AI is not one thing
A chatbot drafting an invitation and a system affecting healthcare or liberty should not share one risk model. Context and consequence matter.
The question is delegation
Human Scale is neither anti-AI nor acceleration-at-any-cost. It asks what humans should delegate, what should remain meaningfully human-controlled, who gains capability, who gains power, what skills remain necessary, and what happens when the system is wrong.
AI should expand human capability without making human judgment, responsibility, skill, privacy, economic agency, or practical exit disappear by default.
Diagnosis
Augmentation ≠ automation
Assisting a human and replacing the human role are different designs. Ask what capability disappears and whether it still needs to exist when automation fails.
Fluency ≠ reliable knowledge
Generative systems can sound confident while wrong. Verification intensity should rise with stakes.
Human-in-the-loop can be theater
A reviewer without evidence, time, expertise, override authority, or appeal power is not meaningful control.
AI can scaffold skill or replace practice
Ask which skills people still need for judgment, safety, learning, creativity, and recovery when the tool is absent.
Agents change authorization
When AI can send, spend, modify, operate, and transact, permissions, limits, logs, confirmation thresholds, and revocation become core design.
Memory creates continuity and surveillance
Persistent context can remove repeated explanation and also become sensitive profiling. Inspection, correction, deletion, portability and use boundaries matter.
AI can expand accessibility
Captioning, speech, description, translation, alternative interfaces and cognitive assistance can increase independence—when dependable enough for the people relying on them.
Education must redefine mastery
If AI can produce the output, schools still need to decide which knowledge and skills people require to judge, question, learn, and function without it.
Productivity has distribution
AI gains can become wages, prices, profits, output, ownership, public benefit, fewer jobs, or human time. Technology does not choose the distribution.
Capability can concentrate or spread
Frontier systems require capital, chips, energy and talent; APIs, local models and open ecosystems can distribute capability. Contestability matters more than ideology.
AI has a physical footprint
Chips, data centers, electricity, cooling, networks and supply chains make digital intelligence real infrastructure.
Synthetic media changes trust
Cheap generation expands creativity and lowers the cost of impersonation, fraud and propaganda. High-stakes provenance becomes more valuable.
AI companionship is real as experience
People can feel attachment or comfort without AI relationships becoming equivalent to embodied, reciprocal human relationships.
AI consciousness is unsettled
Human-like self-report is not proof of subjective experience; certainty that machine consciousness is forever impossible also exceeds what we know.
Advanced risk deserves serious uncertainty
Cyber, biological, autonomous, military, infrastructure and systemic risks deserve evaluation without pretending catastrophic outcomes are either certain or impossible.
Field Guide
Match verification to stakes. Grocery lists and legal decisions do not need the same evidence burden.
Keep one important skill alive. Let AI extend a skill without eliminating all practice when that skill still matters during failure or independent judgment.
Separate drafting, deciding, and acting. Good generation quality is not automatically permission to execute consequential actions.
Review memory and permissions. Inspect stored context, connected accounts, spending authority, recipients, file access, device control and revocation.
Preserve direct human channels. Essential systems should allow escalation to a responsible person when the case is wrong, unusual, unsafe, inaccessible, or rights-bearing.
Run an AI workflow audit. Measure time, error, rework, skill, privacy, vendor dependency, money, output, and what happens when AI fails.
Program
Risk-tiered governance
Entertainment, drafting, medicine, credit, criminal justice, critical infrastructure, weapons, and dangerous capabilities justify different controls.
Appealable consequential decisions
When AI materially affects housing, employment, healthcare, education, credit, benefits, insurance, or liberty, notice and meaningful correction/appeal matter.
Agent accountability
Systems that act should support scoped permissions, limits, identity, logs, confirmation thresholds, revocation, anomaly detection, and rollback where feasible.
Portability & competition
Prompts, memory, workflows, evaluations and tool integrations should be movable where practical so useful intelligence does not become a permanent vendor trap.
Legible public-sector AI
Use AI to translate, explain, search, prefill, and track while linking authoritative sources, exposing uncertainty, and preserving human escalation.
Labor transition without false certainty
Test training, portable benefits, ownership, competition, wage growth, shorter work, income support, entrepreneurship, and broad AI access without assuming either universal unemployment or frictionless job creation.
AI that develops students
Tutoring, translation and teacher support should coexist with foundational knowledge, independent reasoning, privacy, assessment integrity, and human relationships.
Provenance for high-stakes media
Authenticate official statements, evidence, records and consequential communications without stigmatizing ordinary synthetic creativity.
Frontier evaluations & incident learning
Evaluate cyber, deception, tool use, biological assistance, security, robustness and control as capabilities evolve; scorecards should not become safety theater.
Compute is infrastructure
Chips, data centers, grids, water, land, transmission, energy prices and supply chains belong in AI planning.
Keep the right to a person
Routine automation can be excellent. Rights, safety, unusual cases, accessibility failures, and relational care often still require responsible human access.
Frameworks worth using
NIST AI Risk Management Framework
Govern, map, measure, and manage AI risk across technical and socio-technical systems. NIST ↗
NIST Generative AI Profile
GenAI-specific risk considerations and actions extending the AI RMF. NIST ↗
OECD AI Principles
International principles around inclusive benefit, rights and democratic values, transparency, robustness, safety, and accountability. OECD.AI ↗
UNESCO Recommendation on AI Ethics
An international normative framework emphasizing dignity, human rights, fairness, transparency, oversight, sustainability, and governance. UNESCO ↗
These frameworks are reference points, not scientific proof of Human Scale’s values. AI governance remains a moving technical, legal, economic, and political problem.
A Human Scale test for AI & automation
What human goal is the AI serving?
Is this augmentation, automation, or delegation?
What happens when it is wrong?
How reversible is the action?
Are stakes low or high?
Does human review have real authority and evidence?
What skill may weaken through disuse?
What capability becomes more accessible?
What data or memory is retained?
Can important data be inspected, corrected, deleted, or moved?
Can the user change providers?
Who receives the productivity gain?
What happens to workers and learners?
What physical resources does the system require?
Does synthetic output need provenance?
Are children or vulnerable people involved?
Is there meaningful appeal?
Are agent permissions scoped and revocable?
Can a responsible person be identified?
What evidence would make us change the design?
The direction
Writing remembers more than one mind. Machines lift more than one body. Computers calculate faster than one brain. AI extends this pattern into language, perception, reasoning, creation, and action.
We should use it—to remove drudgery, translate, teach, build, navigate bureaucracy, expand science and medicine, and make expertise more accessible—while continuing to ask what humans should still know, decide, practice, own, and be responsible for.
The best AI should make human beings more capable of living their own lives — not merely more efficient components inside systems controlled by someone else.