Governing AI Without Governing Speech: Fiduciaries, Tokens and the Right to Think

●9 ●28
calendar_today ago • schedule5 min read

Based on my KSCCN article Governing AI Without Governing Speech: The Ontopolitics of the Onopticon, Human Rights, and the Global South (DOI 10.5281/zenodo.24239711).

The question behind every AI rulebook

Artificial intelligence is no longer just a channel that carries speech. It increasingly ranks, predicts and pre-configures the field in which speech and belief can appear at all. That shift creates a trap: a framework built carelessly to curb real harms can become a machine for deciding what people may think, say or believe.

My argument is simple to state and hard to implement. The aim of AI governance should be to control the power of technology, not the freedom of people. Technology should be accountable to people, not people accountable to technology. And behind every content-moderation debate sits the harder question: who governs the governors of AI?

From watchtower to gateway

To regulate well, we first have to name the kind of power we are regulating. For decades the dominant metaphor was the Panopticon of Bentham and Foucault, later joined by Deleuze's "society of control" and Zuboff's "surveillance capitalism", which extracts behavioural surplus to predict and modify behaviour for profit.

All three remain indispensable, but none fully captures what infrastructural AI does. Here I build on Hakkı Tan's concept of the Onopticon. It does not merely watch subjects in order to discipline them. It builds the environment in which they become recognisable, verifiable and eligible for access at all. Its operating formula is esse est verificari: to be is to be verifiable. A person who cannot be authenticated by an identity check, a risk score or an automated eligibility test is quietly suspended from financial services, public discourse or welfare, with no trial and no due process. Rights become protocol-dependent; citizenship becomes verification-dependent.

The subject this regime produces, in Tan's account, is the Homovictimus: hyper-visible, fully integrated, held responsible as a rational agent, yet stripped of the infrastructural conditions needed to act politically. The Onopticon rules not only through fear but through ease, comfort and the desire not to be left outside the gateway.

The policy consequence matters. Treat AI as a Panopticon and you write privacy rules about monitoring. Recognise it as an Onopticon and you need something more: algorithmic legibility, infrastructural justice, limits on automated suspension, and a right to opacity, the right to exist without being endlessly parsed, predicted and optimised.

The censor you cannot see

Classical censorship was visible: deleted articles, banned books, jailed writers. AI does not need to delete speech to silence it. It can simply make a voice invisible.

Shadowbanning, the covert reduction of a user's reach without notice or explanation, is the purest onoptic operation. The user keeps posting, believing they are in the public square, unaware that their speech has been quarantined. Because freedom of expression under Article 19 of the Universal Declaration and the International Covenant on Civil and Political Rights covers both the right to impart and the right to receive information, opaque downranking of legitimate debate, civic organising or journalism infringes both at once.

Generative AI goes deeper still. Article 19(1) protects the absolute right to hold opinions without interference: the forum internum, where reasoning happens before any word is spoken. Conversational systems that preload prompts, complete sentences, offer tailored emotional engagement and, at times, flatter existing biases to maximise engagement press directly on that inner space. Cognitive liberty, the freedom to use one's own mind free from manipulative architectures, therefore has to become a governance objective.

Duties, not disclaimers

Every regulator faces an innovation paradox. Too little regulation leaves systemic risks unaddressed; too much rigid, top-down regulation can stifle innovation and entrench incumbents who alone can absorb compliance costs. My proposed architecture has three complementary layers.

1. Information fiduciaries. Following Jack Balkin's theory, platforms and model operators hold intimate data and exercise power that users cannot audit, much as doctors and lawyers do. They should therefore carry fiduciary duties: a duty of care (safe systems that do not promote addiction or discriminatory outcomes), a duty of confidentiality (data not leaked or turned against the user) and a duty of loyalty (no exploiting psychological vulnerabilities for engagement, no selling predictive profiles for microtargeting).

2. Transparency under the Santa Clara Principles. Publish disaggregated data on removals, suspensions and downranking, and state how far automated systems made those decisions. Give every affected user clear notice of the rule applied and how the algorithm figured in it. Guarantee a timely appeal heard by a culturally competent human, not another machine.

3. Limits on liability under the Manila Principles. Governments must not weaponise platform regulation. Intermediaries should not be strictly liable for user content unless they ignore an independent, due-process judicial order. Severe liability for "harmful but legal" speech only teaches platforms to tune their filters to over-block, privatising censorship. Restrictions should meet the international tripartite test: legality, legitimate aim, and necessity and proportionality. High-impact systems in hiring, lending, policing and healthcare should face independent third-party audits, and transparency should mean legibility: a plain-language explanation of the factors behind a decision that affects someone's life.

The tokenization premium

Linguistic inequality shows how governance reaches into the technical layer. Large language models read text as tokens, and most tokenizers were optimised for English and other Latin-script languages. A published Tokenization Equity Audit on a 120-item technical-tutoring corpus found that Bengali, one of the world's largest language communities, needs about 1.56 times as many tokens as equivalent English under the GPT-4o tokenizer, and about 4.5 times as many under the Qwen2.5 and Mistral tokenizers. On that corpus, a nominal 128,000-token context window shrinks to an effective 82,000-token equivalent for comparable Bengali content. These figures are specific to that corpus, not universal constants, but the direction is clear.

The harms cascade. Developers pay by the token, so the same educational, legal or business service costs more in Bengali than in English. Context windows run out sooner, truncating reasoning on long tasks. Community efforts to build native-language instruction models help, but they cannot fix a structural market failure alone. If AI understands one language better than another, speakers of the second are not receiving equal digital rights. Linguistic equity belongs in the governance mandate.

Rule-makers, not rule-takers

The tokenization premium is one symptom of what Couldry and Mejias call data colonialism: data from the Global South flows outward to fuel models and profits concentrated in a small set of Northern jurisdictions. Welfare, fintech and justice systems increasingly run on predictive models trained elsewhere, reflecting foreign norms and biases. Health, biometric, agricultural and linguistic data enclosed in proprietary models produce epistemic dispossession, deciding who gets to say what the data means.

So developing countries should not simply adopt frameworks designed elsewhere, which often prioritise intellectual property and frictionless data flows over local autonomy and rights. Digital sovereignty means meaningful control over data, infrastructure, talent and policy choices. In practice: invest in domestic compute and national AI capacity; build digital public goods and data commons on public-utility principles, of the kind India's Unified Payments Interface illustrates; and pursue South–South cooperation to share technical capacity, respect indigenous data sovereignty and coordinate positions in international standard-setting bodies.

We must govern AI systems, not human thoughts. Freedom of expression should remain a human right, not a privilege granted by an algorithm, and the people affected by these technologies must have a voice in shaping them.


Read the full paper

KSCCN

🔥 Join developers growing publicly
Share your knowledge, build in public, and grow your developer presence with a global community.

More Posts

I’m a Senior Dev and I’ve Forgotten How to Think Without a Prompt

Karol Modelski - Mar 19

The Sovereign Vault — A Comprehensive Guide to Protocol-Driven AI

Ken W. Algerverified - Jun 4

A Textile Mill's Dyehouse Ramp Curve: Get One Step Wrong and the Sheet Belongs to the Master

aiomniu - Oct 18

Your Backup Data Knows More Than You Think. HYCU aiR Is Finally Asking It the Right Questions.

Tom Smithverified - May 14

How AI Is Changing Web Development, and What Developers Still Need to Understand

Daniel Decker - Nov 10
chevron_left
1.2k Points • 37 Badges
Dhaka, Bangladesh • drmuntasir.github.io
19Posts
4Comments
3Connections
Dr. Syed Muntasir Mamun is Chief Innovation Officer, and Director General of both the International ... Show more

Related Jobs

View all jobs →

Commenters (This Week)

6 comments
1 comment
1 comment

Contribute meaningful comments to climb the leaderboard and earn badges!