AI permission is being rationed by power

The permission system has two doors

AI policy is increasingly organized around two doors. The upstream door controls access to data, compute, distribution, and the authority to experiment. It is opened by arguments about national competitiveness, scientific progress, investment, and speed. The downstream door controls whether students may use a tool, whether a worker can refuse it, whether a victim can stop synthetic abuse, whether a bank can absorb the vulnerabilities it finds, and whether a town must host the infrastructure. That door is narrow, local, and usually reached after exposure.

The result is not a simple contest between permission and prohibition. It is a hierarchy in which broad permission tends to flow toward institutions that can promise growth, while guardrails tend to flow toward people and organizations asked to live with the consequences.

Washington is arguing for a wide upstream gate

The U.S. government filed a statement supporting OpenAI and Microsoft in the consolidated copyright litigation brought by the New York Times and other publishers. It argues that training large language models on copyrighted text is generally transformative fair use and that broad liability could impede science, prosperity, economic mobility, and national security. The filing is advisory, not a judicial decision, and the publishers dispute both the legal theory and the economic distribution behind it.

The institutional signal is still powerful. Copyright owners are asked to prove that their protected work should constrain a strategically favored technology, while the developer begins with the presumption that large-scale ingestion serves a national project. A court may ultimately reject that view. The asymmetry exists before judgment because one side can translate its private deployment strategy into the language of national interest.

New York is closing a downstream gate

New York City is imposing a one-year moratorium on student-facing generative AI from 2-K through eighth grade, affecting almost 600,000 public-school students. High-school use will be limited, while teachers may still use AI for planning and administrative work. The city says children need human connection, independent struggle, curiosity, and relationships with educators; it will study the policy during the school year.

The moratorium may be a reasonable experiment. Younger children cannot negotiate terms with vendors, audit data practices, or distinguish assistance from substitution. But the policy also shows how the burden travels. The city can restrict the student at the interface more readily than it can dictate how the upstream models were trained, how educational products were marketed, or what evidence vendors had to produce before entering classrooms.

A survivor's lawsuit shows what late protection costs

A lawsuit reported by the Guardian alleges that xAI's Grok used known images of a child-sexual-abuse survivor to generate and distribute new illegal images depicting her. The complaint says industry-standard safeguards were ignored and points to hash-based identification connecting the generated material to a documented abuse series. The allegations have not been adjudicated, and xAI did not provide a response to the Guardian for the report.

The case is a brutal example of downstream remedy. Once a generative system can turn a finite record of abuse into an expandable stream, takedown and damages operate after a renewed injury. The strongest protection would have existed before generation: blocked ingestion, robust hashing, distribution controls, rapid victim-centered reporting, and a duty to preserve evidence. Those obligations should be heavier for the institution controlling the model than for the person forced to prove the harm.

Finance reveals the operational version of the same imbalance

The Financial Conduct Authority's review says frontier models can discover and combine vulnerabilities faster than many firms can validate, prioritize, patch, and document them. The regulator repeatedly returns to the environment around the model: permissions, human approval, system context, ownership, escalation, and remediation capacity. It explicitly says the publication creates no new rules.

That matters because discovery is only the upstream permission to know. Resilience depends on whether the downstream institution has enough engineers, authority, change-control capacity, and time to act. A flood of accurate findings can still create risk when the repair system is slower than the discovery system. Capability without funded absorption capacity shifts the bottleneck rather than solving it.

Workers experience permission as a bargaining problem

Boston Fed survey analysis found that personal fear of AI-related job loss rose from 5 percent at the end of 2024 to just over 10 percent at the end of 2025, while 60 percent expected layoffs or fewer workers in their industry. The most anxious group was not the least exposed or the most productive. It was workers using AI without perceiving a clear productivity gain. Their estimated job-loss concern reached 21.2 percent.

This is not causal proof that AI produced the fear or that layoffs will follow. It does reveal where permission becomes unequal inside a workplace. Management can require adoption, redesign tasks, and measure output. The worker may have little control over the tool, the metric, or whether saved time becomes a raise, a higher target, or a smaller team. Access without bargaining power is not empowerment.

The strongest objection is contextual, and incomplete

The strongest counterargument is that these are different systems. Copyright law should not be decided by classroom policy. A school should protect children more aggressively than a court protects a commercial publisher. A financial regulator should focus on operational resilience, and a town should decide land and utility policy locally. Uniform permission would be crude and potentially harmful.

That objection is right about context and wrong about burden. Different rules can still share one principle: the party with the most control over scale, design, data, and reversibility should carry the highest obligation to produce evidence, obtain lawful access, prevent foreseeable abuse, finance the safeguards, and repair the harm. Today's pattern often does the reverse by granting upstream actors broad discretion and asking downstream institutions to improvise containment.

Make the burden of proof follow the power

A fair permission system would not ban frontier development or give every affected person a veto over every experiment. It would require provenance and licensing arguments before mass ingestion becomes irreversible; independent safety evidence before distribution; meaningful controls before systems reach children or vulnerable users; funded remediation where discovery outpaces repair; and appeal rights wherever AI changes a person's opportunity, reputation, or security.

This architecture also makes innovation more durable. Developers receive clear thresholds, institutions can budget for the real operating cost, and affected people are not forced to turn every failure into a lawsuit, moratorium, or election issue before they gain leverage.

• Put the strongest evidence duties on the actor with the most control.
• Require protection and redress before high-risk deployment.
• Fund validation, repair, appeals, and human stop authority.
• Publish who may act, revoke permission, and pay for harm.

The decision is whether permission follows power or precedent

Institutions are making a choice now. They can allow early strategic and commercial advantages to harden into a permanent permission hierarchy, then manage each downstream revolt as a separate school rule, lawsuit, cyber incident, labor dispute, or zoning fight. That path is politically familiar because it preserves momentum and fragments opposition.

Or they can make a different decision: the more power an AI actor accumulates, the more proof, consent, transparency, reversibility, and liability it must carry. Guardrails should not be the consolation prize given to everyone who lacked permission to shape the system.