Human centred AI governance and design stack

Public critique of AI now resembles the way people talk about meat and flying, full of moral gestures and lifestyle signalling while the underlying systems stay largely intact. Social media platforms reward this; outrage, virtue and counter‑signals all convert into engagement and that reward loop frames how most people encounter AI. Two loud poles dominate: one celebrates enthusiastic adoption and waves away data centres, labour conditions and cheating, the other sees those clearly and turns refusal into a badge of belonging. The leverage sits elsewhere, in how we design and govern data centres, models and institutions under material limits, where procurement, regulation and infrastructure choices actually get made.

I work on a human centred AI governance and design stack that treats models and interfaces as part of concrete economic and political infrastructures. My focus lies on small, local first systems that can run on owned or community controlled hardware under clear energy and water constraints and that serve as demanding partners in thinking and writing. Success means more capable, more agentic people who can interrogate their tools, see where they fail and turn them off while keeping their own practice intact.

Vision

• AI must stay small, situated and accountable. We build models that fit inside real power and water budgets and we keep them owned, inspected and replaceable.
• 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 be designed for a finite world. Let constraints on energy, water, hardware and labour set the boundaries for every model and infrastructure choice.
• 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.

Governance as infrastructure

• Treat the dominant, principle only approach to AI ethics as inadequate for learning systems and for the social, political and environmental impacts they create.
• Build richer ethical models grounded in context, institutions and material constraints and use those models as the basis for any serious AI governance work.
• Build governance around responsibility, traceability and contestation inside real institutions and 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.
• 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.

Consent and provenance in training data

• Treat the training corpus as appropriated labour until proven otherwise. Build on data that authors, artists and creators have consented to, licensed or released and record the provenance of what goes in.
• Pay and credit the people whose work trains a model. Where a corpus draws on living creators, route value back to them 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 ask what a model was trained on and trace an output’s debt back toward its sources.

Radical legibility

• AI must stay legible to the people who live with its consequences. Give users, auditors and organisers concrete ways to see how outcomes are produced and where data flows.
• 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.

Frugal, situated computation

• 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.

Against slop and substitution

• Cut AI slop. Limit mass generation, expose the material cost of computation and put 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.

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, take a deliberate counter move 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.

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.