In development · Working draft v0.5 · 4 October 2026. I am developing this framework in the open. It is a proposal, not yet fully verified, piloted or validated — and it will change. Please do not use it to certify, approve or reject an AI system.
Responsible AI & Human Agency · Framework · In development
The Responsible AI & Human Agency Framework
From ten studies to one assessment method: how to establish whether an AI system is technically responsible while preserving human autonomy, capability, dignity and accountability.
What this note does
The ten studies in this series each answer one question. This note turns them into a single assessment method. It fixes six things before any workbook, questionnaire or pilot is built:
- What is assessed — the unit of assessment.
- The structure — an entry gate, three layers and a loop.
- The decision rules — red lines, weakest link and five outcomes, never a single score.
- The evidence standard — what counts as proof, and how strong it must be.
- Who assesses — independence, participation and conflicts of interest.
- What already exists — a map against law and standards, showing where this framework adds something and where it must not duplicate.
The framework assesses two things at the same time: the quality of the technology, and the quality of the human situation it creates — for the people who use it, the people affected by it and the organisation deploying it.
1 · The unit of assessment is a use, not a model
The same model can be proportionate in one setting and unacceptable in another. So the framework never assesses “model X”. It assesses an AI use, defined by six elements:
- Purpose
- What the use is for, and what it is not for.
- System
- Model, version, configuration, data, retrieval sources, tools and supplier.
- Workflow
- Where AI sits in the work, and where authority sits.
- People
- Users, people affected, and people who carry the work or the risk.
- Organisation
- Owner, incentives, oversight and remedy routes.
- Context
- Sector, legal regime, stakes and reversibility.
A change to any element — a new model version, a new population, a new target — is a reason to reassess. That is what links the method to continuous assurance.
2 · Architecture: a gate, three layers and a loop
The gate comes first because the layers can only tell you how to use AI responsibly. Only the gate can tell you whether to. A use that fails the gate is not assessed further — the outcome is do not use or human-primary.
Human rights are not a fourth layer. Privacy, non-discrimination, autonomy and remedy each run through the technical, human and organisational layers, so they are assessed as a cross-cutting lens using the Rights-to-Evidence method.
3 · Two entry-gate tests and eighteen assessment domains
Two tests come first, at the entry gate. Then each of the three layers has six assessment domains. Each has one core question and at least one measurable indicator. The study column shows where the reasoning and evidence for that domain are set out.
| Domain | Core question | Example indicator | Study |
|---|---|---|---|
| Entry-gate tests | |||
| Proportionality | Is AI necessary, suitable and the least intrusive option for a legitimate purpose? | Documented comparison against the best realistic non-AI alternative | 01 |
| Purpose integrity | Does the activity keep the human good it exists to provide? | Affected people's own success criteria, recorded before deployment | 01 |
| Layer 1 · Technical responsibility | |||
| Validity & accuracy | Does it work for this task, population and setting? | Local validation against outcome; error rate with confidence interval | 09 |
| Robustness & reliability | Does performance hold under change and stress? | Performance under drift, adversarial and edge-case tests | 10 |
| Fairness & subgroup performance | Who does it serve worse? | Error rates disaggregated by relevant groups and languages | 03 |
| Uncertainty & abstention | Can it signal when not to rely on it? | Calibration curve; tested abstention rate | 09 |
| Security & privacy | Are access, data and actions proportionate and protected? | Least-privilege review; red-team findings closed | 10 |
| Traceability | Can an output and an action be reconstructed? | Version register; complete logs for consequential outputs | 06 |
| Layer 2 · Human responsibility | |||
| Retained capability | Can people still do the work, and explain, challenge and correct outputs, without AI? | Unaided performance before and after sustained use | 04 |
| Appropriate reliance | Do people accept correct outputs and reject wrong ones? | Correct and harmful acceptance and rejection, measured on seeded cases | 09 |
| Meaningful oversight | Do reviewers have the information, time, competence and authority to disagree? | Override rate and quality; review time per case against workload | 05 |
| Autonomy & choice | Can people refuse, opt out or reach a human without penalty? | Availability and use of a non-AI route; time to reach a human | 05 |
| Dignity & relational integrity | Does the design misrepresent competence, agency or care? | User testing of what people believe the system is and does | 09 |
| Human impact & distribution | Where does saved time go, and who carries the cost? | Workload, job quality and the share of gains reaching workers | 07 |
| Layer 3 · Organisational responsibility | |||
| Accountability | Is there a named owner who can answer, decide and stop? | Owner with documented stop authority | 06 |
| Incentives & culture | Do targets reward responsible outcomes, or only speed and adoption? | Responsible-outcome measures in the same scorecard as adoption | 08 |
| Participation & inclusion | Did affected people have influence before the decision? | Changes made because of affected-people input | 03 |
| Legal compliance | Are legal duties met — as a floor, not the finish line? | Legal register; impact assessments completed and current | 02 |
| Independent challenge | Can someone without a stake in the deployment question it? | Independent review with evidence access and escalation rights | 10 |
| Incident response & remedy | Can harm be reported, corrected and compensated? | Appeal success rate; time to remedy; near-miss reporting volume | 06 |
How the human layer is evidenced. Claims about people need more than surveys. Measuring Human Agency in AI Systems sets out the method in development: from perception to demonstrated behaviour and outcome, with an instrument library and fourteen indicators for piloting.
Rights-respecting measurement. The method used to measure a human right must itself respect human rights. Measuring people without surveilling them. Human indicators should be collected at team or cohort level, with consent, minimal data and a clear purpose — never turned into individual performance telemetry. A framework that protects human agency cannot be implemented through covert monitoring of the people it is meant to protect.
4 · Decision rules: no single score
Each domain receives one of four findings, together with its evidence level:
Findings are combined by five rules:
- Red lines block. A prohibited practice, an illegitimate purpose, or a severe and irreversible rights or safety harm without effective remedy means do not use — whatever the other findings say.
- Weakest link, not average. Strong performance in one domain cannot offset a red flag in another. There is no total score to average out.
- Uncertainty raises the bar. The higher the stakes, vulnerability and irreversibility, the higher the evidence level required — and the more “not evidenced” counts against use.
- Layer 2 cannot be waived. A use with strong technical and organisational findings but unassessed human domains is not ready for scale.
- Every outcome is time-limited. Each decision names its conditions, its owner and its next review date.
The rules produce one of the five outcomes from study 01: do not use · human-primary · restricted pilot · use with strong safeguards · proportionate use. Reassessment can move a use in either direction.
5 · Evidence classes, not a ladder
The framework reuses the evidence classes from continuous assurance, so assessment and monitoring speak the same language:
- E0 · Assertion — what is claimed: a policy, vendor or team says it is true.
- E1 · Controlled evidence — what happens under defined test conditions.
- E2 · Production evidence — what happens in actual use.
- E3 · Independent challenge — does the claim survive scrutiny from outside the deploying team?
- E4 · Human corroboration — does the protection hold from the perspective and experience of the people affected?
These are complementary classes, not a ladder. E4 is not automatically stronger than E3. To ask whether a model leaks training data, an adversarial test is far stronger evidence than interviews. To ask whether affected people can contest a decision, no technical audit can replace evidence from those people. Evidence is triangulated according to the claim being assessed. For significant rights impacts, a positive finding normally needs a legal or normative basis, controlled evidence, real-world evidence, distributional analysis, and independent or affected-person evidence in proportion to the stakes.
Proposed minimums: E1 for any pilot; E2 for proportionate use; E3 for use with strong safeguards in high-stakes settings; and E4 for the human-layer domains of any use that affects people's rights, livelihoods or access to services. E0 alone never supports a positive finding.
An evidence class describes the kind of evidence — not legal weight, maturity or compliance. A binding legal duty can be met with E0 evidence, such as a named role; a voluntary recommendation can call for E4 evidence from affected people.
The method also distinguishes four kinds of claim and never mixes them: legal requirement, voluntary standard, empirical finding and my proposed principle.
6 · Who assesses
| Role | Responsibility |
|---|---|
| Use owner | Provides evidence and is accountable for the decision — but does not sign off alone. |
| Independent assessor | Tests the evidence, sets findings, has stop and escalation rights, and has no stake in the deployment's success. |
| Users and frontline professionals | Supply human-layer evidence: workload, reliance, capability and whether override is real. |
| Affected people or their representatives | Shape success criteria and purpose integrity before the decision, not only after deployment. |
| Worker representatives | Review workplace effects, surveillance risk and the distribution of gains. |
| Legal, privacy and security | Confirm the legal floor and the red lines. |
Every assessment records conflicts of interest — including commercial relationships with the AI supplier — and how they were managed.
7 · How this framework relates to existing law and standards
The framework is designed to complement, not replace, the EU AI Act, the GDPR, the NIST AI RMF, ISO/IEC AI standards, the OECD AI Principles and the UNESCO Recommendation. I mapped all eighteen domains against them, provision by provision.
Existing instruments provide substantial coverage of technical risk, governance processes, fundamental rights and selected human impacts. Coverage is less systematic and less operational for retained human capability, appropriate reliance, relational integrity, practical autonomy, the distribution of gains and burdens, organisational incentives, and the degree of influence affected people have over AI-use decisions.
Coverage is not the same as legal applicability: many AI Act duties apply only to high-risk systems, and the GDPR only to personal-data processing.
See how the human layer is measured →
See how human rights are evidenced →
Apply the framework
The framework is designed to be used in real deployment decisions — not only read. Two practitioner instruments turn it into structured questions, evidence requirements, findings and decision records: an Organisational Questionnaire that gathers claims and evidence, and a Practitioner Workbook in which an assessor tests them.
Download the framework overview — PDF, 14 pages, working draft v0.5 →
8 · Validation: three contrasting pilots
| Setting | Why it tests the framework | Domains under most pressure |
|---|---|---|
| Enterprise AI assistant | Low stakes per task, high scale, strong adoption incentives | Retained capability, human impact & distribution, incentives |
| AI-assisted recruitment | High stakes for people who never chose the system | Fairness, participation, meaningful oversight, remedy |
| Customer-facing conversational AI | Relational design, vulnerable users, contested outcomes | Relational integrity, appropriate reliance, autonomy & choice |
The pilot protocol sets out ten phases, a minimum evidence pack, ground rules and the criteria the framework must meet. Each pilot should report what the framework found that existing assessments did not, what it missed, how long it took, and whether any finding changed a decision. A framework that never changes a decision has not been validated — it has been rehearsed.
9 · What comes next
- Check the crosswalk clause by clause against each instrument.
- Refine the indicators — every domain now has at least one; pilots will test which are practical and which change decisions.
- Build the practitioner workbook and an organisational questionnaire.
- Run the three pilots and publish what changed.
- Publish version 1.0 as an open method, with a public log of changes.
Development path
- v0.1 · Working draft, published in developmentArchitecture and decision rules proposed.
- v0.2 · Clause-level standards mappingAI Act, NIST, OECD and UNESCO citations checked against primary sources; ISO clauses and a second reviewer still pending.
- v0.3 · Measurement methodsHuman-agency measurement and the Rights-to-Evidence method.
- v0.4 · Indicators and pilot protocolEvery domain has at least one indicator; the pilot protocol is ready for a first partner.
- v0.5 · Practitioner instruments designedYou are here. Questionnaire and workbook previewed online as a research prototype; not yet released for use.
- Substantive reviewDomains, indicators, thresholds and decision rules reviewed and revised.
- Three pilotsEnterprise AI assistant, AI-assisted recruitment, customer-facing conversational AI.
- Documented revisionsWhat the pilots found, what they missed, what changed in the framework.
- v0.9 · Public betaA verified and piloted method, open to challenge.
- v1.0 · Validated methodOnly once pilots show the framework can change real decisions.
Version history
- v0.5 · 4 October 2026
- Practitioner instruments designed: an Organisational Questionnaire (a 32-question rapid screen and nine stakeholder tracks) and a Practitioner Workbook (charter, entry gate, domain modules, registers, decision record and continuous assurance plan). Previewed online only; release path to v0.9 and v1.0 defined.
- v0.4 · 4 October 2026
- Indicators added for the two entry-gate tests and the technical and organisational layers, so every domain now has at least one. Pilot protocol published: ten phases, minimum evidence pack, report questions and pass criteria.
- v0.3 · 4 October 2026
- Human rights & evidence added as a cross-cutting lens with the Rights-to-Evidence method, six evidence perspectives and the rights-in-practice path. Evidence levels renamed evidence classes: complementary and triangulated, not a ladder. Rights-respecting measurement made an explicit principle.
- v0.2 · 4 October 2026 (later)
- Human-agency measurement method added: evidence chain, six candidate dimensions, instrument library and fourteen pilot indicators.
- v0.2 · 4 October 2026
- Clause-level standards mapping added. Proportionality and purpose integrity named as entry-gate tests, separate from the eighteen domains. Coverage claim narrowed. Evidence levels clarified as evidence type, not legal weight. Unverified ISO clause numbers removed; ISO/IEC 42005 and 42006 added. Participation mapping corrected: the AI Act identifies affected groups but does not require their involvement.
- v0.1 · 4 October 2026
- First working draft, published in development.
Method, limits and independence
Independent and vendor-neutral. The framework is not affiliated with, commissioned by or endorsed by any employer, supplier, institution or standards body — including those referenced in the crosswalk.
Methodology disclosure. This framework is my own synthesis of the ten studies in this series, which draw on legislation, standards, regulatory guidance and empirical research. Referencing an instrument does not imply that its authors agree with this method. The architecture, domains, indicators, decision rules, evidence minimums and roles are proposals and have not been validated. The crosswalk is indicative.
Version status. All v0.x versions are developmental. They are not validated assessment standards and should not be used to certify, approve or reject an AI system.