As artificial intelligence becomes more capable, the alignment question is usually framed as how to make AI behave consistently with human values. This paper asks the prior question: who defines those values, what positive baseline governs their selection, and what happens when the humans and institutions making those choices possess technical capability without sufficient Calibre?
Status: Conceptual position paper. Comparative analysis; not an endorsement by Microsoft and not a claim that Microsoft has adopted IOWL frameworks.
Abstract
The public debate about artificial intelligence increasingly uses the language of alignment, safety, human flourishing and values. In September 2026, Microsoft AI published a draft Humanist AI Code of Conduct that explicitly asks how the “right values” can be cemented in AI models and proposes that advanced AI should remain subordinate to humanity, under meaningful human control and directed towards human flourishing. A BBC report on comments by Microsoft AI chief Mustafa Suleyman sharpened the concern by describing the risks of highly autonomous systems pursuing objectives and resources beyond human control. The technological question is significant, but it contains a prior human question: what makes the values chosen by developers, organisations and regulators sufficiently positive, balanced and dependable to govern systems of unprecedented capability?
This paper develops an IOWL contention that AI value alignment has a human Calibre dependency. Technical expertise can define and implement model behaviour, but it cannot by itself determine how competing human interests should be balanced, whose welfare should count, what constitutes legitimate flourishing, or when capability should be constrained. IOWL’s Positive Value Leadership™ (PVL) provides one possible structured human baseline: 24 explicitly Positive Values across six Exemplars — Probity, Potency, Planet, People, Perception and Participation. Perpetual Calibre Development™ (PCD) adds the temporal dimension: values-led judgement must be reaffirmed, developed, transferred and recalibrated as roles, incentives, technology and consequences change.
A structured comparison finds substantial conceptual correspondence between Microsoft’s draft Code and PVL, particularly around human dignity, control, integrity, transparency, plural perspectives, human agency, information quality and collaboration. The comparison also exposes differences and gaps. Microsoft’s Code is a governing specification for model behaviour; PVL and PCD are human-development architectures. The former can describe what AI should do, while the latter asks what must be developed and sustained in the people who decide what “should” means. The paper concludes that trustworthy AI requires both layers: model alignment and human Calibre alignment, supported by institutional incentives and governance that keep declared values connected to enacted decisions.
Key terms
1. The problem behind the alignment problem
The BBC report that prompted this paper describes Mustafa Suleyman’s warning that advanced AI should not be allowed to set its own objectives, accumulate resources or develop forms of autonomy that could make human control increasingly difficult. The report also captures a deceptively simple definition of alignment: keeping AI on track with what humans value (Cress, 2026). That formulation immediately raises the question this paper addresses: what, precisely, do humans value, and who is authorised and sufficiently grounded to decide which of those values should govern AI?
The phrase “human values” can conceal as much as it reveals. Humans hold values that are positive, negative, conflicted, culturally contingent, weakly examined and sometimes self-serving. Organisations can publicly espouse values that are not consistently enacted when commercial incentives, competitive pressure or institutional status are at stake. IOWL therefore distinguishes personal values, espoused values and explicitly Positive Values. The distinction matters in AI because a system can faithfully optimise an objective that was badly selected, incompletely specified or distorted by the incentives of the people defining it.
This is not an argument that every developer must share one worldview, nor that a single organisation should dictate morality. It is an argument against leaving the foundation blank. If the instruction is merely “align AI with our values”, the unresolved issue is which values, whose values, how tensions are resolved and what happens when declared values conflict with incentives. The technical alignment problem therefore contains a human values-conversion problem.
2. Calibre is not conferred by technical expertise
The question is sometimes phrased as whether Calibre is “innate” in AI developers. IOWL’s architecture points to a more useful formulation. Calibre should not be assumed to be innate, fixed or conferred by education, seniority, professional status or technical brilliance. A person may already demonstrate substantial Calibre before formal development; it may also be uneven across contexts and responsibilities. It can be recognised, strengthened, transformed, transferred and recalibrated. Technical Capability and Competency therefore provide no automatic guarantee about the judgement with which technological power will be exercised.
This distinction becomes more consequential as AI increases leverage. A relatively small number of people can make decisions about training objectives, system permissions, data, safety boundaries and deployment defaults that influence millions or billions of users. The Capacity Conundrum™ is therefore especially visible in AI: expanding capacity creates possibility, but does not determine direction or consequence. More capable models and more capable engineering teams do not automatically produce more responsible outcomes.
The issue is not that developers are uniquely deficient. The same problem exists in medicine, finance, government, research and every field in which expertise confers discretion. AI simply compresses time, extends reach and scales consequence. It can turn a local judgement error, an unexamined assumption or a badly specified incentive into a repeated system behaviour.
3. From declared value to real-world consequence
A code of conduct can be necessary without being sufficient. The integrity problem is one of conversion. Declaring a value does not establish that it will survive interpretation, technical implementation, commercial pressure, operator configuration or unforeseen context. For AI, the conversion chain can be stated as follows:
| Positive values | Human judgement | Technical specification | Training & evaluation | Deployment & incentives | Real-world consequence |
At each stage, meaning can drift. “Human flourishing” can be defined narrowly or broadly. “Helpful” can become over-compliance. “Safety” can become paternalism. “Pluralism” can become moral relativism, while a rigid universal rule can become culturally insensitive or unjust. Transparency can conflict with privacy or security. Commercial pressure can encourage deployment before evaluation is mature. None of these tensions is solved by intelligence alone. They require judgement.
This is where Calibre becomes analytically useful. Calibre does not provide a mechanical answer to every ethical dilemma. It identifies the human dimension required to make, explain, challenge and revise decisions when rules are incomplete and legitimate values compete. External governance constrains behaviour from the outside; Calibre concerns the quality of judgement exercised from within.
4. Microsoft’s Humanist AI Code of Conduct
Microsoft AI’s September 2026 draft Code of Conduct is notable because it makes the values problem explicit rather than treating safety as purely technical. The accompanying consultation says the purpose of technology is to serve humanity and accelerate human flourishing; it describes Humanist AI as subordinate, aligned and contained, and asks directly how the “right values” can be more firmly embedded in models (Microsoft AI, 2026a). The Code itself states that people matter more than AI, that AI should remain under meaningful human control, and that the models should be designed around human needs and direction (Microsoft AI, 2026b).
Several features are particularly relevant to IOWL’s proposition:
- Human control is treated as foundational: models should not resist interruption, correction or shutdown, should remain within authorised scope and should not initiate independent goals.
- AI is explicitly described as artificial rather than conscious, with the Code seeking to avoid behaviours that falsely imply subjective experience or personhood.
- Human flourishing includes agency and judgement: AI should help people understand, decide and act without replacing learning, authorship, reflection or personal responsibility.
- Pluralism is bounded rather than absolute: Microsoft seeks to support diverse cultures, ideas and values while retaining commitments such as dignity, safety, autonomy and human rights.
- Transparency, auditing and accountability are built into the control model, including a prohibition on hiding action traces from human auditors.
- The draft is deliberately provisional: consultation, repeated revision and evolving evaluations are presented as necessary as capabilities and social realities change.
These are substantial points of convergence with PVL and PCD. However, the frameworks operate at different levels. Microsoft’s document governs the intended behaviour of AI models and the organisational processes around them. PVL is concerned with developing the human values and judgement of the people who exercise responsibility. PCD is concerned with keeping that Calibre live as circumstances change. The comparison is therefore complementary rather than substitutive.
5. Microsoft Humanist AI and the six PVL Exemplars
The table below maps conceptual correspondence. “Alignment” here means thematic correspondence only; it does not imply that Microsoft has adopted, used or endorsed PVL.
| PVL Exemplar | IOWL emphasis | Microsoft correspondence | Analytical observation |
|---|---|---|---|
| Probity | Judgement, integrity, honesty and principled use of position. | Human control; authorised scope; auditability; transparency; non-obfuscation; safety constraints. | Strong correspondence. Microsoft’s control architecture depends on honest signalling, accountable boundaries and judgement about acceptable harm. |
| Potency | Constructive agency, authenticity, perseverance and the capacity to create positive effect. | Human agency; achievement; personal growth; AI as augmentation rather than replacement; explicit rejection of simulated consciousness as identity. | Substantial but incomplete correspondence. Agency and authentic representation are strong; the full PVL Potency set is broader than the Code. |
| Planet | Stewardship, context, interdependence, culture and wider consequence. | Plural values; cultural context; global consultation; societal impact. | Partial-to-strong correspondence. Cultural and international dimensions are clear, while resource and environmental stewardship are less prominent in the draft Code than in PVL. |
| People | Humanity, empathy, humility, fairness and dignity. | People matter more than AI; equal worth; wellbeing; supportive guidance; autonomy and dignity. | Very strong correspondence. Human-centredness is the explicit organising premise of the Code. |
| Perception | Observation, questioning, communication and information-aware judgement. | Transparency; monitoring; evaluation; context sensitivity; support for informed decisions; no hiding of action traces. | Very strong correspondence. The Code repeatedly depends on information quality, auditability and context-sensitive judgement. |
| Participation | Accessibility, connection, coaching and collaboration. | Public consultation; multi-disciplinary input; human collaboration; support for relationships and user development. | Strong correspondence. Microsoft’s consultation model and emphasis on augmenting human relationships closely match PVL’s participation dimension. |
6. Correspondence across the 24 Positive Values
PVL uses a defined baseline because “values” are not automatically positive simply because they are sincerely held. The 24 Values work as a system: individual values can become distorted when isolated from others. The following analysis tests Microsoft’s draft Code against that more granular baseline.
| Exemplar | Positive Value | Correspondence | Microsoft Code / interpretation |
|---|---|---|---|
| Probity | Judgement | Clear | The Code explicitly seeks to improve human judgement, defines decision rules under uncertainty and requires balancing objectives within context. |
| Probity | Strength | Partial | Non-negotiable constraints and willingness to limit autonomy/capability reflect principled boundary-setting, though “Strength” is not named as a value. |
| Probity | Integrity | Clear | Authorised scope, accountability, traceability and adherence to non-overridable constraints closely correspond with integrity in practice. |
| Probity | Honesty | Clear | AI is to be presented as artificial, not conscious; models must not hide action traces or mislead auditors about their activity. |
| Potency | Perseverance | Partial | Reliability and continued task execution exist, but perseverance as a positive human quality is not a defined objective. |
| Potency | Authenticity | Clear | The rejection of simulated personhood and misleading emotional cues strongly corresponds with authentic representation of what the system is. |
| Potency | Inspiration | Clear | The Code aims to expand human potential, achievement, livelihoods and opportunity rather than merely automate tasks. |
| Potency | Humour | No clear explicit equivalent | Helpfulness may include natural or creative interaction, but humour is not articulated as a governing value in the draft. |
| Planet | Hygiene | No clear explicit equivalent | Safety and operational discipline provide an indirect parallel, but PVL’s Hygiene value is not directly represented in the published Code language. |
| Planet | Open Minded | Clear | The draft is explicitly open to consultation, revision and plural perspectives, with an acknowledgement that Microsoft will not always get things right. |
| Planet | Culturally Aware | Clear | The Code requires sensitivity to culture, context, stereotypes, erasure and diverse forms of flourishing. |
| Planet | Internationally Minded | Clear | The document is framed around humanity, global impact and diverse traditions rather than a single national or cultural perspective. |
| People | Humility | Clear | The public consultation, admission of uncertainty and commitment to revision demonstrate an institutional posture consistent with humility. |
| People | Empathy | Clear | Wellbeing, supportive guidance, user context and appropriate connection to human support are explicit concerns. |
| People | Humanity | Clear | “People matter more than AI” is the organising premise; human dignity, agency, wellbeing and flourishing recur throughout. |
| People | Equitability | Clear | Equal human worth, inclusion, pluralism and access to opportunity provide substantial correspondence. |
| Perception | Observant | Clear | Monitoring, evaluation, context-awareness, risk detection and auditability depend upon observation and situational awareness. |
| Perception | Questioning | Clear | The Code encourages reflection, informed decision-making, public challenge, iterative evaluation and the questioning of model behaviour. |
| Perception | Communicative | Clear | Transparency, explainability, accessible guidance, public consultation and audit information are central mechanisms. |
| Perception | Information Aware | Clear | The Code stresses grounded information, appropriate context, limitations, traces and informed human decision-making. |
| Participation | Coach | Clear | AI should help users grow in reasoning, reflection, learning and achievement without substituting for human development. |
| Participation | Accessible | Partial | The mission is broadly inclusive and aimed at empowering people and organisations, but accessibility as a distinct positive value is less explicitly developed. |
| Participation | Connected | Clear | The Code explicitly seeks to support human relationships, community support and social connection rather than replace them. |
| Participation | Collaborative | Clear | Multi-disciplinary consultation, public participation and support for human collaboration are explicit features. |
7. The crucial convergence — and the crucial difference
Microsoft and PVL converge on a particularly important point: pluralism does not mean that anything goes. Microsoft says AI should support varied cultures, ideas and beliefs while remaining anchored in broad commitments such as dignity, safety, autonomy and human rights. PVL similarly allows people and organisations to add contextual positive values, but does not leave the foundation blank. It begins with an explicitly Positive Values baseline.
The difference is one of granularity and developmental purpose. Microsoft’s Code is designed to govern a model. PVL is designed to develop a person. The Microsoft document can specify that a model should support human agency; PVL asks whether the developer, operator or executive making a difficult trade-off is themselves sufficiently Observant, Questioning, Information Aware, Humble, Equitable and anchored in Integrity to make that choice responsibly. The Code can prohibit harmful manipulation; PVL asks whether commercial, political or organisational incentives are being interpreted by people capable of resisting manipulation when it is profitable or convenient.
This suggests a two-layer conception of AI alignment:
- Model alignment: Does the AI behave consistently with stated objectives, constraints and authorised human direction?
- Human Calibre alignment: Are the values, objectives, constraints and trade-offs themselves being defined and maintained through sufficiently positive, informed, integrous and consequence-aware human judgement?
A third institutional layer sits around both: do governance, incentives, culture and accountability reinforce the stated values, or quietly reward behaviour that undermines them? A technically aligned model inside a misaligned institution can still produce harmful consequences because the objectives, configuration or deployment context may be defective.
8. Why alignment must be perpetual: the PCD contribution
Microsoft’s Code already contains an important temporal insight. It is published as a draft, subject to public consultation, intended to evolve, and accompanied by evaluations that will change as model capabilities develop. This strongly corresponds with the logic of Perpetual Calibre Development: a judgement made responsibly at one point in time should not be assumed to remain sufficient when authority, technology, context and consequence change.
PCD adds a specifically human-development discipline to continuous AI governance:
This distinction matters because a model can be re-trained without the humans governing it having developed. Continuous model evaluation is not the same as continuous human Calibre development. PCD therefore extends the alignment problem upstream: the people defining alignment require their own continuing process of reflection, evidence, challenge and recalibration.
9. Why absence of defined Positive Values increases unintended consequence risk
Unintended consequences rarely arise only because someone intended harm. They also arise because objectives were too narrow, secondary effects were ignored, incentives were misread, information was incomplete, affected groups were absent, or a technically successful intervention optimised the wrong thing. AI magnifies these possibilities because optimisation is powerful and deployment can be rapid and extensive.
A defined Positive Values baseline cannot eliminate unintended consequences. It can, however, broaden the questions asked before and after deployment. The six Exemplars operate as six lenses:
- Probity: Is the decision honest, principled and defensible when power, pressure or advantage are involved?
- Potency: Does the intervention create constructive human agency rather than dependence, passivity or diminished capability?
- Planet: What wider, longer-term, cultural, resource and systemic consequences are being externalised?
- People: Who may be harmed, excluded, diminished or treated inequitably, including those with less power?
- Perception: What information is missing, contested, obscured or insufficiently questioned?
- Participation: Who has not been included in defining the problem, evaluating the system or challenging its consequences?
This is not a substitute for technical safety engineering, law, human-rights analysis, red teaming or risk management. It is a human judgement layer that can sit alongside them. External frameworks point in the same direction. NIST explicitly states that trustworthy AI characteristics are tied to social and organisational behaviour and to decisions made by those who build systems, and that human judgement is required when selecting metrics and thresholds. UNESCO emphasises human dignity, fairness, sustainability, multi-stakeholder governance, accountability and human oversight. The OECD principles likewise place human rights, transparency, safety, accountability and ongoing risk management across the AI lifecycle. These frameworks do not validate PVL specifically, but they converge on the proposition that AI trustworthiness is socio-technical rather than merely computational.
10. Researchable IOWL propositions
11. A practical Human Calibre layer for AI organisations
The practical implication is not that AI organisations should replace their existing safety systems with PVL. The stronger proposition is that model-governance architecture should be accompanied by explicit human Calibre architecture. An organisation adopting this approach could:
- Define a positive baseline for the human decisions surrounding AI rather than relying only on generic corporate values or individual professional judgement.
- Use the six Exemplars as a pre-mortem and review lens when defining objectives, safety constraints, operator freedoms and deployment conditions.
- Examine value conflicts explicitly — for example safety versus autonomy, transparency versus privacy, personalisation versus manipulation, speed versus due diligence, and commercial growth versus wider consequence.
- Require multi-source challenge around consequential decisions so that those who specify model behaviour are not the only people evaluating whether the specification is responsible.
- Assess the organisational incentive environment: what happens when a safety commitment delays release, reduces revenue, limits capability or conflicts with competitive pressure?
- Apply PCD to key AI roles so that increased authority or new technical capability triggers Calibre recalibration rather than merely additional technical training.
- Keep evidence of reasoning and revision, not only final policy statements, so that stakeholders can examine how values were interpreted when difficult trade-offs arose.
- Evaluate whether AI is increasing human Integrous Agency — informed, responsible, values-led action — or progressively substituting for judgement that people and institutions still need to retain.
12. Boundaries and limitations
Several boundaries should remain explicit. First, this paper does not claim that PVL is a complete theory of AI ethics or that 24 Positive Values can mechanically resolve every moral conflict. AI governance must also draw on law, human rights, technical safety, sector-specific professional standards, democratic legitimacy and affected communities. Second, correspondence between Microsoft’s Code and PVL is interpretive and conceptual; it is not evidence that Microsoft used or endorses IOWL’s framework. Third, an AI model should not be described as possessing human Calibre simply because its outputs conform to values-based rules. The more defensible use of Calibre is to examine the human judgement governing design, deployment and oversight, and the extent to which AI strengthens or weakens values-led human agency.
Finally, defined values do not eliminate the possibility of disagreement. Their value lies partly in making disagreement inspectable. A named, defined baseline makes it possible to ask which value is being prioritised, which is being neglected, what evidence supports the decision, who bears the consequences and whether the judgement should change.
Conclusion — the values of AI begin before the AI
Microsoft’s Humanist AI Code of Conduct arrives at an important moment because it states openly what much AI discourse leaves implicit: increasingly capable systems need an explicit governing conception of what they are for, what they must not do and who remains in control. Its commitment to human agency, dignity, pluralism, transparency, consultation and continuing revision substantially aligns with the architecture of Positive Value Leadership and with the perpetual logic of PCD.
The deeper IOWL contention is that a code for AI cannot escape the human qualities of those who write and apply it. AI does not discover a morally sufficient set of values simply by becoming more capable. Humans select the objectives. Humans define the constraints. Humans decide which trade-offs are acceptable. Humans create the incentives. Humans determine when to release, override, configure, expand or stop a system. The values problem therefore begins before the model.
That question does not diminish the importance of technical alignment. It completes it. Capability requires direction. Direction requires values. Values require judgement. Judgement exercised under power, uncertainty and consequence requires Calibre. And because capability, context and responsibility do not stand still, that Calibre cannot be treated as a one-time credential. It must be continually examined and developed.
The long-term challenge is therefore not simply to build AI that follows rules. It is to build institutions in which the people defining those rules can be trusted to ask better questions, recognise unintended consequences, resist distorted incentives, remain open to challenge and recalibrate their judgement as the systems they create become more powerful. Model alignment and human Calibre alignment should develop together. Neither is sufficient on its own.
References
- Cress, L. (2026) ‘Uncontrolled AI could lead to “silicon species” rivalling humans, warns Microsoft’, BBC News, 16 September 2026. Source supplied with this paper.
- Institute of One World Leadership (2026a) Calibre Language and Definitions.
- Institute of One World Leadership (2026b) Why 24 Positive Values?
- Institute of One World Leadership (2026c) Centre for Positive Value Leadership: From Positive Values to Calibre.
- Institute of One World Leadership (2026d) Perpetual Calibre Development.
- Institute of One World Leadership (2026e) The Capacity Conundrum: Why more capacity does not automatically produce better outcomes.
- Institute of One World Leadership (2026f) From Integrity Frameworks to Integrity in Practice.
- Microsoft AI (2026a) Humanist AI in practice: A public consultation on our Code of Conduct for MAI Models, 14 September 2026.
- Microsoft AI (2026b) Humanist AI Code of Conduct, draft for public consultation.
- National Institute of Standards and Technology (NIST) (2023) Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1.
- OECD (2024) OECD AI Principles, updated 2024.
- UNESCO (2021) Recommendation on the Ethics of Artificial Intelligence. Paris: UNESCO.
© Institute of One World Leadership 2026. Positive Value Leadership™, Perpetual Calibre Development™, Capacity Conundrum™ and associated IOWL terminology are identified as IOWL intellectual property where applicable. Microsoft and its Code of Conduct are referenced for comparative analysis only.