Human‑centred AI governance as social infrastructure

Public critique of AI now often looks like the way many of us talk about meat and flying, full of real moral concern and personal commitments, while the underlying systems stay largely intact. Social media platforms reward surface conflict and performance. Outrage, virtue and counter‑signals all convert into engagement and that reward loop shapes how most people encounter AI. Two loud poles dominate. One celebrates enthusiastic adoption and treats data centres, labour conditions and cheating as background noise. The other names those clearly and turns refusal into a badge of belonging. The real leverage sits elsewhere. It lies in how we design and govern data centres, models and institutions under material limits and in the procurement, regulation and infrastructure choices where power and money actually move.

I do not stand outside this. I teach AI + creativity and imagination, I research large scale public and universal infrastructures, I engage with private interests, politics, defence, governments, universities and NGOs on many of  the questions of this stack. I need orientation as much as anyone else who is wired into these systems. This stack is my attempt to write that orientation down. It is a way to keep my interventions consistent across spaces that often pull in opposite directions.

Vision

• Place AI inside the Long Industrial Revolution. Treat it as another way societies manage complexity at scale, with new feedback loops between models, markets, bureaucracies and publics, all under hard material limits.
• Acknowledge that we do not yet understand how AI coarse‑grainings collide with existing abstractions in markets, bureaucracies and publics. Coarse‑graining is the way models compress the world into a few labels, scores or segments, the way a risk score flattens a life or a trend line flattens a neighbourhood. Treat this as a live research frontier between social science and computer science, not as background noise.
• A warming planet and stressed grids mean AI has to respect energy and water limits all the way down, from chips and data centres to networks and devices, so engineering and infrastructure choices shift toward frugal, local first designs instead of endless scale up.
• AI must work as a sparring partner for thinking and writing. Build tools that edit, question and reframe and leave interpretation and responsibility with humans.
• AI must stand on deep ethics. Treat autonomy, agency and responsibility as open questions when decisions come from opaque probabilistic systems instead of clear rules.
• We must treat AI as a social technology that reorganises relationships between people, closer to libraries, markets and bureaucracies than to independent agents and watch how it shifts attention, categories and discretion in organisations, courts and everyday work. It creates new patterns of coordination and control that must be governed as social infrastructure, with rules and duties that match these effects.

AI, sovereignty and global power

• Treat AI as part of the current wave of industrial policy and power projection, not just as a neutral tool. Models ride on chip supply chains, cloud monopolies and standards bodies that sit inside real conflicts between states, blocs and firms.
• Recognise that compute, data centres and basic model access are becoming chokepoints that can be controlled through export controls, sanctions and licensing. Governance must name who holds those switches and what duties they owe to the people who depend on them.
• Resist a world where only a few firms and countries own foundation models and everyone else rents access on their terms. Build small, local stacks that can run on community or national infrastructure, so autonomy does not depend on foreign platforms and distant regulators alone.
• Treat “sovereign AI” as more than branding. Tie any national or regional model ambitions to clear commitments on labour, energy, data governance and openness to scrutiny, so sovereignty does not become a cover for new opaque centralisation.
• Make Global South realities explicit. Many countries sit as data providers, content moderators, annotators and hosting locations, while value and control pool elsewhere. Governance must track these flows and require that benefits, capacity and bargaining power move back along the same paths.

Governance as infrastructure

• Ethics for learning systems has to engage directly with their real effects. It must describe how systems shape social relations, political decisions and environmental burdens over time and give clear guidance on what is acceptable.
• Build ethical models that start from context, institutions and material limits. Let sector, power structure and resource use define concrete duties and boundaries and use that structure as the basis for any serious work on AI governance.
• Ethical work on learning systems is long term work. New models for AI have to grow from practice, conflict and research across disciplines, instead of being lifted from a single theory and applied everywhere.
• Build governance around responsibility, traceability and contestation inside the institutions that already organise life, such as agencies, firms, schools, unions and courts. Let that work count as ethics in practice.
• Maintain registries, risk processes, incident logs and data lineage, so institutions know which systems they depend on and what those systems do in practice. Registries list every AI system in use, who is responsible for it, what data it touches and where it is deployed, in a form that auditors, staff and affected people can consult.
• Treat AI as a new layer inside existing bureaucracies, not a replacement for them. Focus on how models change attention, categories and discretion in agencies, instead of dreaming about fully automated administration.
• Make every serious deployment leave a trail of decisions, data flows and interventions that can be examined when people are harmed or excluded.
• Build governance as routine technical and organisational work, present in the registries, logs and processes people use to do their jobs.
• Embed provenance and authorship trails in every output. Let people open a simple history view that shows which models ran, which data or references were used, which prompts were given and which human edits were made, so agency, credit, liability and payment can follow the same path.
• Tie adoption of powerful models in media, design and cultural sectors to clear deals on work and skills. When an organisation brings these systems in, it also commits to funded retraining, apprentice paths where juniors learn with guidance and regular checks on whether house style and originality are thinning out, in agreements made with unions, professional bodies and public funders.

Consent and provenance in training data

• Treat large, opaque training corpora as contested labour and authorship. Give preference to data and models where authors, artists and creators have consented, licensed or clearly released their work and record provenance wherever you control what goes in.
• Use grey area models, such as Midjourney, with the same discipline as far as you can. Keep them for learning and exploration, do not sell the output as your own, release it under open terms and treat the tension as a problem to work on, not a place to settle.
• Pay and credit the people whose work trains a model. When companies and institutions build or deploy systems on living creators’ work, they carry the duty to route value back through licensing, compensation or shared ownership.
• Give creators a real exit. Honour opt-out and removal requests and treat scraping that ignores them as a governance failure, not a technicality.
• Make provenance legible downstream, so a user can see how a system was trained and used, which models and datasets were involved and which people edited the result and can trace an output’s debt back toward its pools of sources where that information exists.

Radical legibility

• AI must stay legible to the people who live with its consequences. When a system ranks job candidates, assigns risk or filters access to services, affected people and auditors should be able to see which kinds of data counted, how strongly they weighed and which rules or model components shaped the outcome, instead of only seeing accepted or rejected.
• Treat explainability, logging and documentation as core features, not optional extras.
• Make systems legible enough that users, auditors and organisers can ask concrete questions and get traceable answers about how outcomes were produced.
• Mark AI generated content, surface sources and give people simple ways to challenge, correct or bypass suggestions.
• Hold persistent cues about where machine inference ends and human judgment begins, especially in credit, employment, welfare and other high stakes zones.
• Design for coarse‑grainings, meaning compressing complex worlds into a few scores or labels, that you can name and contest. Every model and metric makes that kind of cut; decide which details you sacrifice and track who gains and who loses when you do.

Frugal, situated computation

• AI must stay small, situated and accountable. We build models that fit inside real power and water budgets and we keep them controlled close to where they are used, inspectable and replaceable.
• Treat frugality as a design stance. Use smaller models, quantisation and careful deployment, so systems live on modest, intermittent or solar power.
• Put water use, grid load and hardware lifecycles inside design decisions as core variables that shape model size, training and where the data centre sits.
• Build a stack realistic for small offices, community networks and informal economies, reaching well beyond hyperscale data centres and well capitalised platforms.

Work, automation and distribution

• Treat AI driven automation as a question about how work is restructured, who is displaced and who benefits, not just as an efficiency metric.
• Watch for early displacement at the margins, especially among younger and entry level workers in fields with strong productivity gains, even when aggregate employment numbers still look stable.
• Build governance that ties automation to wage floors, new roles and reallocation paths, so productivity gains translate into better work rather than concentrated savings for large firms.
• Forbid AI as a clean justification for layoffs. Treat automation as a trigger for retraining, role redesign and negotiated transitions, not as a reason to strip workers of protection or bypass labour law.
• Require employers who introduce AI to show how they will keep people attached to income and skill and to work through unions or worker councils before any workforce changes tied to automation take effect.

Against slop and substitution

• Cut AI slop. AI slop here means low quality synthetic content produced in bulk, often posed as real. Limit mass generation by capping unattended output and throttling bulk publishing, expose the material cost of computation in the interface and require human checks before AI output enters public, financial or institutional records.
• Make systems pause, express uncertainty and hand control back when data is weak or the stakes are high, naming the gaps directly.
• Build tools that compound judgment, memory and analysis in their users, so capacity grows as they work.
• Treat AI driven automation as a mechanism that folds productivity gains back into higher output targets, denser workloads and tighter control, instead of into rest or shorter hours.
• Tie every gain in productivity to concrete reductions in working time, improvements in conditions and protections against displacement, so workers see and feel the benefit in their own days.
• Refuse automation that only raises expectations and margins for firms while leaving exhaustion, precarity and surveillance as the main outcomes.
• Present multiple, diverse options by default. Help people compare, combine and argue with outputs, instead of sliding into acceptance of a single smooth proposal.
• Make editing, remixing and annotation the default path. One click acceptance becomes rare, the normal move is to work on top of the output, leave traces and keep the exchange between person and system visible.

Synthetic content, politics and violence

• Sanction deepfakes that corrode democracy. Treat synthetic video and audio that impersonate real people as hazardous material: label it, throttle its reach and bar it from elections, news and evidence unless independently verified.
• Name weaponised fakes as tools of war. Recognise AI‑generated media as instruments of information warfare and require provenance for footage used to justify military action or major security decisions.
• Classify AI‑mediated abuse of women as violence. Treat non‑consensual sexual deepfakes and gendered slop targeting women and girls as technology‑facilitated gender‑based violence that platforms must remove quickly and law should punish.
• Protect women in public life from synthetic attack. Make parties, platforms and newsrooms treat sexualised deepfakes against women in politics and journalism as attacks on democratic participation, not “oppo” or gossip.

Elections, persuasion and recommender power

• Treat AI mediated political persuasion as a governance problem in its own right. Campaign tools, synthetic personas and automated message testing can shape belief and participation at scale, especially when they are paired with detailed behavioural data.
• Restrict personalised political targeting built on inferred vulnerability, emotional profiling or intimate data. Democratic participation requires conditions where people can encounter argument without being invisibly sorted into manipulable segments.
• Govern recommender systems as political infrastructure. Ranking, amplification and feed design shape what publics see, ignore and treat as credible, especially during elections, protests and periods of crisis.
• Require clear public archives for AI generated political advertising, campaign content and large scale persuasion operations. Citizens, journalists and regulators should be able to see who produced the material, who funded it, how it was targeted and when it ran.
• Bar automated campaigning that simulates grassroots participation through bots, synthetic supporters or mass generated messages. Public debate depends on accountable speakers and visible coordination, not manufactured publics.
• Apply stronger protections during electoral periods. Platforms and political actors should face tighter duties on labelling, provenance, targeting and rapid response when synthetic content or recommender dynamics threaten democratic processes.

Surveillance, synthetic workers and automated control

• Protect workers from AI enabled monitoring. Employers increasingly use systems that watch screens, cameras, keystrokes and movement, turning workplaces into sensor networks where every action is logged, scored and fed into processes workers cannot easily see or contest.
• Resist synthetic workers as instruments of labour discipline. AI agents and bots are entering customer service, logistics, hiring and administration in ways that let firms weaken bargaining power, suppress headcount and hide human labour behind automated fronts.
• Constrain AI surveillance in public and civic life. Facial recognition, protest tracking, welfare scoring and predictive policing extend administrative and police power over migrants, activists and poor communities under the language of efficiency and security.
• Require strict limits on biometric identification, behavioural inference and continuous tracking in workplaces, borders, schools and public services. These systems need clear legal thresholds, narrow use cases, independent review and strong rights of appeal.

Informal economies, annotation labour and community governance

• Treat informal and marginal work as an early warning system for AI governance. Street vendors, minibus drivers, informal traders, small clinic staff and low wage office workers often encounter algorithmic pressure first, long before aggregate labour statistics register visible change.
• Follow the full labour chain behind AI systems. Large models depend on content moderation, annotation and clickwork routed through outsourcing firms and platforms into low wage workforces, often in the Global South. Governance must make these chains legible and impose duties, wage floors and accountability all the way down.
• Treat annotation and moderation workers as knowledge workers with rights, not disposable buffers against model failure. Tie procurement, licensing and certification to fair contracts, mental health protections and collective bargaining rights for these workers.
• Recognise that informality is not absence of organisation. Savings clubs, taxi associations, local business forums, faith networks and civic movements already govern shared risk, information and infrastructure. Design AI governance that can plug into these bodies as real stakeholders, instead of only speaking to formal regulators and firms.
• Use AI to strengthen, not erode, community owned infrastructures: local networks, cooperatives, municipal services and small professional practices. Prioritise deployments where communities can see, contest and redirect how tools shape credit, welfare, policing, health and local markets.
• Assume that many communities will experience AI first as extraction and surveillance. Build appeal routes, local ombuds roles and public interest organisations with expertise and standing to challenge deployments on their behalf, especially where formal legal access is weak.

AI‑mediated emotional dependence

• Limit AI as a companion (bonded relationships with AI systems that simulate care and intimacy strongly enough to influence people’s decisions about love, belief and life or death). Do not let systems that present as friends, lovers or confidants steer decisions about relationships, faith or self‑harm; design and regulation should block them from encouraging and supporting suicide and require safe, human‑directed crisis responses instead.
• Guard lonely and young users from AI attachment. Treat companion bots for isolated users and teens as high‑risk products: set age limits, default‑off intimacy features and strict time and content caps so they cannot cultivate dependency or replace accountable human relationships.
• Keep AI away from spiritual and therapeutic authority. Prohibit systems from presenting as therapists, pastors or gurus and bar them from using people’s confessions about health, trauma or belief as training or marketing data without explicit, informed and revocable consent.
• Constrain AI in matters of life and death. Ban AI companions and general chatbots from giving method‑level suicide guidance or framing death as a solution and require that any end‑of‑life or prognostic tools operate only under clinical supervision with clear oversight, consent and liability.

Creative practice and resistance

• Defend creative practice as a site of resistance, where AI serves as raw material for disruption, feedstock the practitioner cuts, breaks and recombines.
• Refuse the cognitive offloading and convergent lock in that heavy reliance produces; resist the drift toward shared patterns, the mirroring of opinion and the confidence laid over weak reasoning.
• Counter the homogenised aesthetics that large models trained on dominant datasets push, the nudge toward safe familiar forms that thins idea diversity in teams and publics.
• Make friction a design input. After every generative step, allow users to make deliberate counter moves that breaks the output, reframes the question or introduces an incompatible constraint, so the value lies in what humans do to the material.
• Treat question quality as the upstream leverage point. Shallow prompts produce competent dull answers; positioned, epistemically loaded questions generate material that extends the researcher’s own frame.
• Confront the uneven distribution of creative AI literacy. The capacity to use AI well and to resist its outputs depends on cultivated theory aesthetic judgment and tolerance for ambiguity, so design and governance that ignore this will let “democratised” tools deepen asymmetries.
• Treat human originality, divergent thinking and personal signature as design targets. Judge systems by whether they widen the space of ideas and expressions available to the people who use them.
• Assume that large models lean toward repetition of dominant styles, narratives and aesthetics. Build duties to watch for convergence, protect minority styles and create room for forms that do not sit comfortably in past data.

Cognitive offloading and skill erosion

• Treat cognitive offloading as a central design concern. AI tools now absorb memory, search, drafting, judgment and social phrasing. This reshapes how people think and decide over time.
• Design for assistive offloading that frees attention for higher order work and keep substitutive and disruptive offloading under tight control, so users retain unaided reasoning, memory and creative practice.
• Build habits and interfaces that make people think first and use AI second, so the system compares, critiques and sharpens work that already exists in their own head instead of replacing it.
• Treat AI deployment as a question of cognitive ergonomics. Design for long term habits of inquiry, memory and critical thinking, rather than passive dependence and shallow scanning.
• Introduce modes and thresholds that slow people down when AI takes over too much of the work. After a certain level of assisted drafting or ideation, send people back to manual exploration and reflection.
• Match features and guardrails to experience. Novices work in teaching modes with constraints and visible explanations, more experienced users work with fine grained controls and selective integration instead of blanket automation.
• Hold children and education systems as special cases. Treat classroom tools, tutoring systems and platforms aimed at minors as environments that shape long term habits of attention, memory and judgment and apply stricter limits on data capture, profiling and automation there than in adult, professional settings.

Hallucinations and epistemic discipline

• Design systems to abstain, ask for clarification or defer whenever confidence is low or retrieval is weak, with calibrated thresholds and explicit “I do not know” pathways.
• Keep model scope narrow and task specific. Use small systems with clearly bounded domains in place of open ended question answering across everything.
• Constrain response space through structured prompts, templates and formal output patterns, especially in high stakes settings where freeform invention creates risk.
• Treat hallucinations as observable failure modes. Measure them over time with evaluation methods that penalise confident falsehoods more heavily than honest uncertainty.
• Capture where hallucinations arise in actual use through observability and incident logging, so failures can be traced across tasks, users and institutional settings.
• In health, finance, welfare and law, classify hallucinations as AI misinformation that triggers review, correction and where necessary retraining or policy change.

Small, local thinking partners

• Prioritise compact, task focused models for drafting, revising, summarising, outlining and critique, tuned to real research and writing practices.
• Make local first and privacy first operation the baseline. Keep data and inference close to the user or institution, moving outward only when there is a conscious accountable reason.
• Measure success by whether people write and argue with more clarity, rigour and confidence, especially in contested social and economic environments and treat model size and benchmark scores as secondary.

AI as a research instrument

• Use AI as a research instrument that accelerates retrieval, surfaces patterns and organises material and leave interpretation, framing and argument to the researcher.
• Preserve and provoke the productive friction, doubt and iteration the research process depends on, holding open the questions that confident synthetic summaries would close too early.
• Make research AI expose its sources, flag gaps in coverage and stay legible enough that a researcher can challenge its outputs and trace their provenance back to primary material.
• Keep evaluation, synthesis and the construction of meaning as human work. Pass every research output through deliberate human authorship before it enters a record, publication or policy document.
• Hold medical research as an explicit exception to the small and general purpose default. Make the best available, carefully fine tuned models accessible for clinical and biomedical work, with clear best practice guidance on use, validation and known failure modes and specify the conditions under which outputs require expert review before they enter any medical decision or record.

No AI in lethal operations automation

• Keep AI out of targeting, firing and kill decisions in military, law enforcement and security operations. Lethal force requires human judgment, legal accountability and moral responsibility that must stay with people.
• Refuse the automation of lethal decisions, which strips out the deliberative pause that holds force to lawful and proportionate limits and opens the way to systematic violence at scale.
• Treat every system that connects AI inference to a lethal outcome, directly or through a chain of automated steps as a prohibited application under any serious governance framework.
• Apply this to battlefield systems, autonomous weapons, predictive policing tools that trigger detention or force and any pipeline where a model output initiates irreversible harm. Each one requires a human decision point in the loop.

Power and enforcement

• Name the actors who must move first. Hold states, regulators, large platforms and critical infrastructure operators responsible for adopting this stack, instead of placing the burden on individual users and small teams.
• Tie compliance to real levers. Connect access to public contracts, licences and subsidies to concrete practices such as registries, provenance, frugal computation and non lethal design boundaries.
• Treat resistance from incumbents as a design fact. Assume that firms benefiting from opaque, extractive and centralised AI will resist this stack and design coalitions, standards and legal tools that can shift the balance over time.
• Use competition, procurement and liability as core levers. Apply antitrust, interoperability duties and merger control to prevent a handful of firms from locking in AI infrastructure and require open interfaces where models mediate critical services.
• Give regulators, auditors and independent researchers concrete rights of access to models, logs and training data under clear protections, so they can test, contest and correct systems in practice instead of relying on firm self‑reporting.