How this editorial can be challenged
What becomes governable when policy moves below the AI interface and targets the chokepoints that actually distribute power?
AI power is exercised through enabling systems: schools define acceptable assistance, model operators set abstention thresholds, utilities allocate scarce generation and cooling capacity, and data custodians determine whether neural signals can be recombined into new inferences. Rules at those chokepoints shape behavior before a user ever sees an output.
Infrastructure-level governance can harden the position of incumbents because large companies are better able to absorb technical standards, negotiate grid access, build compliance systems, and influence the definition of acceptable evidence. Product-level rules may be more flexible and easier for smaller innovators to understand.
That risk is real, but it argues for open, contestable infrastructure rules rather than retreating to interface disclosures. Shared testing methods, public-interest data access, transparent capacity contracts, portable audit records, and proportional duties can lower compliance barriers while preventing private chokepoints from becoming unreviewable authority.
The PISA relationships are observational rather than causal; the language-model study isolates factual multiple-choice abstention rather than open-ended agency; Malaysia's figures capture a hot-weather demand spike and official capacity planning; and the European neuro-AI statement is expert advice, not enacted law or measured outcomes.
This argument would weaken if product-level notices and user controls repeatedly prevented high-consequence harms without changes to underlying institutions, or if infrastructure standards consistently reduced access and competition while producing no measurable gains in safety, accountability, resilience, or public bargaining power.
The screen is where the public looks
AI debate is organized around visible encounters. A student opens a chatbot. A model answers or refuses. A user clicks through a privacy notice. A company announces a data center. These are the moments people can photograph, regulate, praise, or blame.
Yet each encounter is the last step in a longer chain of institutional choices. The important question is not only what the interface says. It is who designed the surrounding system, who controls the scarce resource, which threshold changes behavior, and whether an affected person can contest the decision.
Earlier networks moved power below the consumer product
Electricity, railways, telecommunications, and modern finance all became socially decisive when the network beneath the individual transaction shaped access, price, reliability, and exclusion. The light switch mattered, but generation and rate design mattered more. The train ticket mattered, but track ownership and scheduling determined the market.
AI is crossing the same boundary. Treating each chatbot or model as a self-contained product misses the network of education rules, compute allocation, energy supply, evaluation systems, and data rights that makes the product consequential. Governing only the screen is the digital equivalent of regulating the shape of a plug while ignoring the grid.
The classroom is an assessment infrastructure
PISA 2025 reports that 45.5% of students across OECD countries use AI at least weekly to help them learn. Weekly use was associated with science performance similar to non-use after adjusting for socio-economic background, while very frequent and occasional users tended to score lower. The report warns that those relationships do not establish causation.
The more important result is institutional. Students who had opportunities to assess AI-generated information showed a more promising learning pattern, while disadvantaged students were less likely to report receiving that practice. The advantage is not the chatbot. It is a school system that teaches when to question it and evaluates reasoning that cannot be outsourced invisibly.
Confidence is a control surface, not a feeling
A peer-reviewed study found causal evidence that internal confidence-related signals help determine whether several language models answer or abstain. Steering confidence changed refusal behavior, and instructed thresholds changed the decision policy. But verbal confidence was less predictive of correctness even while it independently influenced abstention.
That distinction matters for autonomous systems. A confidence display is not a truth meter. The operational question is who sets the threshold, how costs for errors and unnecessary refusals are encoded, and whether an independent evaluator can test the resulting behavior in the setting where the model will act.
Heat exposes the physical stack
Malaysia's energy regulator said data centers reached 9.3% of national electricity consumption during the second week of a hot August, above a 7% average for 2026. Officials linked the increase to cooling demand as low hydro levels tightened supply and described a 9-gigawatt gas-fired capacity gap to fill by 2032.
The figure turns the cloud back into infrastructure. The decisive governance choices are interconnection queues, weather-sensitive demand forecasts, cost allocation, firm capacity, water and cooling plans, and emergency curtailment. A sustainability badge on an AI service cannot answer who pays when heat and computation peak together.
Neurodata turns cognition into a shared system
European ethics advisers argue that neurotechnology and AI are converging into infrastructures that collect, combine, process, and reuse neurodata across contexts. Their recommendations reach beyond device safety and raw-data privacy to cover derived inferences, consequential uses, public-interest capacity, and a review of the wider legal framework.
That infrastructure framing is overdue. A person may consent to a device for rehabilitation without understanding what future models could infer from combined neural signals. Governance must travel with the data and the inference, not end at the device or the first purpose named on a form.
The incumbent trap is real
Rules for infrastructure can become tollbooths. A dominant laboratory can afford a complex evaluation regime. A hyperscaler can secure long-term power. A large platform can define the data standard. Smaller competitors, public researchers, schools, and patients may be left with obligations they cannot meet and systems they cannot inspect.
The remedy is not to leave the infrastructure ungoverned. It is to make the rule itself contestable: publish test methods, require portable records, open public-interest access, disclose capacity and cost assumptions, and scale duties to authority and consequence. A standard that only an incumbent can satisfy is not public protection. It is market design disguised as safety.
- Publish the threshold and evidence required before an AI receives more authority.
- Make resource contracts disclose who pays for expansion, resilience, and failure.
- Protect derived neural inferences as seriously as the raw signals that produced them.
- Fund public evaluators and schools so governance capacity is not purchased only by incumbents.
Look beneath one ordinary AI interaction
Picture a student asking a model for help on a hot afternoon. The answer appears in seconds. Beneath it sits a school policy that may reward the answer or the reasoning, a confidence threshold that decides whether the model speaks, a data center drawing more power for cooling, and a legal architecture deciding which traces can be kept and recombined. The screen makes the exchange feel individual. The infrastructure makes its consequences collective.
That is where AI governance must learn to look. Interfaces will keep changing, and product names will keep arriving faster than legislation. Chokepoints persist. The society that can see, measure, and contest them will have a chance to shape AI. The society that regulates only what appears on the screen will discover that the real decisions were made somewhere underneath.
Read the reporting
Opinion is ours. The factual record is linked below.
OECD — PISA 2025 Results, Volume I OECD — Student school life, digitalisation and AI Nature Machine Intelligence — Causal evidence that language models use confidence to drive behaviour Reuters — Malaysian data centres consume more power as temperatures rise European Commission — Experts call for a new approach to Neuro-AI governance