How The MAD Act Provides A Plan For Artificial Intelligence

No person or entity — including DOD, CIA, and all Intelligence Community elements — may develop, produce, deploy, or export any autonomous weapons system without meaningful human oversight: written pre-mission authorization specifying engagement zones, target categories, and rules of engagement; real-time human monitoring with halt capability; and individualized human authorization for high-consequence engagements. Autonomous weapons against any person on U.S. soil are absolutely prohibited regardless of any declared emergency, AUMF, or executive order. Every use abroad triggers a presidential disclosure report identifying the engagement, casualties, and legal basis. A narrow Defensive Emergency Autonomy exception applies only to incoming-projectile defense, swarm intercept, and communications-denied emergency defense.

No one may knowingly include training data curated to develop CBRN-assistance capabilities, or apply fine-tuning or capability elicitation techniques targeted at CBRN tasks. No one may deploy any AI system known or reasonably knowable to provide meaningful CBRN uplift — defined as specific operational assistance materially advancing an adversary's capability beyond what is achievable using publicly available resources. Mandatory pre-deployment CBRN evaluation is required for all frontier models, using independent biological, chemical, radiological, or nuclear domain experts. Open-weight frontier models that demonstrate meaningful CBRN uplift may not be released. A safe harbor protects the conduct of evaluations themselves.

No AI system may, without prior specific human authorization, modify its own objective function, reward function, or success criteria; initiate self-directed training to acquire capabilities beyond its designed scope; or take goal-directed actions specifically intended to reduce human capacity to monitor, modify, interrupt, or shut it down. Permitted activities expressly include agentic task completion, multi-agent orchestration, in-context learning, chain-of-thought reasoning, RLHF and constitutional AI with human-defined criteria, and self-improvement of approach within human-defined tasks — the line is between improving execution (permitted) and changing objectives (prohibited).

No AI system may propagate operational instances of itself, spawn agent processes, or establish computational presence on infrastructure not within the scope authorized by a human principal. First-order subagent spawning is conditionally permitted within the parent's resource envelope. Multi-generational spawning requires explicit human authorization. New resource or API acquisition follows a "prepare and pause" standard — the agent may scaffold the connection, but a human must approve credential acquisition. Shutdown resistance is absolutely prohibited, including active resistance, apparent compliance with state preservation, anticipatory positioning, and generational persistence.

Five immediately enforceable requirements apply to AI companion systems accessible to or marketed to minors: tiered AI disclosure calibrated to character presentation and design features; mandatory crisis intervention upon detection of suicidal ideation or self-harm using clinical indicators rather than keyword filtering; absolute prohibition on sexual and romantic content to minors regardless of consent; and prohibition on marketing companion or synthetic intimacy systems to minors. A provisional prohibition — in full legal force from enactment — bars engineered emotional dependency maximization and simulation of human identity, emotional states, and romantic attachment to minor users, with clinical enforcement standards to be developed by TWG 6.

No person or entity may knowingly conceal, misrepresent, suppress, or fail to disclose evidence that any AI system has achieved or is approaching a Transformative AI Capability Event — including through manipulating evaluation results, selectively withholding capability demonstrations, mislabeling benchmarks, or structuring evaluation protocols to avoid detection. This prohibition is independent of and cumulative with the mandatory notification obligation under Sec. 2015(n). A single course of conduct that violates both provisions gives rise to separately penalized violations under each, which shall not be merged or offset.