Law works through people and institutions
AI systems do not sit outside ordinary accountability. Developers, deployers, vendors, officers, public agencies, and users make decisions about training, access, integration, and response. Law can assign duties, require evidence, create remedies, and authorize regulators to intervene. Its effectiveness depends on scope, enforcement capacity, technical competence, and whether obligations reach the actor that can actually prevent harm.
Regulation is not automatically faster or more reliable than public pressure, and voluntary practice is not automatically useless. Rulemaking and litigation can be slow; standards and procurement can move sooner but may lack democratic accountability. A durable strategy combines enforceable floors with flexible technical implementation and public scrutiny.
Match the rule to the lifecycle
A training rule might require reporting, security, or evaluation for exceptionally large or capable development runs. A model-weight rule might impose protection or controlled release where theft or unrestricted copying creates severe risk. A deployment rule can govern a hosted service’s users, safeguards, and incident monitoring. An inference or use rule can restrict an application in employment, health, credit, infrastructure, or weapons.
These are not interchangeable. Prohibiting one deployment does not erase weights. A training-compute threshold may miss a smaller model enhanced through tools or inference-time computation. An application rule may protect affected people without regulating general research. Good law identifies the actor, control surface, evidence trigger, and remedy.
The current landscape is no longer “voluntary only”
The European Union AI Act creates binding obligations. General-purpose AI provider duties began applying on August 2, 2025, including technical documentation, information for downstream providers, a copyright policy, and a training-content summary. Providers of models classified as presenting systemic risk face additional evaluation, mitigation, incident-reporting, and cybersecurity duties (European Commission GPAI overview). The AI Office’s relevant enforcement powers began applying on August 2, 2026 (EU enforcement framework).
In the United States, Executive Order 14110 was revoked in January 2025, so articles should not present its reporting program as current federal policy (2025 executive order). Federal policy remains distributed across existing agencies, sector law, procurement, export controls, and state rules rather than one comprehensive AI statute.
California enacted SB 53, requiring covered large frontier developers to publish frameworks, address catastrophic-risk thresholds and mitigations, report specified incidents, and protect covered whistleblowers (California Attorney General SB 53 guidance). New York’s RAISE Act and 2026 chapter amendments created additional frontier-model transparency and reporting requirements; the official bill record shows the chapter amendment was signed in March 2026 (New York Senate S8828). These laws are narrower than a general safety guarantee, but they disprove the claim that binding frontier-model governance does not exist.
Liability and remedy
Liability can make preventable harm costly, but slogans about assigning all cost to “the party best positioned” leave hard questions. A developer may control training and safeguards; a hospital or employer controls context and human review; a downstream vendor controls fine-tuning; a malicious user may violate every rule. Responsibility may be shared.
Useful regimes define standards of care, preserve evidence, allow harmed people to obtain explanations or challenge decisions, and avoid blanket immunity for foreseeable misuse. At the same time, unlimited liability can deter beneficial open research or concentrate the market among firms able to insure. Safe harbors can reward documented testing and prompt incident response without becoming immunity for checking a box.
Remedies should fit the harm: compensation, correction, deletion, injunction, product recall, license restriction, civil penalty, or criminal enforcement for intentional misconduct. High-stakes systems also need rapid administrative appeals because years of litigation cannot restore a missed benefit or urgent medical decision.
Procurement can create operational standards
Governments and large enterprises can require protections before purchase. U.S. Office of Management and Budget guidance for federal AI acquisition calls for cross-functional planning, performance monitoring, data and intellectual-property considerations, and risk management proportional to impact (OMB M-25-22).
A strong contract should specify:
- the approved model and whether the vendor may change it automatically;
- relevant pre-deployment tests and independent evaluation access;
- audit logs, documentation, data retention, and privacy;
- breach and AI-incident notification deadlines;
- tool permissions and human-approval points;
- rollback, suspension, credential revocation, and safe decommissioning;
- data export, interoperability, and transition assistance;
- subcontractor and cloud dependencies;
- remedies when performance or safeguards fail.
Procurement conditions can spread through the market because vendors reuse capabilities for multiple customers. They can also entrench incumbents if requirements are vague, proprietary, or disproportionately expensive. Open standards, shared testing resources, and scaled obligations help smaller suppliers compete.
Audits need access and consequences
An audit is only as good as its scope. A paper review cannot validate real-world behavior. Auditors need the relevant deployed configuration, documentation, logs, known incidents, and authority to test plausible failure conditions. Independence, competence, conflict disclosure, and protection for sensitive information all matter.
Audit findings should connect to remediation deadlines, deployment limits, follow-up tests, and public summaries appropriate to the risk. A confidential full report may be necessary for security or privacy, but secrecy should not allow a company to advertise “audited” without stating what was examined.
Regulators also need technical staff and secure access. Mandating reports that no agency can interpret creates compliance theater. Fees on regulated entities can support oversight if structured to preserve institutional independence.
Antitrust has a genuine safety tradeoff
Concentration in chips, cloud, data, distribution, and foundation models can create lock-in, shared failure points, and private political power. The FTC’s 2025 study of cloud-AI partnerships identified access to compute and talent, switching costs, exclusivity, and sensitive-information sharing as competition concerns (FTC staff report).
More competition can reduce dependency and allow users to leave unsafe providers. It can also intensify release races, multiply poorly secured weight repositories, and make coordinated safeguards harder. A single large provider may invest more in security while creating a catastrophic single point of failure. Antitrust policy should therefore protect contestability, interoperability, labor mobility, and access to essential inputs without assuming that the largest or smallest market structure is automatically safest.
Safety coordination among competitors can be beneficial—shared incident formats, benchmark methods, or security intelligence—but can also conceal exclusion or price coordination. DOJ and FTC sought updated public guidance on competitor collaboration in 2026, explicitly including algorithmic pricing and information sharing (DOJ/FTC inquiry). Clear rules can allow genuine safety cooperation while preserving competition.
Incident response and emergency authority
Organizations should have legal duties to preserve logs, notify appropriate authorities and affected parties, contain harm, and cooperate across dependencies. Regulators need graduated powers: request information, require mitigation, restrict a function, or order suspension when evidence shows urgent danger.
Emergency measures need due process. They should identify the decision-maker, threshold, scope, duration, review, and restoration conditions. A broad indefinite “AI shutdown” is neither technically coherent nor compatible with accountable government.
The practical conclusion
Legal leverage is strongest when it is specific. It turns “be responsible” into duties attached to training, weights, deployment, inference, procurement, and high-impact use. It gives affected people remedies and gives operators clear authority to audit, revoke privileges, roll back systems, and respond to incidents.
No jurisdiction has solved AI governance. But by September 2026, binding EU and state-level frontier obligations, federal procurement rules, competition enforcement, and sector law provide real tools. The task is to improve their evidence, interoperability, enforcement, and democratic legitimacy—not to describe the field as a choice between voluntary promises and dramatic confrontation.