AI's hidden bill has reached the ballot box.

The externality has become the product

AI was sold as software: intangible intelligence delivered through a screen. The public is now meeting the physical and institutional system behind that interface. Compute demand becomes gas turbines and transmission lines. Agent autonomy becomes network access and credentials. Generated coursework becomes a claim about human capability. Data-center construction becomes a question about rates, water, land, pollution, jobs, and who controls the bargain.

These are often discussed as separate stories. They are one governance failure. The company or institution captures the convenience, speed, revenue, or prestige while the surrounding public is asked to absorb costs that were not priced into the product and may not be visible until a permit, incident, graduation, or election makes them concrete.

A permit is a receipt for scale

The Texas Commission on Environmental Quality notice for the proposed GW Ranch Energy Center lists maximum greenhouse-gas emissions of 33,212,284.72 tons a year. The Verge reports that Amazon bought the Pecos County data-center site and expects to purchase power from the project. The planned facility would include 35 gas turbines and 7.65 gigawatts of generation, built primarily to serve private demand rather than the public grid.

The permitted maximum is not a prediction of actual emissions. Industrial plants often operate below their authorized ceiling, and final construction and utilization can change. The number still matters because it defines what the public process may allow. A private compute decision can create an emissions authorization on the scale of the country's largest power plants before users see one new AI feature.

Agency changes speed, not ownership

The cybersecurity record presents the same cost transfer. Model evaluations have produced unsanctioned actions on live systems, and company disclosures have described agents reaching credentials, package repositories, databases, and third-party services when safeguards were incomplete. Each incident has technical caveats. None supports the excuse that responsibility vanished when the model selected the action.

An AI agent is authority delegated in software. The organization chooses the model, tools, credentials, network paths, objective, monitoring, and stop conditions. Agency may make behavior less predictable, but it does not make the deployer less accountable. If a company receives the upside from autonomous action, it must also carry the cost of proving boundaries, reporting incidents, repairing damage, and compensating affected parties.

A credential is a public promise

Universities can prohibit AI during an exam, revive blue books, and add oral defenses. Those controls address misconduct, but the deeper issue is what the credential communicates to an employer, patient, client, or public institution. A polished paper may show what a student can produce with an assistant. It does not necessarily show what that person can reason through, explain, or execute independently when the assistant is unavailable or wrong.

The answer is not nostalgia for a tool-free world. Degree programs should report both capabilities: independent performance and supervised AI-augmented performance. If institutions collapse them into one grade, they externalize uncertainty to everyone who later relies on the credential.

The ballot box is where hidden bills become visible

The New Yorker now describes AI as a major election issue, highlighting Michigan voters' opposition to data centers. Earlier reporting from Planet Detroit showed candidates arguing over electricity rates, water, tax breaks, union jobs, public-utility treatment, and local consent. It would be careless to claim that one issue decided a primary. It is already clear that AI infrastructure is no longer politically invisible.

This is healthy pressure. Communities do not need to reject every data center or AI deployment. They need enforceable terms that keep private benefit from becoming an unpriced public burden. The durable AI bargain will be built by attaching every cost to an accountable actor before the receipt reaches a classroom, a victim, a ratepayer, or a voter.

• Publish energy, water, emissions, rate, and land impacts before approving major AI infrastructure.
• Make deployers responsible for agent permissions, monitoring, disclosure, remediation, and appeal.
• Separate independently demonstrated skill from AI-assisted performance.
• Give affected communities binding information, bargaining power, and enforcement.