15 Nov 2026 The committees that decide what can be known
Joel Naoki Christoph
Most of the decisions an AI system embodies are made before you ever type a question. The deepest are written into technical standards that almost nobody outside the engineering world reads. If communication is a human right, its defenders need seats at the table where those standards are drafted.
By the time a chatbot answers a question, the decisive choices are already made. Some are in the training data and the company’s policies. Others are buried deeper, in technical standards: the specifications that determine which languages a machine can process, what information accompanies an image, and whether your data can go from one service to another. These documents are rarely discussed in public, though their effects can last for decades.
In July 2026, I took part in a panel on “The Politics of Knowledge in the AI Age” at the WSIS Forum in Geneva, convened by the International Telecommunication Union (ITU), the UN human rights office, and Harvard Kennedy School’s Carr-Ryan Center. The table gathered voices from government, a large video platform, and human rights work. Two of them matter for what I want to argue here. One was a technologist who helped write an open standard for content provenance. Another spoke for a platform that will have to obey it. The conversation kept returning to an uncomfortable point. The documents that decide much of what can be known in the AI age are ones almost nobody reads.
Knowledge has never been a neutral mirror of the world. Someone always decides what gets recorded, what gets trusted, and what gets found. Libraries decided which books to keep. School curricula decided which histories children would learn. Broadcast licences decided which voices reached the public. These were political choices, whoever made them and however they were described.
What has changed is the venue and the scale. Decisions about recording, trust, and retrieval now occur within digital systems at an enormous scale, and most of those systems are privately governed. When you ask a chatbot a question, something has already been determined about what it can retrieve, what it considers plausible, and what it will refuse to say. The politics of knowledge is the set of rules behind that determination.
Where the rules live
Many of those rules are written in documents called technical standards. A standard is an agreed-upon specification that enables independently built systems to work together. Standards do not determine every answer a system gives. They do something more basic. They define the formats, identifiers, and metadata through which information becomes legible, portable, and trusted. The ITU’s Landscape of ICT Standards, which aggregates specifications from standards bodies worldwide, lists more than 8,000 of them, from fibre-optic transmission to connected vehicles.1
In practice, they read like legislation for machines. That overstates it only a little. Standards are voluntary and agreed upon by consensus among participating organisations. Once a specification is embedded in products, procurement rules and infrastructure, it can work as de facto law for the systems that follow it, because a product that ignores the standard simply does not connect. A choice made once can govern for decades, since replacing installed infrastructure is slow and costly.
Nobody in these rooms is secretly legislating. The consequences are political all the same, whether or not anyone intends them. Three features explain why. Standards are durable, so an early choice lasts for years. They are written in rooms with uneven attendance. They operate beneath democratic scrutiny too, arriving as defaults that almost nobody changes.
The uneven attendance has a cause. Firms participate in standard-setting because interoperable standards open markets and reduce their costs, thereby strengthening the commercial incentive to participate. Affected communities, smaller economies, and most civil society groups have no comparable incentive and often no budget for travel and the slow hours of draft review. Some of those rooms are also gated. At the ITU, the study groups that draft standards are open to member states and to fee-paying members, so participation runs mainly through membership. There are narrower routes in: an organisation can join a Member State’s national delegation, and the ITU can waive fees for civil society or Global South groups. These openings depend on a government’s discretion or an administrative decision, which is a weaker footing than a standing right to attend. The result is a table that tilts toward those who profit from the outcome.
What can be known: the case of language
In his 2025 essay collection Decolonizing Language and Other Revolutionary Ideas, his final book, the Kenyan writer Ngũgĩ wa Thiong’o described how colonisation denied native languages the status of valid sources of knowledge, an argument he set out most famously in Decolonising the Mind in 1986.2 Digital systems can repeat that denial politely, through omission.
A language can be formally encoded yet remain digitally marginal, because the keyboard layouts, datasets and evaluation tools that make it usable are missing. Character encoding and keyboard layouts are themselves standards work, handled largely by ISO and IEC. What is machine-readable can be used to train AI models. What is absent from that data fades from the answers those models produce. Thousands of languages are served poorly today, and whole bodies of knowledge are close to invisible to the systems that now mediate what people find. That is a political outcome masquerading as a technical gap.
The Global Digital Compact, adopted by UN member states in 2024, calls for artificial intelligence that supports diverse cultures and languages.3 Encoding tables and dataset specifications will decide whether that promise means anything.
Who controls knowledge: the case of provenance
Consider a newer example. The Coalition for Content Provenance and Authenticity (C2PA) develops an open standard that attaches verifiable information to images and video, recording when a file was made and how it was edited.4 For a journalist, provenance can help verify the source of a photograph of an atrocity and whether it has been edited. In an information environment flooded with synthetic media, that capability protects the credibility of human rights documentation.
The same standard, designed without care, does the opposite. The metadata that authenticates a video can also identify the person who filmed it. Verifiability that protects a newsroom becomes a surveillance trail that endangers a dissident. Human rights groups understood this early. The organisation WITNESS joined the standards process itself and co-chairs the coalition’s work on threats and harms. That is why the technologist and the platform manager were at the same table in Geneva. The outcome depends on the details of the specification more than on the intentions of the people who wrote it.
When agency becomes a question of format
Between the language a system can process and the safeguards it offers, lies the question of human agency, and in standards, that question becomes concrete. Can people export what they have made, or is it captured within a single service? Can an automated decision be logged and questioned, or does it arrive without explanation or appeal? Can someone leave a platform without losing their contacts, history, and audience? A right to challenge a decision that lives only in a company’s terms of service can vanish with the next update. The same right, supported by a documented format and a portability requirement, remains in place after a change of management. Specifications can make these freedoms possible or make them impossible.
Communication rights arrived here first
Readers of this journal will recognise the pattern. The communication rights tradition has long argued that communication infrastructure is never neutral, as evidenced by the New World Information and Communication Order debates. The demand voiced at the WACC Partner Forum in Kathmandu, “nothing about us without us”, applies directly to standardisation.
Communities are already acting on that insight. A recent issue of Media Development described the Kasepuhan Ciptagelar community in West Java, which built and governs its own internet infrastructure, and the indigenous mobile operators Wiki Katat and TIC OMV in Mexico, which sell telephone and internet service nationwide on their own terms. These initiatives show that communities can build and govern their own connectivity rather than only consume it. The next step is harder to see. A community network still has to use the global formats. Whoever writes those formats decides some things for every network that uses them.
What a rights-based approach means
Momentum already exists. In 2023, the UN human rights office published a report on human rights and technical standard-setting, with recommendations for building rights considerations into the work of standards bodies.5 The Internet Research Task Force has published guidance for reviewing Internet protocols for their effects on human rights.6 The Geneva panel itself brought together the ITU and the UN human rights office. The two now run a joint work plan on standards and human rights.
Three measures matter most.
1. A one-page human rights note attached to an early draft, written in plain language, so that a specification’s likely effects are visible while the text can still change.
2. Funded participation for civil society. Because there is no commercial incentive for community groups to attend, someone has to pay for their seats. Funders who support media development can treat standards participation as fundable work, covering travel, fees and observer places.
3. Procurement rules that require privacy, portability and contestability safeguards in the standards that governments and large buyers adopt. Most firms follow the cheapest available option. The UN Guiding Principles on Business and Human Rights already assign companies a responsibility to respect rights.7 Procurement can make the rights-respecting option the cheapest one.
Proportionality keeps this credible. Many standards raise no evident human rights questions, and a short early screen can record that in a single line, so that fuller review concentrates where the stakes are highest: identity, content, and data flows. The screen should decide that, rather than an assumption about the subject matter. Even a transmission standard can affect access to information, and accessibility standards exist to serve disabled users. Inside companies, there is one more inexpensive step. Human rights teams and standards teams rarely speak with each other by default. Introducing them may be the most cost-effective measure available.
Where readers come in
None of this requires waiting for permission. Faith-based organisations, community media and civil society organisations built the communication rights agenda from the outside. The same approach applies now. The obstacle is that standards agendas are dense and hard to follow. What the field needs is a shared translation service that turns the technical agendas of the main standards bodies into readable alerts for non-technical groups, naming which draft touches identity, provenance or content moderation, and by when a comment would matter. A start exists. The Civil Society Alliances for Digital Empowerment project runs a hub of this kind, though it needs far more support to cover the field.8 Universities and training programmes can teach standards literacy the way they once taught media literacy.
There is also a role for the multilateral system. Standards bodies based in Geneva work a short walk from the UN human rights office. Regular exchange between the two communities, in both directions, would cost little and make a big difference.
We assume that great decisions about freedom are made in courts and parliaments, after the harm has occurred. In the AI age, many of them are made earlier, in a format, a field, a default. The editorial of a recent issue of this journal insisted that other technological worlds are possible. That is true. Some of those worlds are being written now, in committee rooms with empty chairs. The task is to fill them.
The author thanks Olivier Alais of the ITU for his comments on an earlier draft. The views expressed here, and any errors, are the author’s own.
AI Use Disclosure: In line with the editorial standards of Media Development, the author discloses that the AI assistant Claude (Anthropic) was used for structural and copy-editing assistance, working from the author’s remarks at the WSIS Forum panel and his own notes. The argument, the choice of examples and the final text are the author’s, who reviewed and revised the article and takes full responsibility for its content.
Notes
1. International Telecommunication Union. ITU-T Landscape of ICT Standards (accessed 2026): itu.int/itu-t/landscape
2. Ngugi wa Thiong’o (2025). Decolonizing Language and Other Revolutionary Ideas. New York: The New Press. The argument builds on his Decolonising the Mind (1986).
3. United Nations (2024). Global Digital Compact, adopted as an annex to the Pact for the Future: un.org/global-digital-compact
4. Coalition for Content Provenance and Authenticity: c2pa.org
5. Office of the UN High Commissioner for Human Rights (2023). Human rights and technical standard-setting processes for new and emerging digital technologies, A/HRC/53/42. See also the OHCHR study, Making technical standards work for humanity: ohchr.org
6. Internet Research Task Force (IRTF), Human Rights Protocol Considerations Research Group (2024). RFC 9620: Guidelines for Human Rights Protocol and Architecture Considerations, updating RFC 8280 (2017): rfc-editor.org/rfc/rfc9620
7. Office of the UN High Commissioner for Human Rights (2011). Guiding Principles on Business and Human Rights.
8. Civil Society Alliances for Digital Empowerment (CADE), co-funded by the European Union: cadeproject.org
Joel Christoph is a political economist working on AI governance, technical standards and human rights. He has been a Technology and Human Rights Fellow at Harvard Kennedy School’s Carr-Ryan Center and a fellow at the Centre for the Governance of AI.