FUURAA™ · Vision depth page

Six decisions that keep people in control of AI

Turn human-centred intent into six decisions that can be rejected, reviewed and corrected: know how AI is involved; choose or refuse; correct data and memory; intervene before consequence; challenge outcomes and reach an accountable human; and exit with residuals closed.

Enter the six decisions ↓See the minimum evidence record

Published21 August 2026Public statusLong-term vision · public research methodEvidence basisMethod synthesis across Singapore public-value policy and international ethics, risk and privacy frameworks

The evidence boundary of agency

Agency is not a button—and “human in the loop” does not prove it.

Human agency depends on whether people can change the system’s path at consequential moments: understand, refuse, correct, intervene, challenge and exit. An interface may display approval while hiding alternatives, allowing effects before review or losing responsibility inside an automation chain.

Applicability boundaryThis page is a public research method within FUURAA’s long-term vision. It does not state that any FUURAA product implements these controls, and is not a compliance claim, legal-rights conclusion, safety certification, audit opinion, medical advice or other professional advice.

Six rejectable decisions

Each names the real question, minimum evidence and point where inference must stop.

Know when and how AI is involved

Decision question: Can affected people see the AI system’s role, purpose, consequence, data use and accountable owner before those facts matter?
Minimum evidence: A user-facing notice in context; exact system boundary; AI and human roles; consequence class; data purpose; current accountable owner.
Stop the inference: A policy page elsewhere does not prove that notice reached the person, appeared at the right moment or was understandable.

Meaningfully choose or refuse

Decision question: Where the context permits, is there a usable alternative—and are time, cost, friction and consequences visible before a person chooses?
Minimum evidence: The alternative or non-AI path; refusal outcome; necessity rationale; comparative time and cost; coercion, dark-pattern and accessibility tests.
Stop the inference: A click or consent record is not meaningful choice when refusal is hidden, punitive, inaccessible or impracticable.

Correct data and memory

Decision question: Can a person identify what shaped the system, correct material errors and see whether the correction reached derived and downstream state?
Minimum evidence: Source and provenance; correction route; affected records and models; downstream propagation; derived-state handling; retention and deletion state.
Stop the inference: Editing one visible field does not prove that caches, embeddings, derived memory, agents or downstream systems were corrected.

Intervene before consequence

Decision question: Can a suitably authorised person pause, change or reject a consequential action while intervention can still alter the outcome?
Minimum evidence: The human checkpoint; actual decision authority; intervention timing; fail-closed behaviour; escalation path; tested pause and rejection outcomes.
Stop the inference: Review after an irreversible effect is accountability evidence, not proof of pre-effect human control.

Challenge an outcome and reach an accountable human

Decision question: Can an affected person understand the relevant basis, contest the outcome and reach someone authorised to investigate and provide remedy?
Minimum evidence: An explanation suited to the decision; appeal or recourse route; case owner; response timing; reviewed evidence; remedy and closure record.
Stop the inference: A chatbot loop, generic explanation or contact form without ownership is not operational recourse.

Exit, revoke and close

Decision question: Can a person end participation, revoke delegated authority and obtain evidence that pending and downstream effects were addressed?
Minimum evidence: Authority revocation; agent and tool stop state; export or portability where appropriate; retention and deletion disposition; downstream propagation; residuals and closure owner.
Stop the inference: Deleting an account or disabling an interface does not prove that copies, delegated authority, derived state or pending effects are closed.

Minimum human-agency evidence record

Turn interface promises into a record that can be checked across products, data and chains of responsibility.

01Exact system, release, interface and decision context
02Affected people and groups, including accessibility needs
03Purpose, consequence class and prohibited uses
04AI and human roles, authority and current owner
05Available choices, alternatives, refusal outcome and comparative friction
06Data, memory, provenance, correction and propagation state
07Intervention checkpoints, recourse path, remedy and response timing
08Exit, revocation, retention, residuals, closure and next review

Four principles across the path

Agency comes from a real ability to change the path—not notification after the fact.

Legibility

People need the right information in the context and at the time where it can change their action.

Real alternatives

Choice is measured by the usable path left after refusal—not by the presence of a button.

Consequential timing

Human involvement matters only when the person has authority and enough time to alter the result.

Accountable recourse

Explanation without an owner, investigation and possible remedy leaves agency incomplete.

FUURAA analysisAgency is not one feature. It is a system property distributed across interfaces, data, memory, authority, timing and institutional responsibility. If a person cannot understand, refuse, correct, intervene, challenge or exit, “human in the loop” can become ceremonial. Trustworthy AI should prove not merely that a person appeared, but that at the consequential moment they had choices, authority, time and recourse capable of changing the result.

Singapore and international primary sources

Public value, rights-based ethics, risk and privacy each supply one part—not an inflated conclusion.

Sources rechecked 21 August 2026. Each identifies publication timing, methodological role and non-transfer boundary.

Introduced in 2024 · page updated 1 April 2025Singapore Smart Nation 2.0

Directs digital development toward outcomes that benefit people and uphold shared social values, rather than treating deployment as the outcome.

BoundaryA national digital vision for Singapore, not an operational AI test, product endorsement or proof that a system preserves agency.

Open primary source ↗Adopted May 2019 · updated May 2024OECD AI Principles

Connects trustworthy AI with human rights, democratic values, transparency, robustness and accountability across the lifecycle.

BoundaryIntergovernmental principles guide policy and practice; they do not certify an implementation or exhaust jurisdiction-specific duties.

Open primary source ↗Adopted 23 November 2021UNESCO Recommendation on the Ethics of Artificial Intelligence

Places human dignity, rights, oversight, determination, inclusion and recourse around the social use of AI.

BoundaryA global normative recommendation, not a substitute for law, contextual impact assessment or participation by affected communities.

Open primary source ↗Published 26 January 2023NIST AI Risk Management Framework 1.0

Provides voluntary lifecycle functions for governing, mapping, measuring and managing AI risks in context.

BoundaryUse-case agnostic and non-prescriptive; adoption does not prove that choices, oversight or recourse are meaningful in a specific system.

Open primary source ↗Published 16 January 2020NIST Privacy Framework 1.0

Frames privacy risk around problems people may experience across data collection, use, transformation, retention, sharing and disposal.

BoundaryA voluntary, outcome-based risk tool; it is not legal advice, a universal consent model or evidence of compliance with any jurisdiction.

Open primary source ↗W3C Statement · 12 December 2024W3C Ethical Web Principles

States that the web should support dignity and personal agency, enhance individual control and remain usable through people’s chosen tools.

BoundaryA non-normative TAG finding for the web platform, not an AI standard, accessibility conformance test or implemented product safeguard.

Open primary source ↗

Continue into adjacent evidence layers

Connect human choice to the AI life journey, system governance and public evidence.

Return to Vision Open the AI life journey Open safety and governance Evaluate meaningful human oversight Enter AI Evidence Atlas