Making AI Ethics Executable

1.The gap is between commitment and consequence

Consider an AI assistant that produces a fluent, accurate summary. The person receiving it cannot tell which statements are evidence, which are interpretation, or who can authorise the recommended action. The output may be useful, but the relationship between information and accountable action is incomplete.

CCD asks a question at that boundary: what does the complete system make possible – or impossible – for the people affected by it?

This question belongs within an existing field. Value Sensitive Design integrates conceptual, empirical and technical inquiry into human values, including the interests of direct and indirect stakeholders.^[1] NIST's AI Risk Management Framework 1.0 explicitly treats AI as socio-technical and addresses risks arising from its organisational and social context.^[2] The OECD AI Principles include human agency, oversight and accountability.^[3]

CCD does not claim to have discovered these concerns. It proposes a particular way to connect the participant's experience of a system with the organisation's authority to deploy and change it.

Here, executable means translated into explicit requirements, permissions, interaction behaviours, accountable decisions and testable controls. It does not mean that ethical judgement can be fully automated, that instructions alone enforce rights, or that the complete CCD architecture already exists as production software.

2.The person is not merely an input

In this paper, consciousness refers to lived human attention, perception, interpretation, meaning and choice. It is not a claim about machine sentience or a comprehensive scientific theory of consciousness.

CCD uses participant deliberately. The term brings consent, contribution, correction, refusal and reciprocal obligation into the design brief. This is a design commitment, not a claim that the word user is inherently unethical or that changing vocabulary changes power.

Participation does not imply voluntary involvement. Employees, customers and citizens may have little choice; people affected without using the system also count. An opt-out carrying an unreasonable penalty is not meaningful agency. Where participation cannot be optional, necessity, limits, accountability and challenge routes must be explicit.

Selected CCD commitments guide the analysis. Divergent Kind publishes their working form as four rules the work runs on.

The first commitment requires meaningful control over participation, attention and representations made about a person. A useful interpretation must not become an unchallengeable institutional fact.

The second connects declared purpose, actual incentives and observed behaviour. Alignment with a harmful purpose is not ethical alignment; purpose itself remains subject to rights and legitimate challenge.

The third asks what a system returns for the data, effort, attention and responsibility it requests. The ambition exceeds avoiding harm: participation should strengthen capability rather than manufacture dependency. Gains for one party must not conceal burdens displaced onto others, future participants or the environment.

The third commitment also asks whether the system expands what people can understand, choose, create and contribute, including through support they choose to use. Adaptation runs both ways. People should have the opportunity to strengthen judgement and take on new capability. Organisations should remove burdens the work does not need and allow different forms of participation. This is not a requirement to produce ever more, or to work without support. Entitlement does not depend on demonstrated improvement. The question is whether the system makes more possible for the people it affects, without conformity or dependency becoming a condition of participation.

The fourth requires relevant distinctions without reducing a person to a label. Uncertainty should remain visible. A request for clearer pacing is not permission to infer a diagnosis or restrict opportunity.

These commitments do not resolve every conflict. Trade-offs require affected-party input, a documented decision, accountable authority and review – not a claim that values harmonise automatically.

3.One commitment, six deployment decisions

The following example is hypothetical, not a client case or a tested intervention.

An organisation proposes an assistant that summarises customer escalations, identifies unresolved work and recommends next steps. The central commitment is that people should retain meaningful agency over information and decisions affecting their work.

First, establish purpose before access. The assistant supports coordination; it does not evaluate employee performance. Designers identify who benefits, who contributes data and who could be affected indirectly. They test whether the problem needs AI at all. A simpler process repair remains a valid outcome.

Second, restrict what it can read. The deployment names permitted sources, excluded material, retention arrangements and who can inspect the record. Private messages are not silently repurposed. Permission to access a channel does not itself justify every inference that could be made from it. Optional reflection must remain separate from required operating evidence.

Third, adapt presentation without changing entitlement. A participant may request a diagram, a shorter sequence or additional context. The system should offer these forms without inferring a fixed cognitive identity. Material evidence and uncertainty remain available to every authorised participant; simplification must not conceal inconvenient information. People can change the presentation or return to the source.

This commitment is expressed in Divergent Kind's proposed adaptive-presentation work as adapting form, not truth or rights. But form is not neutral: ordering and emphasis can influence interpretation. It therefore requires testing and contestability, not an assumption that presentation changes are harmless. Existing human–AI interaction guidelines already recommend dismissal, correction and cautious adaptation.^[4]

Fourth, make authority operational. Recommendations identify who can accept, reject or escalate them. Reviewers need time, relevant evidence and actual discretion; a nominal approval button is insufficient. Low-risk delegated actions may be bounded explicitly, but consequential commitments require the appropriate human authorisation. There must be a tested suspension and handback route.

Fifth, preserve correction and dissent. A worker can challenge an incorrect attribution or explain why an apparently stalled action is appropriately paused. The system records the correction and its disposition. It does not classify disagreement as resistance to adoption. Customers and other affected parties need appropriate routes to question consequential outcomes too.

Sixth, test who received the benefit. Evaluation examines whether coordination improved after including verification, correction and review effort. It asks whether reduced work for managers created extra work elsewhere, whether people retained the ability to work without the assistant, and whether they became better at recognising its limits. Faster summaries alone do not settle those questions.

This is ethical traceability: a commitment becomes a requirement, a configuration choice, an accountable practice and something observable. The interface and operating constitution are two connected sites of implementation, not separate ethics exercises.

4.What we read, what people experience, what we observe

A framework becomes harder to scrutinise when dimensions, outcomes and metrics use interchangeable names. This paper separates them.

Divergent Kind's public Coherence Lens supplies four analytical dimensions. As this paper reads them, Agency concerns the practical ability to understand, choose, act and contest. Reciprocity concerns the distribution of benefits, burdens, response and responsibility. Alignment concerns consistency between legitimate purpose, incentives and behaviour. Signal Integrity concerns whether relevant information, provenance and uncertainty remain usable through interpretation and action.

These are questions to examine, not a validated universal scale. They can inform inquiry at different organisational levels, but conceptual reuse does not establish that one instrument or threshold works at every level.

Participant coherence describes conditions at the receiving end of the interaction:

• Signal Landing: relevant information and material uncertainty are accessible in a usable form.
• Meaning Formation: the participant can understand the claim, its basis and its limits, and articulate a different interpretation. Agreement is not required.
• Action Coherence: the participant can take an informed action, seek clarification, defer or refuse through a legitimate route.
• Trust Calibration: confidence and actual reliance are proportionate to demonstrated reliability, uncertainty, context and consequences. They may appropriately increase or decrease.

The Lens describes what is examined; the participant conditions describe what that examination looks for in an experience. Their relationship is many-to-many, not four dimensions mapped mechanically to four conditions.

Observable indicators come afterwards. Depending on the question, these might include accurate recognition of uncertainty, successful correction of an attribution, appropriate refusal, time spent checking outputs, or reliance on reliable versus unreliable recommendations. Such indicators require defined methods; they are not interchangeable with the wider Qualitative Coherence Indicator (QCI) measurement family.

The earlier Participant Coherence essay called its fourth condition Trust Continuity. This paper proposes Trust Calibration instead. Appropriate reliance, rather than maximum trust, is an established concern in automation research.^[5] An explanation that helps someone trust an unreliable system less may be an ethical success.

The revision also clarifies Meaning Formation. Understanding an institution's claim must not be confused with accepting its preferred meaning.

5.Coherence must preserve disagreement

A system can be internally consistent while serving an unjust purpose. A team can appear aligned because dissent has become costly. An interface can feel reassuring while making its limitations harder to recognise.

Coherence is therefore not sufficient evidence of ethical quality. Within CCD, it must be constrained by agency, reciprocity, rights and contestability.

Coherence is not conformity. A legitimate objection must not count as a defect merely because it interrupts a workflow. Nor should participation require performing calmness, positivity or a designer's preferred emotional state. Support can be offered; it must not become a condition for being heard.

Similarly, a high average cannot excuse a serious harm to a smaller group. Evaluation should preserve distributional differences and adverse findings rather than allowing an aggregate result to erase them.

6.What would test the proposition?

The test is whether CCD-informed requirements improve observable outcomes compared with a credible alternative – not whether participants find the vocabulary appealing.

A useful comparison would hold the model and task constant while varying interaction and authority configuration. The comparison must include established responsible-design practices, not a deliberately weak baseline. Primary outcomes, foreseeable harms and stopping rules should be specified in advance.

Comprehension can be tested through application of information, not just satisfaction ratings. Agency can be examined through whether correction and refusal routes work. Reciprocity requires accounting for whose labour increased or decreased. Trust calibration requires observing reliance when a system is reliable and when it is not – not simply asking whether people trust it more.

Participant accounts should inform interpretation, alongside observed behaviour and operating records. Repeated observation matters because an initially helpful adaptation may later create dependency or additional work. Evaluation should include people for whom the configuration works poorly and assess whether benefits survive outside the original setting.

Evidence against the proposition would include no meaningful improvement over the comparison, increased inappropriate reliance, reduced visibility of uncertainty, inaccessible challenge routes, or benefits achieved by moving unacknowledged burdens elsewhere. Those findings should lead to revision or rejection of the relevant design, not a redefinition of coherence that protects the framework.

Construct definitions, hypotheses, failure criteria and evaluation logic belong in public. Independent evaluation additionally needs sufficient methodological access to scrutinise specific claims. Protected implementation detail cannot substitute for that access; where it is unavailable, the claim must remain correspondingly limited.

7.Governance continues after launch

Runtime ethics means that commitments remain inspectable when models, permissions, workflows or organisational incentives change. The deployment needs change records, exception handling, accountable review and a way to reverse or suspend a harmful configuration. They support ethical judgement, not replace it.

The provider's commercial structure also deserves scrutiny. Finding a problem should not automatically oblige the client to buy remediation. Assessment, implementation and review should be separately commissioned where appropriate. Capability transfer and a workable exit should be explicit objectives.

Producer–reviewer separation matters when others will rely on an assurance claim. A different agent or report title does not alone establish independence. The reviewer needs appropriate evidence access, predetermined criteria, protection from conflicts and the ability to reach an adverse conclusion. These are requirements for the proposed review arrangement, not a claim that every Divergent Kind engagement already provides independent assurance.

Evidence status and disclosure

This paper presents a normative argument, a hypothetical design example and a proposed evaluation approach. It reports no new experiment and makes no client-specific empirical claims. The wider CCD programme includes applied organisational work, but that does not validate every construct, application or proposed technical component. A specification is not a demonstrated runtime; neither an intellectual-property filing nor agreement among AI models is empirical validation.

Richard Lipp founded Divergent Kind, which develops commercial services and intellectual property related to the framework. That interest should be considered when evaluating the claims. The work was developed through Richard-led design practice and iterative AI-assisted research, analysis and drafting. The author is responsible for factual verification, the final argument and the released text. References establish relevant prior work, not endorsement of CCD.

The commitment that must survive

The practical test is whether a participant can recognise the ethical commitment in what the system permits, refuses, explains and repairs – and whether an accountable organisation can demonstrate how it upheld that commitment.

A person who questions a recommendation, withdraws unnecessary data or chooses not to automate may be exercising precisely the agency the design should protect.

Making AI ethics executable is not making people more governable by machines. It is making the design and deployment of machines more answerable to people.

References

1. Friedman, B., & Hendry, D. G. (2019). Value Sensitive Design: Shaping Technology with Moral Imagination. MIT Press. See also Value Sensitive Design Lab, About Value Sensitive Design: overview of the approach, its tripartite methodology, stakeholders and value tensions. vsdesign.org/vsd
2. National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. doi:10.6028/NIST.AI.100-1. Executive summary: airc.nist.gov/airmf-resources/airmf/0-ai-rmf-1-0 (The cited text is the January 2023 framework, which NIST describes as a living document subject to periodic review.)
3. Organisation for Economic Co-operation and Development. (2019; updated 2024). Recommendation of the Council on Artificial Intelligence. OECD/LEGAL/0449. legalinstruments.oecd.org/en/instruments/oecd-legal-0449
4. Amershi, S., et al. (2019). Guidelines for Human-AI Interaction. Proceedings of CHI 2019. doi:10.1145/3290605.3300233. Research overview and guidelines: microsoft.com/en-us/research/project/guidelines-for-human-ai-interaction
5. Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80. doi:10.1518/hfes.46.1.50_30392. pubmed.ncbi.nlm.nih.gov/15151155

Related essay

Lipp, R. (2026, 1 June; revised 5 September 2026). Participant Coherence: The Missing AI Product Metric. Divergent Kind. divergentkind.com.au/insights-article. The revision of its meaning, action and trust definitions is dated on that page.