Digital Narrative Care

The first Global Dialogue on AI Governance would be successful if it moved the international conversation beyond high-level principle and into the practical conditions under which AI systems shape human consequence. One essential outcome would be recognition that AI governance must protect not only systems, models and outputs, but also the institutional accounts those systems create: records, summaries, classifications, risk flags, case notes and decisions that travel through public and private institutions. These "travelling accounts" increasingly influence care, liberty, eligibility, education, employment, financial access, migration, social protection and redress. A successful Dialogue should therefore establish that transparency, accountability and human oversight are not abstract commitments. They require practical mechanisms: provenance, contestability, correction, review, explanation, escalation and aftercare. A person affected by an AI-shaped record should be able to know that the record exists, understand what it says, see where it has travelled, challenge what is wrong or misleading, and have corrections propagate to the places where the earlier account may already have been relied upon. The Dialogue should also recognize that the ethical question is not only whether AI systems treat people fairly at the point of output, but whether the accounts they create allow people to remain recognizable, contestable and correctable as they move through institutions. The strongest outcome would be a shared governance direction: AI must not only be safe at the point of deployment. It must remain accountable as its outputs become institutional memory.

• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Protection and promotion of human rights
• Transparency, accountability, and human oversight

These priorities belong together because the human consequence of AI increasingly appears through institutional records. An AI system may draft a clinical note, summarize a social care visit, classify a complaint, rank a risk, translate a consultation response, support a policing file, or help assess a student or worker. The immediate output may look technical, administrative or routine. But once that output enters an institutional pathway, it may shape how a person is understood, treated, believed, supported, sanctioned or denied. This is why "safe and trustworthy AI" cannot be limited to model performance. Trustworthiness must include the integrity of the account that travels. Was the AI origin visible? Was the account checked before use? Was uncertainty preserved? Was context flattened? Did the output distinguish fact from inference? Could the person affected understand and challenge it? If corrected, did the correction travel to downstream users? The social, ethical, cultural and linguistic implications are equally central. AI systems often compress human complexity into forms institutions can process. Accent, idiom, disability, trauma, distress, hesitation, cultural context and contested meaning can be thinned or misread. The harm may not be spectacular; it may be a plausible but unfaithful account becoming the version institutions inherit. Human rights protections require practical accountability. Meaningful oversight must include the right to contest the account, not only the decision. Transparency must make the chain from capture to output to record to later use visible. The Dialogue should therefore treat record integrity, meaning integrity and temporal integrity as part of the global AI governance agenda.

Yes. A key cross-cutting issue is the governance of the travelling account. Current AI governance often focuses on models, datasets, risk categories, transparency registers, impact assessments and decision systems. These are necessary. But they do not fully capture what happens when AI-generated or AI-mediated outputs become institutional records. The emerging risk is not only that an AI system produces a bad answer. It is that a plausible account enters a workflow, becomes a record, informs another decision, is summarized again, and later reappears as inherited institutional memory. At that point, the person affected may no longer be able to see where the account came from, what was inferred, what was omitted, who relied on it, or how to correct it. This creates three linked governance needs. First, record integrity: what entered the record, how it was produced, whether AI involvement is visible, and who is accountable at the point of creation. Second, meaning integrity: whether the account remains faithful to the person, situation or evidence it represents as it is summarized, translated, classified or reused. Third, temporal integrity: whether the account remains true over time, and whether there is a governed correction route when facts, circumstances or interpretations change. A further emerging issue is the way AI systems increasingly produce the social grammar of understanding, care, apology, confidence or accountability. In high-trust settings, the danger is not only error, but the appearance that understanding has taken place when a system has merely produced a plausible account. Governance should therefore distinguish between performed understanding and accountable human judgement.

In the United Kingdom, AI adoption is already moving into record-intensive public services and high-trust institutional settings. This creates both opportunity and risk. In healthcare, ambient scribing and AI-assisted documentation are being explored to reduce administrative burden and improve clinical workflow. In social care, transcription and summarization tools are beginning to shape case records. In policing and justice, AI may assist transcription, evidence management, case-file preparation and risk analysis. In higher education, AI is already embedded in student practice while institutional governance, assessment design and professional formation are still catching up. The opportunity is real. AI can reduce administrative burden, improve access to information and support stretched services. But governance often remains strongest at procurement, privacy and cyber-security level, and weaker at the level of meaning, record travel and correction. The most significant challenge is that institutional accountability can become blurred. Vendors may say the human reviewer is responsible. Institutions may say the tool only assists. Frontline staff may be left with the practical burden of checking fluent outputs under time pressure. The person affected by the record may have little visibility at all. Human oversight should not be treated as a symbolic safeguard. It is meaningful only where the reviewer has time, authority, competence, psychological safety and access to provenance sufficient to challenge, slow, correct or override the system. Otherwise, "human oversight" risks becoming a liability transfer onto frontline staff rather than a protection for affected people. The opportunity is to build governance around the actual pathway of use: what is captured, what travels, who relies on it, and how correction works.

The AI Dialogue can play an important role by helping international cooperation move from shared principles to shared operational questions. Many governance frameworks already refer to safety, human rights, transparency, accountability, inclusion and human oversight. The difficulty is that these terms can remain too abstract when AI systems enter real institutional workflows. The Dialogue can help translate them into practical tests that are usable across sectors and jurisdictions. One such test should be: what happens to the account produced by the system? International cooperation should not seek a single global regulatory template for every setting. Context matters. A clinical note, welfare assessment, school record, asylum summary, police file and employment profile each carry different risks. But the underlying governance questions often repeat: who or what interpreted the person; what was captured; what was inferred; what was changed; what travelled; who relied on it; and how can it be corrected? The Dialogue could therefore advance cooperation by developing shared language for AI-mediated records and institutional memory. It could support common expectations around provenance, correction routes, human review, contestability, explanation and downstream accountability. It should also create a bridge between technical governance and human-rights governance. Provenance standards, audit logs and transparency registers are important, but they are not enough unless they connect to the lived ability of affected people and communities to understand, challenge and correct what institutions make of them. The Dialogue's value lies in being universal without becoming vague: a forum where governments, civil society, technical experts, researchers and affected communities can name recurring governance problems and turn them into practical obligations.

The Dialogue should build on existing human rights, data protection, public-sector transparency, digital cooperation and AI governance initiatives, while adding a sharper operational focus on how AI-shaped outputs become institutional records. Relevant foundations include the Global Digital Compact, UNESCO's work on AI ethics, OECD AI principles and related policy work, human rights mechanisms addressing technology and discrimination, regional data protection and digital rights frameworks, and emerging public-sector algorithmic transparency standards. These efforts provide important scaffolding. The Dialogue should also recognize the growing role of arts, humanities and social science research in responsible AI. In the UK, the BRAID programme — Bridging Responsible AI Divides — offers a useful example. It recognizes that responsible AI is not only a technical agenda, but also a cultural, social, ethical and interpretive one. Arts and humanities expertise are not decorative additions to AI governance. They help make visible questions of meaning, memory, context, dignity, representation and public trust. Digital Narrative Care builds from that same recognition but focuses on the missing operational step: how to govern AI-mediated meaning once it enters workflows as records, summaries, classifications, assessments, decisions and institutional memory. The Dialogue should also learn from sectors with mature correction and redress traditions: medical records, child and adult social care, archives, administrative justice, legal disclosure, complaints systems and data protection. AI governance should not reinvent these fields, but it must update them for a world in which machine-shaped accounts can travel faster, further and with greater apparent authority than before.

Different stakeholders should contribute according to where they sit in the life of an AI-shaped account. Governments can identify where AI systems are already being used in public administration and set duties for provenance, review, correction and redress. Regulators can clarify how existing rights apply when AI outputs become records. Technical experts can explain what can be logged, tagged, versioned and audited. Civil society can identify harms that are invisible from the system owner's perspective. Frontline practitioners can show where human oversight is meaningful and where it is only nominal. Affected communities can say what it is like to meet an institutional account that does not recognize them. The Dialogue should be structured to prevent the most powerful stakeholders from defining the problem too narrowly. It should not only host general panels on AI safety or innovation. It should include practical, pathway-based sessions: for example, "AI in health records," "AI in migration and welfare decisions," "AI in education and assessment," "AI in policing and justice," and "AI in employment and workplace management." Each session should ask the same core questions: what does the system produce; where does the output travel; who relies on it; what rights does the affected person have; and how can the account be corrected? The Dialogue should publish summaries that preserve minority and dissenting views rather than smoothing them into artificial consensus. It should also maintain transparent version histories of draft outputs, so participants can see how their contributions were interpreted, condensed or excluded. The process itself should model the governance it recommends: provenance, contestability, correction and inclusion.

The most underrepresented voices are often those most likely to be (mis)represented by AI-shaped records. This includes patients, disabled people, care-experienced people, children and young people, migrants, refugees and asylum seekers, people in contact with police, courts or prisons, people receiving welfare or social protection, students subject to assessment or misconduct processes, workers subject to algorithmic management, and communities already exposed to surveillance, profiling or administrative suspicion. Participation should follow the principle of "Nothing About Us Without Us." Lived experience is not interchangeable with academic, technical or policy expertise. Knowing how to analyse a system is different from knowing what it means to survive one. This is especially important for disabled people and other communities whose lives are often treated as problems to be fixed by technology rather than as evidence of disabling infrastructures, institutional misrecognition or inaccessible design. Global AI governance must also avoid smuggling in a narrow "default human": white, Western, able-bodied, literate, institutionally fluent, professionally legible, and presumed rational unless proven otherwise. In high-consequence settings, many people arrive before systems already expecting to be misread. Responsible AI governance must therefore ask not only whether a system works for "humans" in general, but whose humanity the system has been trained, designed and authorized to recognize. Inclusion should be designed around pathways of consequence, not only stakeholder categories. The Dialogue should ask: whose lives are being interpreted, compressed, classified or remembered by these systems, and how can they participate before those accounts harden?

The Dialogue should use engagement formats that make AI governance concrete. One useful format would be "record journey" workshops. Participants would follow a fictional but realistic AI-shaped account through a pathway: for example, a patient consultation summary, a social care case note, a police incident record, a student misconduct file or a welfare eligibility assessment. At each stage, participants would ask what changed, who relied on it, what was visible, what became harder to challenge, and how correction would work. A second format would be "affected person review panels," where people with lived experience examine how AI governance language appears from their side of the record. Can they understand the system? Can they see what was inferred? Could they challenge it? Would correction reach the places where the account has already travelled? A third format would be "frontline oversight simulations." These would test whether human oversight is realistic under workload, time pressure, hierarchy and uncertainty. This is important because oversight often fails not through bad faith but through poor workflow design. The Dialogue should also use deliberative methods that preserve disagreement. AI governance outputs often smooth conflict into neutral language. But disagreement can reveal where legitimacy is fragile. Minority reports, dissent logs and unresolved-issue registers should be used alongside consensus documents. Finally, the Dialogue should maintain a public contribution map showing how submissions have been read, grouped and reflected in outputs. This would model provenance and accountability within the Dialogue itself. Meaningful participation is not only about who is invited. It is about whether their meaning survives the process.

Several practical approaches should inform effective AI governance. First, impact assessments should be expanded to include interpretive impact. Before procurement or deployment, institutions should ask what human situation a system interprets, what it treats as signal, what it compresses or excludes, whether AI origin is visible, and how accountability is exercised. Second, AI-shaped records should carry durable provenance. Where an AI system drafts, summarizes, classifies or materially reshapes a record, that origin should remain visible as the account travels across systems, teams and decisions. Third, high-consequence records should require documented human review before use. Review should include not only factual accuracy, but meaning integrity: whether the account remains faithful to the person, situation and uncertainty it represents. Fourth, correction routes must reach downstream users. It is not enough to correct the surface record if earlier versions have already informed referrals, decisions, summaries, risk flags or future AI outputs. Correction aftercare should become a governance expectation. Fifth, effective governance should include designed interruption points. Where provenance is uncertain, where a record has been over-redacted, where the source account is missing, or where an AI-shaped summary has begun to travel without adequate verification, systems should not smooth the gap into apparent continuity. They should mark the discontinuity, pause reliance where necessary, and require human review before the account is inherited further. Sixth, AI governance should include workforce preparation. People asked to oversee AI need role clarity, time, authority, training and psychological safety to challenge outputs. Digital Narrative Care offers one approach: protect record integrity, meaning integrity and temporal integrity before AI-shaped accounts travel.