Humanity's Mirror

Artificial intelligence is often discussed as though it has introduced an entirely new political problem: an intelligence may acquire influence over human life without sharing human values, accepting human responsibility or remaining under meaningful human control.

The technology is new. The problem of power is not.

Human societies have repeatedly created institutions capable of shaping lives at a distance. Governments administer populations. Corporations organise work, information and capital. Financial systems determine access to credit and economic participation. Military organisations possess destructive force. Digital platforms influence what billions of people see.

Each can produce immense public value. Each can also become dangerous when its authority expands beyond clear limits, its decisions become difficult to inspect and the people affected lose any practical route of challenge.

This does not make a government, company or court equivalent to an artificial intelligence. Human institutions contain relationships, histories, laws and moral agents that a computational system does not possess. AI introduces its own problems of speed, replication, automation and technical opacity.

But the safeguards now proposed for AI reveal something important about humanity itself. We are asking how powerful systems can remain limited, supervised, auditable, reversible and accountable. These are not merely engineering questions. They are unresolved questions of institutional design.

AI has become a mirror. What unsettles us in the machine is partly what we already recognise in ourselves.

Capability is not legitimacy

A system may be highly capable without possessing the legitimate authority to make a decision.

An AI model might identify patterns in medical images, estimate the probability of fraud or rank applications faster than a human team. None of those capabilities answers whether it should determine a person's treatment, investigate their finances or restrict their opportunities.

The same distinction applies to human institutions. Wealth, expertise, force and administrative reach create capacity. They do not automatically create the right to use that capacity without limit.

Institutional decline can begin when this difference becomes blurred. A system assumes that because it can act, it is entitled to act. Temporary authority expands into permanent jurisdiction. Technical complexity discourages scrutiny. Responsibility fragments across departments, contractors and procedures until no one appears to have made the final decision.

The problem is not simply that malicious people sometimes gain power. Any durable system must assume that participants will include the mistaken, the ambitious, the frightened, the loyal, the careless and the self-interested. Good institutional design cannot depend on every person remaining wise. Its purpose is to limit the damage possible when they are not.

The same design problem

Credible AI-governance frameworks differ in legal force and scope, but several safeguards recur. NIST's voluntary AI Risk Management Framework organises work around governing, mapping, measuring and managing risk. The OECD principles emphasise human rights, transparency, robustness and accountability. UNESCO connects oversight to human dignity, traceability and the preservation of ultimate human responsibility.¹²³

These ideas have recognisable institutional equivalents.

Safeguard for AI	Equivalent institutional principle
Clearly limited authority	Defined jurisdiction and constitutional limits
Human approval for consequential actions	Lawful authorisation and accountable judgement
Independent evaluation and adversarial testing	External audit, inspection and judicial scrutiny
Transparent operational records	Reasons, decision logs and information rights
Minimum access to critical systems	Separation of powers and least-privilege access
Reversible decisions	Appeals, injunctions, correction and compensation
Independent systems checking one another	Courts, legislatures, regulators, media and civil society
Protection against majority harm	Fundamental rights and due process
Named responsibility	Liability that cannot disappear inside a process

The comparison is not exact. It exposes a shared principle: the greater the consequence of a decision, the stronger the evidence, authority, oversight and route of appeal should be.

Human approval is not human control

Placing a person somewhere in a workflow does not guarantee meaningful oversight.

A reviewer may lack time, expertise or access to the evidence required to challenge an automated recommendation. They may approve hundreds of outputs until approval becomes habitual. They may believe the model is more objective than they are, or understand that disagreement creates delay and personal risk. Under those conditions, a signature becomes ceremonial: the appearance of responsibility without the conditions needed to exercise it.

Meaningful oversight requires more. The reviewer must understand the decision and the system's relevant limitations; have enough time and evidence; possess genuine authority to intervene; and be protected when raising a justified objection. The record should show what information was available, what decision was made and who accepted responsibility.

Research also warns against treating human–AI behaviour as one simple tendency. Three Dutch experiments involving 2,854 participants did not find a general pattern of greater deference to algorithmic advice than to equivalent human advice. One experiment did find stronger adherence when advice matched group stereotypes, while a later civil-servant sample did not reproduce that pattern after a major public scandal had heightened awareness. The result is not that automation bias is imaginary. It is that context, incentives, prior beliefs and institutional learning shape how oversight works.⁴

A universal instruction to “keep a human in the loop” therefore says too little. The useful questions are whether that person can recognise error, refuse the output and remain answerable for the result.

Visibility and independent challenge

Power becomes harder to govern when its operation cannot be reconstructed.

Transparency does not require publishing every line of code, confidential record or security procedure. It requires enough visibility for an authorised and independent body to determine what occurred, what rules applied, what evidence was used and whether the system performed as claimed.

That creates several distinct requirements: records must exist; they must be comprehensible and retained; access must be controlled and logged; and the reviewer must not depend entirely upon the body being reviewed.

An internal assurance team can improve a system, but it cannot always establish legitimacy on its own. The institution being examined may control the evidence, define the test and decide whether the result is released. Independent scrutiny matters because incentives matter.

Disclosure without comprehensibility can conceal as effectively as secrecy. Thousands of pages released without structure, responsibility or a route of challenge may technically satisfy transparency while leaving the affected person powerless. The objective is not information for its own sake. It is the ability to establish responsibility and correct error.

Authority, access and reversibility

A system should know what it is permitted to do and what remains outside its authority.

For AI, that means more than preventing access to a database. The permitted purpose, affected population, operating conditions and prohibited uses must also be defined. A model approved to assist a trained professional in one setting has not automatically earned authority to operate independently in another.

Least privilege is therefore a political as well as a cybersecurity principle. A tool that drafts correspondence does not need authority to send payments. A diagnostic system does not need unrelated personal records. Human organisations also become safer when information, money, coercive authority and final judgement are not silently accumulated in one place.

Some friction is protective. Independent approval, divided authority and an appeal can slow a decision. The relevant question is whether that delay is proportionate to the harm it is intended to prevent.

Reversibility must also be designed before deployment. Who can suspend the system? Can the previous state be restored? How are affected people notified? Is compensation available? Can essential services continue if the supplier disappears or the system is withdrawn?

A theoretical right to complain is inadequate if thousands of people receive decisions faster than any appeal body can review them. An institution that cannot be stopped, corrected or exited has converted service into dependency.

Rights beneath preference

Human control over AI is sometimes presented as though transferring authority from a machine to a majority resolves the moral problem. It does not.

Majorities can support discrimination, collective punishment or the removal of inconvenient rights. Public approval may confer political authority, but it does not make every use of that authority just.

This is why accountable systems need principles beneath preference: human dignity, due process, proportionality, equal protection and defensible rights for minorities. UNESCO's recommendation deliberately leaves “quality of life” open to individuals and groups, but places human rights, fundamental freedoms and dignity around that freedom. It also states that responsibility must remain attributable to human actors and that consequential decisions require explanation, auditability and traceability.³

No framework can eliminate moral judgement. If wellbeing means only what the most powerful group presently desires, an AI optimised for “human wellbeing” may simply automate existing inequality with greater consistency.

Alignment must therefore ask more than whether a system follows instructions. It must ask whose instructions, under what authority, within which limits and with what protection for people who did not consent.

Law, guidance and the danger of false assurance

Governance documents should not be treated as though they all carry the same force.

NIST's framework is voluntary guidance and is being revised. The OECD AI Principles and UNESCO Recommendation are influential intergovernmental standards, not self-executing statutes. The European Union's AI Act is binding law whose provisions apply on a staged timetable: governance and general-purpose-model duties began applying in August 2025; enforcement powers and important transparency requirements followed in August 2026; many high-risk-use rules are now scheduled for December 2027, with product-embedded high-risk rules following in August 2028.¹²³⁵

The Council of Europe Framework Convention is a binding treaty in design, but at the Treaty Office's September 2026 status date it had not entered into force. Entry requires five ratifications, including at least three Council of Europe member states. Signature, ratification and commencement are different events.⁶

These distinctions are not legal trivia. Calling voluntary guidance “law” exaggerates protection. Calling a signed treaty operational hides the work still required. Calling a partly commenced law fully operational makes assurance sound more complete than it is.

Trust should follow evidence

Institutions frequently describe declining trust as a communications problem. Sometimes communication is poor. But trust can also decline because people cannot see how decisions are made, do not feel able to influence them and encounter no effective route of correction.

The OECD's 2024 trust survey covered 30 countries. Thirty-nine per cent of respondents reported high or moderately high trust in national government, while 44 per cent reported low or no trust. The sharpest contrast concerned political agency: 69 per cent of people who felt they had a say trusted national government, compared with 22 per cent among those who felt they did not.⁷

The survey is cross-national and observational; it cannot prove that giving any individual a formal consultation will create trust. It does show that perceived voice, evidence use, accountability and fair treatment are closely connected to institutional confidence.

Trust should not be demanded as an act of faith. It should emerge from repeated evidence that a system is competent, honest about its limits, responsive to challenge and capable of correction.

This changes the objective from making people trust technology to making technology and institutions trustworthy.

Responsibility cannot end at the interface

An AI decision may involve a model developer, data supplier, cloud provider, deploying organisation, procurement team, operator and final user. Every participant can describe itself as only one part of the chain.

Distributed production must not produce dissolved responsibility.

Before a consequential system is used, responsibility should be assigned for design, testing, deployment, monitoring, incident response, appeal and compensation. The affected person should not have to reconstruct an entire supply chain merely to discover who can correct a decision.

This principle applies equally to human institutions. Contractors, committees and technical systems can distribute work, but the public still needs an accountable authority. Delegation may move a task. It should not erase the duty attached to it.

A practical standard

When evaluating an AI system or powerful institution, ask:

1. What precise authority has been granted, and what remains outside it?
2. What evidence supports the claimed capability, and who can test it independently?
3. Who can stop or override a consequential decision?
4. Can the affected person understand, challenge and reverse the outcome?
5. What information, money or infrastructure can the system access?
6. Which rights remain protected even when a majority prefers otherwise?
7. Who is named as responsible for design, deployment, monitoring and remedy?
8. What record will allow an event to be reconstructed later?

These questions do not guarantee justice. They make evasion harder.

Neither rejection nor surrender

Artificial intelligence can improve scientific work, accessibility, infrastructure, education and administration. Rejecting it completely would surrender genuine capability. Deploying it everywhere because it is efficient would confuse capability with wisdom.

The more useful position is conditional adoption. Use powerful systems where their purpose is legitimate, evidence supports the claimed benefit, authority is limited and failure can be detected and contained. Refuse uses whose consequences cannot be inspected, challenged or safely reversed.

The same principle should govern human institutions. Power is sometimes necessary to coordinate infrastructure, enforce law, protect rights and respond to emergencies. The answer to abusive power is not the absence of all authority. It is authority constructed with boundaries, evidence, counterweights and responsibility.

Humanity's mirror does not show that people are incapable of governing intelligent machines. It shows why technical intelligence alone will never solve the problem of governance.

No intelligence—human, institutional or artificial—should possess more consequential power than society can inspect, challenge and safely withdraw.

The difficulty is not understanding the principle. It is accepting its application when the power serves our own interests.

---

Sources

Footnotes

1. National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023; NIST, AI Risk Management Framework status page, accessed 3 October 2026. ↩ ↩²

2. OECD, “OECD AI Principles”, adopted 2019 and updated 2024. ↩ ↩²

3. UNESCO, Recommendation on the Ethics of Artificial Intelligence, adopted 23 November 2021, especially paragraphs 14, 35–43. ↩ ↩² ↩³

4. Saar Alon-Barkat and Madalina Busuioc, “Human–AI Interactions in Public Sector Decision Making: ‘Automation Bias’ and ‘Selective Adherence’ to Algorithmic Advice”, Journal of Public Administration Research and Theory, 33(1), 2023, pp. 153–169. ↩

5. European Commission, “AI Act: Regulatory framework”, implementation status accessed 3 October 2026; official legislative text is linked from the Commission page. ↩

6. Council of Europe Treaty Office, CETS No. 225: signatures and ratifications, status checked 3 October 2026; Council of Europe, Framework Convention overview. ↩

7. OECD, Survey on Drivers of Trust in Public Institutions — 2024 Results, based on nationally representative surveys across 30 OECD countries conducted in October and November 2023. ↩