An institution is a set of people, powers and rules that can keep doing a job after an announcement has been forgotten. For AI, the job might be to give researchers access to equipment, check a supplier’s claims, hear a complaint or replace a failing system.
The design should begin with the missing power. People who cannot run an experiment need usable resources. People who cannot discover that a system affects them need information. People who can see a mistake but cannot change it need a remedy. These problems may be connected, but solving one does not automatically solve the others.
We now have enough evidence to compare several concrete approaches. None has earned a claim to solve the whole problem. Each offers something worth examining, along with a way it can fall short.
Public compute: permission is only the first step
Shared computing programs can give researchers a route to work that would otherwise require a private budget or partnership. Their success should be judged by usable access, including support, rather than the existence of a machine.
In March 2026, the US National Science Foundation reported that the National Artificial Intelligence Research Resource pilot had supported more than 600 research teams and 6,000 students. Its resources combine public and private contributions. These counts establish the administrator’s reported reach. They do not establish how much work would otherwise have gone undone, whether access became fairer, or whether dependence on particular suppliers fell. NAIRR two-year update
A useful evaluation would ask who applied, who received an allocation and who could actually use it. It would record waiting time, abandoned projects, technical assistance and completed work. It would distinguish a team receiving an award from a team running a successful experiment. Those are proposed measures; the participation counts do not answer them.
European and UK arrangements make different access choices. EuroHPC, the European public supercomputing initiative, separates routes by the scale and purpose of the request. Large allocations involve technical and peer review. On the access pages checked in October 2026, some systems were unavailable for new allocations because of demand. Free access for eligible users therefore did not mean capacity without limits. EuroHPC access modes, large-scale access
Guidance for the UK AI Research Resource (AIRR), updated in September 2026, listed Rapid and Gateway access routes, with defined allocation and time limits. Its earlier Innovator route had closed. These are operating rules, not evidence about the quality of completed research. An announced expansion target should also be kept separate from resources already available. UK AIRR guidance
Comparing these programs by advertised GPU hours alone would be misleading. A GPU is a computing chip, and different chips, software and tasks produce different amounts of useful work. A nominal allocation can also be hard to use if the team lacks help adapting its software. The sensible comparison is whether a defined group can complete a defined kind of work at a reasonable total cost.
There is a real tradeoff in selection. A program that funds only the most ready applicants may produce results quickly while leaving new institutions behind. A program that spreads resources very broadly may not give demanding projects enough capacity. Separate entry routes, training and transparent selection can be ways to address that tension. Their effect needs evaluation.
Public resources also require a continuity plan. Researchers should know what happens when credits expire, a partner changes terms or a service closes. Support for a project should not quietly become a requirement to remain with one supplier indefinitely. That is a proposed condition for public investment, not a finding that every existing program creates lock-in.
Independent evaluation needs a path to action
A public evaluator can develop skills and access that individual buyers do not have. The UK AI Security Institute’s December 2025 trends report brings together controlled evaluations of advanced models. It provides evidence about tested capabilities and safeguards, with important limits. Such tests are not a measure of all real-world harm, and the existence of a report does not show that a risky deployment changed. Frontier AI Trends Report
The next institutional question is what happens to a finding. Does a supplier need to respond? Can a buyer delay a purchase? Is there a responsible body with the authority to require a correction? Can a qualified outsider check that the correction addressed the problem?
Evaluation and enforcement can be housed in different organizations. That may protect independence or allow specialist work. It can also create a gap if no one is responsible for moving from a technical result to a decision. The relationship should be explicit.
An evaluation office should be able to report uncertainty. It should not be forced to call every system safe or unsafe when the evidence supports a narrower conclusion. At the same time, uncertainty should not erase responsibility. A decision-maker can explain why a use is permitted, limited or delayed, and what would prompt a review.
A register helps people find the system
A public register can make a hidden administrative tool visible. It can say what the tool does, which organization uses it and where responsibility sits. That is useful groundwork for scrutiny.
The UK’s Algorithmic Transparency Recording Standard requires records for tools within its stated central-government scope, subject to its rules and exemptions. Wider public-sector use is encouraged. The records are descriptions supplied by the organizations using the tools. Publication does not itself prove complete coverage, correct operation or a right to reverse a decision. ATRS scope policy
Two public records show why the distinction matters. The UK tax authority HMRC’s Business Propensity to Pay record describes human involvement but says the tool’s output is not itself appealable. England’s care regulator, the Care Quality Commission, has a Risk Categorisation record that marks the appeals field as not applicable because the category is not shared externally. These are the agencies’ statements about their tools, not findings that a person has no legal remedy against a later decision. They show why internal review and an affected person’s challenge must be examined separately. HMRC record, CQC record
If a model only advises a person, that can be an important safeguard. But advisory status does not answer every question. How much weight does the person give the advice? Can they see why it may be wrong? Can the affected party provide missing information before a consequential decision is made? Which final decision can be reviewed, and by whom?
The answer may lie in an existing appeals process rather than a new AI-specific one. Creating duplicate channels can confuse people and waste resources. A good record should therefore explain the relevant route in ordinary terms, or explain why no individual decision is involved. A general contact address is not automatically an appeal process. A technical monitoring team is not automatically authorized to correct a person’s case.
These questions suggest a way to improve disclosure: connect the technical description to the path a person can actually use. State what can be challenged, the responsible office, what information can be submitted, and what kind of correction is possible. Where the tool does not affect individual decisions, explain that carefully instead of filling every field with a stock answer.
This is a proposal for useful disclosure. The two records do not establish how common a gap is, whether people can navigate the wider system, or how often a challenge succeeds. Those questions need separate research.
An audit rule needs enforcement capacity
A requirement to obtain an audit can create useful scrutiny. It can also create a market for reports whose quality is difficult to judge.
Lara Groves and colleagues’ study of New York City’s algorithmic bias-audit regime used interviews with 16 experts and practitioners. It identified disputes about auditor roles, system scope and access to data. The early, purposive study does not estimate how much discrimination the rule prevented. It helps explain why the design of the audit relationship matters. Auditing Work
A later New York State Comptroller performance audit examined the city’s enforcement. In a re-review of the same 32 companies reviewed by the city department, auditors found potential non-compliance that the department had not identified. The report also raised concerns about complaint routing and technical support. Potential instances are not adjudicated violations, and this selected group does not measure citywide compliance. The findings show why enforcement procedures themselves need examination. Comptroller audit
Few complaints can mean few problems. They can also mean that people do not know a system was used, do not know where to complain, or cannot tell whether the rule applies. Oversight should not choose among those explanations without evidence.
Researchers studying public disclosure under the same city regime made a related point. Searching for reports does not reveal which employers are legally required to publish them. A missing report cannot be treated automatically as a violation when the underlying use and scope are unknown. Null Compliance
An effective audit system therefore needs more than a report requirement. The responsible body needs a way to identify relevant systems, inspect enough evidence, assess auditor competence and obtain corrections. The exact legal design will differ by jurisdiction. The general institutional test is whether a documented failure can lead to a change, rather than another document.
Buy the ability to correct course
Procurement offers a practical point of influence because a buyer can define conditions before dependence grows. A public buyer does not need to build every model itself to ask for inspection rights, usable records, service continuity and a way to replace the system.
These are proposed contract questions, not a summary of legal duties in every country. What evidence must the supplier retain? Who may inspect it? What happens when the model changes? Can the buyer export records in a form another service can use? Who pays for a transition? What remedy is available if the system fails the agreed test?
Some answers will cost money. Keeping alternatives ready requires work. Independent review needs skilled people. A buyer should compare those costs with the consequence of losing a service or being unable to correct a serious problem. The strongest requirements belong where dependence and harm would be greatest.
The same logic applies inside a public institution. Public ownership does not remove the need for scrutiny. A public system can be difficult to challenge or badly maintained. Its managers should face clear duties to explain performance, hear affected people and change course when evidence warrants it.
There is no single institution that can do every job well. A shared computing program, an evaluator, a procuring agency and a review body have different roles. The aim is to make those roles connect without letting responsibility disappear between them.
A program that can learn
Start with a bounded use, a clear comparison and a review date. State who is expected to benefit and who could bear a cost. Record the evidence before the announcement becomes a claim of success.
For public compute, test whether new groups can complete useful work. For evaluation, examine whether findings lead to responses. For a register, check whether people can understand the system and find the relevant route for a challenge. For procurement, test whether records and workflows can actually move. For workplace deployment, measure quality, workload, pay and time separately.
Keep failures visible. A program that reaches fewer people than planned may still teach us where access breaks down. An audit that finds little may be reassuring, or it may reveal that the auditor lacked data. A register may show that existing appeal routes are adequate, contradicting the expectation that a new route is needed. The evaluation should be designed to tell those possibilities apart.
The point of public power is not to control every technical choice from one office. It is to ensure that consequential choices remain answerable to the people who live with them. That requires resources to take part, evidence to judge, a route to challenge and workable ways to change direction.
Tomorrow will contain systems we cannot yet describe. We can still build institutions that make their promises testable and their mistakes correctable. That is a more useful form of ownership than merely being told, after the decisions have been made, that the future belongs to everyone.
Research and source note
This research draft was prepared with sources checked through 14 October 2026. It is a targeted evidence review, not a systematic review. Sources were selected for their relevance to specific claims, their methods and their ability to challenge a simple answer. The chapters distinguish measured outcomes, theoretical arguments, official program descriptions and proposed institutional tests. Links identify the works used; they do not imply that every part of a long report was read or every underlying result reproduced.
The original numerical work is the model-credit analysis in chapter 2. Its calculations were run on a fixed data snapshot, and a separate reconstruction checked the results against the input records. It is descriptive, with no estimate of market shares, corporate ownership or causal effects. The downloadable package states the selection, grouping and attribution rules, source identity, limitations and reproduction steps.
Examples introduced as possible situations are illustrations, not interviews, field reporting or personal memories. The brief includes no external verbatim quotation. The eight-word self-quotation in chapter 4 was checked against the author’s unpublished manuscript and is not counted as newly written prose. That manuscript supplies a stated aim, not independent evidence for factual claims.
A complete research draft is a stage of an argument, open to correction and further editing. It is separate from the finished commissioned PDF edition. No claim of peer review or completed original academic study is made.