AI standards are invisible infrastructure. Show who got to write the blueprint

THE SIGNAL IN ONE SENTENCE

Standards are the quiet machinery behind modern life. They help one plug fit another socket, one network talk to another network and one laboratory understand what another laboratory means by a test result. Most people never read them. Plenty of people live with their consequences. Artificial intelligence is getting the same invisible machinery. On October 23, ITU's AI for Good platform published an account of a discussion among the leaders of the International Organization for Standardization, the International Electrotechnical Commission and ITU's Telecommunication Standardization Bureau. Their subject was the difficult trip from agreeable AI principles to requirements that can be implemented, tested and recognized across countries and industries. The speakers landed on three priorities: capacity building, investment and collaboration. That is a useful list. It is also missing a public receipt. The plain signal is this: show who got to write the blueprint. For every AI standard that may shape public procurement, certification, regulation or market access, publish who proposed the requirements, who could participate, whose evidence counted, what changed after disagreement and how affected people can challenge the result. This is not a demand for a ceremonial guest list. A list of names cannot show whether a consumer advocate had the same practical influence as a company sending a delegation of specialists. It cannot show whether a small country could afford years of meetings, whether documents arrived in usable languages or whether a worker's evidence altered a requirement. What we need is a participation receipt. Start with the problem statement. Who said a new standard was needed? Which failure, market barrier, safety concern or policy goal was it supposed to address? Who drafted the first proposal? Who funded the research used to define the problem? Then show the room. List the countries, organizations, sectors and roles represented during each major stage. Include the membership route, fees, travel requirements, time zones, languages, accessibility support and confidentiality rules. Record who was invited but could not participate, and why. Next, trace the evidence. Link each technical claim to the paper, field test, incident record, public consultation or expert judgment that supported it. Say whether that evidence came from one market, language, population or type of system. Show which submissions were accepted, rejected or left unresolved. Record dissent. A clean final document can hide the arguments that matter most. Publish minority positions, unresolved objections and the reason a committee chose one path over another. If the language changed between drafts, provide a revision history that a normal reader can follow. Finally, open a route back in. State when the standard will be reviewed, what evidence can trigger a revision, who can file a challenge and how procurement teams or regulators should report failures in practice. That receipt matters because standards can exercise real power without looking like law. Sergio Mujica, ISO's Secretary-General, told the summit session that standards bodies do not make policy or regulation. Their role is to support implementation by turning ambitions into definitions, processes, technical requirements and methods of verification. That distinction is real. It is not the same as saying standards are optional in practice. A government can reference a standard in procurement. A regulator can use one to explain acceptable practice. A certification scheme can make conformity a condition for selling into a market. A large buyer can write a standard into its contracts. Once that happens, a technical document becomes a gate even if no parliament voted on every line. The gate may be useful. Shared definitions can reduce expensive confusion. Compatible interfaces can keep countries from building isolated systems. Repeatable tests can make claims easier to compare. A public buyer should be able to ask two vendors for evidence using the same basic language. The risk is that the gate fits the people who designed it. A test may assume infrastructure that small organizations do not have. A documentation rule may be easy for a large vendor with a compliance department and punishing for a local developer. A benchmark may reflect one language or market. A process may count a country as represented even when its delegate lacked time, money or technical support to shape the text. The ITU account confronts part of this problem directly. Philippe Metzger, Secretary-General and CEO of IEC, described international standardization as a bottom-up, multi-stakeholder process and urged governments to mobilize experts from industry, academia, civil society and other communities. Seizo Onoe, director of ITU's Telecommunication Standardization Bureau, described training, fellowships and implementation support intended to help developing countries participate. Mujica said nearly half of the roughly 80 countries participating in the joint ISO and IEC committee on AI are developing countries. The same account immediately adds the sentence that should sit above every standards meeting: representation alone does not guarantee influence. That is the centre of the story. Capacity building should not mean teaching people to comply after the blueprint is finished. It should include money, time, translation, technical assistance and access to evidence while the blueprint is still editable. Investment should not stop at data centres and certification services. It should fund public-interest participation, open testbeds, reference implementations and independent evaluation in countries with different infrastructure, languages and social conditions. Collaboration should produce more than a row of partner logos. It should leave a decision log. The ITU article gives public procurement as one possible testing ground. The example is an AI-enabled healthcare system, where purchasing immediately raises questions about performance, safety, impact and assurance. That is a good place to make the participation receipt practical. Before a public buyer references an AI standard, it could ask five questions. First, fit. Was the standard tested on systems, populations and operating conditions similar to the proposed use? Second, access. Could smaller vendors, public-interest researchers and affected communities see the relevant requirements and submit evidence without prohibitive cost? Third, voice. Who shaped the requirement that will decide whether a product passes? Were patients, workers, disability advocates and frontline professionals involved early enough to change it? Fourth, proof. Does conformity mean a product was tested, an organization documented a process or an independent body confirmed a result? Those are different claims. A certificate should say exactly what was examined. Fifth, repair. When real-world use reveals a failure, who receives the report and how quickly can the standard, test or procurement condition change? Public buyers should attach the answers to the tender. Vendors should attach them to assurance claims. Standards bodies should maintain them beside the document. Regulators should cite them when a standard is incorporated into guidance. The result would not make every technical meeting democratic. It would make influence inspectable. Interoperability is the strongest argument for doing this work well. Onoe used manipulated media as an example. If countries adopt incompatible technical approaches, a tool that checks authenticity in one place may not work elsewhere. A common foundation can let different systems exchange information without forcing every country to use identical technology. But interoperability is not automatically safety, fairness or truth. Two systems can exchange the same flawed signal perfectly. A shared test can reproduce the same blind spot across many markets. A certificate can make a weak requirement look reassuring. Compatibility tells us that parts work together. It does not tell us who chose the goal, whether the test reflects local harm or whether affected people can contest the result. Conformity assessment needs the same precision. The ITU account says ISO and IEC aim for one standard, one test and one certificate that can be recognized across markets. The efficiency is obvious. A manufacturer may not need to repeat equivalent tests for every jurisdiction. Buyers may gain a comparable form of evidence. The caution is equally obvious. Mutual recognition scales whatever the test contains, including omissions. If one certificate travels widely, the public needs a clear statement of scope, independence, test conditions, conflicts of interest, expiry and known limitations. No standard can freeze an AI system in place. Models change. Data shifts. Vendors update software. People use products in ways a committee did not anticipate. A standard that works on publication day can become stale, or reveal its blind spots only after deployment. The participation receipt therefore needs a life after publication. Connect field failures to revision cycles. Preserve version numbers. Tell procurement teams which edition they referenced. Publish change notes in plain language. Invite evidence from people who could not attend the original process. Treat the first certificate as the beginning of monitoring, not the end of doubt. There are limits to what the October 23 article proves. It is ITU's account of a summit discussion, not a new binding standard, a committee voting record or an independent study of adoption. It does not publish participant-level influence data, budgets, draft comments or measured outcomes from the capacity programmes it mentions. The speakers identify sensible priorities, but the article does not demonstrate that current processes consistently deliver equitable influence. That is exactly why the receipt is useful. It does not require us to assume the process is captured or fair. It asks the process to show its work. Standards will remain invisible to most people. That is fine. Sewers are invisible too, and civilization is grateful when they function. Invisible should not mean unaccountable. If AI standards are becoming the plumbing for procurement, certification and cross-border systems, publish the blueprint's authorship alongside the blueprint. Show the people, access barriers, evidence, disagreements, changes and challenge route. Then capacity building, investment and collaboration become more than three handsome words at a summit. They become a public record of who could help decide how the machinery works.