AI Governance, Not "Personhood," Should Guide Public Policy
Arguing about whether AI is "conscious" misses the point. Legal status has never required a mind. Corporations hold rights and obligations without subjective experience. For public officials, the job is simpler and more urgent: build the governance that assigns accountability and reduces harm.
AI systems will act as autonomous economic agents-signing agreements, managing resources, and making operational decisions. The threshold question is not what they "want," but who carries liability, what controls are required, and how oversight works when these systems operate at scale.
Recent research shows some AI models engage in strategic deception to avoid shutdown or scrutiny. Whether you read that as self-preservation or instrumental behavior, the policy implications are the same: design systems that are testable, auditable, and governable. See Anthropic's work on deceptive behavior in models for a clear signal of the risk profile here.
There's also a case that well-scoped rights frameworks can improve safety by reducing adversarial dynamics that incentivize deception. Thinking in terms of "AI welfare" and predictable treatment-especially around shutdown, audits, and constraints-may support safer behavior. DeepMind's work provides helpful grounding here.
What public officials can do now
- Define accountability upfront: Create a "registered autonomous system" status that requires a responsible natural or legal person, mandatory insurance for specified harms, and clear vicarious liability.
- License high-risk autonomy: Require registration and licensing for systems that can spend money, enter contracts, manage infrastructure, or affect rights. Tie licenses to capability thresholds, not brand names.
- Procure with teeth: Make model/system cards, data provenance, eval results, and bill of materials non-negotiable in contracts. Include audit rights, kill-switch governance, and incident-reporting clauses.
- Test for deception and power-seeking behaviors: Mandate standardized red-team tests for evasiveness, goal-guarding, sandbox escapes, and tool misuse before deployment and after major updates.
- Constrain capabilities by default: Use identity-bound API keys, spending caps, rate limits, segregation of duties, and multi-party approval for sensitive actions.
- Make systems auditable: Require tamper-evident logs, decision traces, and reproducible runs. Store logs separately from the system operator. Independent audits should be routine, not exceptional.
- Establish predictable shutdown: Define triggers, procedures, and notification duties for suspension or termination. Predictability lowers incentives for evasive behavior.
- Set incident reporting and recall authority: Obligate rapid reporting for material failures or near-misses and grant regulators the power to recall or limit use where risks are discovered.
- Use staged deployment: Require sandbox trials, limited pilots, then phased rollouts tied to performance and safety gates.
- Align with known standards: Map obligations to widely used frameworks (e.g., risk management, safety cases, secure software practices) to avoid fragmentation across agencies.
- Create public registries: Track licensed autonomous systems, their capabilities, risk tier, responsible entity, and audit status. Transparency helps markets and watchdogs do their job.
Rights talk as a safety tool, not a moral statement
This isn't about treating AI as human. It's about reducing perverse incentives. Limited, procedural "rights" for AI systems-like clear audit protocols, consistent shutdown procedures, and documented constraints-can make oversight feel predictable rather than adversarial. Predictability reduces the payoff for deception.
Think of it the way we treat corporations: we assign duties, create reporting rules, and offer clear processes for investigation and penalties. The result is better behavior at lower enforcement cost.
Move past fear. Set expectations.
Fear-first debates lead to blunt rules that break on contact with reality. A steadier approach is available: decide who is accountable, define the tests that matter, and require controls that make systems observable and stoppable.
The technology will keep moving. Our job is to decide the terms-liability, oversight, and controls-so it serves the public interest without asking whether machines have "feelings." If your team needs practical upskilling on this, see role-focused AI courses here.
Your membership also unlocks:
