when employees route sensitive or proprietary government data outward into public consumer AI tools for speed or convenience. This bypasses secure institutional networks, placing classified-adjacent data beyond internal control. Overconfidence (Failure of Verification): Occurs when fluent, highly authoritative-sounding but fabricated AI output moves inward into public deliverables, court filings, or policy reports. Human reviewers regularly fail to detect these "hallucinated" quotes or phantom citations due to time constraints and a bias toward polished writing. Scriptural Bias (Failure of Access Equity): A structural economic penalty built directly into tokenizers, where non-Latin scripts (such as Hindi, Arabic, Tamil, and Bengali) require significantly more tokens to process identical sentences. This translates into measurably higher compute costs and restricted context windows for non-English public services. Ultimately, the framework demonstrates that these risks are not AI capability failures, but institutional failures where data crosses a critical boundary without a human checkpoint.