Responsible AI governance must confront misalignment between values and outcomes

AI ethics frameworks are failing to address a deeper moral gap because they focus too heavily on procedural rules such as fairness, accountability and transparency while neglecting sincerity, according to a new study published in AI & Society.

The study, titled “Sincerity as ethical alignment to reconstruct the moral foundation of AI ethics,” proposes a sincerity-based ethical framework for AI governance, arguing that responsible AI requires alignment between truth, intention, action and trust, not only compliance with technical or institutional checklists.

AI governance faces a moral gap beyond compliance

The study challenges the dominant structure of AI ethics, where many frameworks are built around fairness, accountability and transparency. These principles remain important, but the paper argues that they do not fully explain how ethical systems should revise themselves when inconsistencies emerge between stated values and real-world outcomes.

A key concern is that AI systems may be transparent and accountable in a procedural sense while still operating in ways that are manipulative, misaligned or ethically incomplete. A system can disclose information, document decisions and pass formal audits, yet still fail to reflect genuine coherence between its declared purpose, its operational logic and its social effects.

Ethics should not be limited to whether an AI system follows rules. It should also ask whether the human and institutional actors behind the system are willing to confront inconsistencies, explain choices, revise objectives and remain answerable to people affected by AI-mediated decisions.

Sincerity is defined as the alignment of four elements: truth, intention, action and trust. Truth refers to cognitive coherence and the quality of justification. Intention refers to whether the stated aims of a system are authentic and not misleading. Action concerns whether those aims are reflected in actual design and deployment. Trust refers to whether institutions remain transparent, revisable and open to public review.

AI systems are treated as socio-technical components that can support governance by recording decisions, surfacing uncertainty, documenting contestation and enabling traceable revision. Moral responsibility remains with designers, deployers, regulators and institutions. The ethical question is not whether AI can itself be sincere in a human sense. The question is whether AI governance can be structured so that human and institutional actors behave sincerely when they design, justify, deploy and revise AI systems.

The paper uses philosophical traditions to show why AI ethics needs more than procedural compliance: it needs institutions willing to correct themselves. It draws on ideas of integrity, relational responsibility, narrative self-understanding, deliberative communication and inquiry-based learning. The result is a framework that shifts AI ethics from static principle-setting to an ongoing process of reflection and correction.

Sincerity-based ethics links truth, intention, action and trust

The proposed sincerity-based ethical framework is organised around four domains: cognitive integrity, emotional authenticity, social reciprocity and institutional transparency. Each domain addresses a distinct part of the moral structure needed for trustworthy AI governance.

• Cognitive integrity concerns whether truth claims are connected to clear justification. In AI systems, this relates to explainability, uncertainty reporting and the ability to show why a model produced a certain output. It is not enough for a system to deliver decisions. It must also expose the limits of its knowledge and avoid overclaiming confidence.
• Emotional authenticity addresses the way AI systems communicate with humans. The study argues that this does not mean AI has emotions. Instead, it concerns whether a system’s signals, tone, confidence and empathy-like cues match its real capabilities and limits. A chatbot or decision-support tool that appears reassuring while hiding uncertainty may be ethically misaligned even if it follows formal disclosure rules.
• Social reciprocity focuses on the people affected by AI decisions. It asks whether they have meaningful ways to understand, contest and influence systems that shape their lives. This connects AI ethics to participation, appeal rights, redress mechanisms and stakeholder involvement. A system that affects hiring, education, healthcare or public services cannot be considered ethically coherent if affected people have no practical route to challenge or correct its outcomes.
• Institutional transparency concerns the ability of organisations to revise norms, policies and systems when problems arise. In this framework, transparency is not only about making information visible. It is about making institutional decision-making open to correction. That means recording why changes are made, documenting how evidence was reviewed and allowing stakeholders to examine the reasons behind revisions.

These four domains work together as a feedback network. Knowledge shapes intention, intention informs action, action affects trust and trust reshapes knowledge. Sincerity operates through this cycle by requiring institutions to respond when misalignment appears.

To support that process, the paper introduces mismatch signals as governance triggers. These signals are not presented as direct measurements of moral worth. Instead, they are designed to flag possible gaps between intention, policy, action and outcome. When such gaps persist, they should trigger review, justification and possible revision.

The study also introduces a well-being and empowerment index as an outcome-oriented monitoring construct. This is not meant to replace moral reasoning or rights-based safeguards. Rather, it gives institutions a way to monitor whether AI systems are affecting people’s well-being and agency in ways that require further ethical review.

The structure reframes fairness, accountability and transparency. The study does not reject those principles. It argues that sincerity strengthens them. Fairness becomes more meaningful when people have contestability and participation. Accountability becomes stronger when explanations are faithful, uncertainty is acknowledged and systems can abstain under epistemic risk. Transparency becomes deeper when institutions disclose not only information but also the reasons and records behind revisions.

Reflective governance could redefine ethical AI

The study points out that ethical AI should not be treated as a fixed checklist completed before deployment. Instead, it should be understood as a learning process in which institutions identify moral dissonance, review evidence, respond to affected people and revise systems over time.

The approach draws on the idea of inquiry, where knowledge evolves through observation, testing and correction. Applied to AI governance, the paper argues that institutions should not only ask whether a system performs efficiently. They should ask whether its goals remain justified, whether its actions reflect declared values and whether its outcomes sustain trust.

The study outlines how this could work in practical settings. During design, organisations would make goals and value commitments explicit. During deployment, they would monitor complaints, exceptions, uncertainty, user harms and signs of mismatch. During audit and review, they would examine whether the system’s effects align with stated purposes. During institutional learning, they would revise policies, models and governance rules with recorded reasons.

In education, for example, an AI system recommending learning pathways may meet formal performance targets but still reduce student agency or increase disengagement. Under a sincerity-based model, those patterns would not be treated simply as performance issues. They would trigger review of the system’s underlying objectives, recommendation logic and effects on learners’ empowerment.

In healthcare, explainability would need to include more than technical clarity. Systems would need to communicate uncertainty responsibly and ensure that human decision-makers remain accountable for final judgments. In public governance, sincerity would require institutions to justify algorithmic choices in ways affected communities can understand and challenge.

The paper also introduces the idea of Society 6.0 as a heuristic label for a future form of ethical and reflective socio-technical governance. It builds on the concept of a human-centered cyber-physical society but adds a stronger moral dimension. The proposed vision is not presented as a fixed historical stage. It is used to describe a society in which technology and governance are designed for self-correction, public reasoning and collective well-being.

The study’s argument has direct relevance for current AI policy debates. Governments, companies and public institutions are increasingly adopting AI ethics principles, audit systems and governance frameworks. But many of these tools risk becoming compliance exercises if they do not require serious reflection on why systems are built, whom they affect and how institutions respond when harm or inconsistency appears.

The framework also raises hard implementation questions. Institutions would need to define who sets mismatch thresholds, who reviews evidence, who has authority to approve changes and how affected stakeholders can contest decisions. Without safeguards, the same organisations that deploy AI systems could control the review process in ways that protect themselves rather than the public.

The study acknowledges some limitations including that its proposed mismatch signals and update rules are conceptual tools, not validated metrics. The well-being and empowerment index needs domain-specific testing. Future research will need to examine whether sincerity-based governance improves accountability, stakeholder understanding and real-world outcomes over time.