Endunamoo Ltd — Drft Literacy

A successful first Global Dialogue on AI Governance would do more than restate existing principles. It would clearly identify the governance gaps that current frameworks still leave unresolved, especially where lived experience, technical design, and policy language do not yet meet. Success would mean three things. First, acknowledging that AI governance is not only about outputs, safety, transparency, and accountability after deployment, but also about what happens earlier in the process — including whether systems faithfully carry a person's instruction, identity, and meaning into generation without undisclosed alteration. Second, creating space for evidence from practitioners, artists, civil society, and communities who experience these systems directly, not only from governments and large technical institutions. Many structural harms become visible first through repeated use, not only through formal audits. Third, producing a clear pathway forward: named priorities, areas for consultation, and a commitment to develop governance language for gaps that are currently under-defined. One of these is instruction integrity — whether users are given what they asked for, or whether systems alter, smooth, or replace important elements before output is produced. For this dialogue to be credible, it should not end with broad agreement alone. It should result in sharper questions, clearer terminology, and an ongoing process that includes those whose evidence has so far sat outside formal governance spaces.

• 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

My selection reflects the specific governance gap my work identifies through Drft Literacy, an independent practice-based research framework developed through repeated testing of generative AI systems. I prioritised safe, secure and trustworthy AI because trustworthiness must include whether a system faithfully carries a user's instruction into generation, rather than altering important elements without disclosure. I selected social, economic, ethical, cultural, linguistic and technical implications of AI because the patterns documented in my research are not only technical. They affect identity, culture, representation, and meaning, especially where systems smooth, replace, or redirect what was originally asked. I selected protection and promotion of human rights because instruction-level alteration can have direct implications for dignity, equality, cultural expression, and fair representation. These issues are often experienced first by the people whose identities are most likely to be reshaped by system behaviour. I selected transparency, accountability, and human oversight because current governance frameworks still focus heavily on outputs, while under-defining what happens between instruction and output. Users are rarely told when their instruction has been materially altered before generation. Without transparency at that layer, accountability remains incomplete. Together, these areas best reflect the urgency of recognising instruction integrity as a governance issue. My contribution to this dialogue is centered on that under-named gap: what happens between what a human asks for and what the system actually begins processing.

Yes. One cross-cutting issue not yet clearly captured is instruction integrity at the point between human intent and system processing. Current AI governance discussions focus mainly on outputs, risk, transparency, accountability, and system behaviour after generation or decision-making. But there is an earlier layer that remains under-defined: whether the system faithfully carries a user's instruction, identity, meaning, and cultural specificity into processing in the first place. In practice, generative systems can alter, smooth, substitute, or redirect aspects of an instruction before output is produced, often without disclosure. This is especially important where identity, culture, language, and representation are involved. A system may appear compliant at the output level while still failing to return what was actually asked. This matters across multiple themes at once. It affects trustworthiness, transparency, accountability, human rights, cultural representation, fairness, and meaningful human oversight. It is therefore not only a technical issue, but also a governance and rights issue. A second emerging issue is the absence of a clear principle of prompt or instruction fidelity in current governance language. At present, frameworks rarely ask whether a user has the right to know when their instruction has been materially altered, softened, or reinterpreted by the system. For that reason, I would encourage this dialogue to consider instruction-level alteration as a distinct governance concern. International AI governance should begin to recognise and define: instruction integrity disclosure of material alteration before output prompt fidelity as part of trustworthy AI This is an emerging issue, but it is already visible in real-world use and deserves formal attention now

In the UK and across Western Europe, the most significant governance gap is that AI policy is moving faster than meaningful protection at the level where people actually experience harm. Current developments are improving awareness around transparency, accountability, safety, and human oversight, but there is still a major blind spot between formal governance language and lived system behaviour. One of the clearest challenges is that governance frameworks tend to assess risk at the level of outputs, system categories, or provider obligations, while ordinary users experience harm much earlier — at the point where their instruction, meaning, identity, or cultural specificity is altered before output is even formed. That gap matters in creative work, education, communication, representation, and public trust. A system may appear functional or even compliant while still failing to return what was actually asked. For my sector as an independent researcher and practitioner working on Drft Literacy, this creates two problems. First, people affected by instruction-level alteration often have no language to name what has happened. Second, current governance approaches do not yet consistently require disclosure when systems smooth, substitute, or redirect aspects of a prompt. The opportunity is that this region is already building serious governance infrastructure. That creates space to improve it before weak assumptions harden into global norms. If instruction integrity, prompt fidelity, and disclosure of material alteration were recognised within transparency, accountability, and human oversight discussions, governance would become more meaningful for real users. The challenge is not only building safe systems. It is making sure governance reflects how systems are actually experienced in practice.

The AI Dialogue can play a useful role if it does more than restate broad principles and instead helps countries, institutions, civil society, researchers, and affected communities identify governance gaps that are being missed across existing frameworks. Its value is not only in consensus, but in surfacing the issues that are still unnamed, under-evidenced, or unevenly understood across regions and sectors. It can help build international cooperation in three ways. First, by creating a shared language for harms and governance problems that are currently treated as isolated or local issues. Many people are observing the same system behaviours without a common vocabulary to describe them. That slows accountability and weakens policy responses. Second, by connecting formal governance discussions with documented lived experience. International cooperation will only be credible if it includes not just states, large companies, and established institutions, but also independent researchers, artists, educators, and civil society actors who are often the first to identify how systems behave in practice. Third, the Dialogue can help prevent narrow regional assumptions from becoming global defaults. If international cooperation is to be meaningful, it must make room for cultural specificity, instruction integrity, representation, and other issues that may not yet be fully captured by current safety and governance models. The strongest outcome would be a process that does not only harmonise existing approaches, but also remains open enough to recognise new categories of harm as they become visible. That is where real cooperation matters: not just aligning what is already known, but making space for what governance has not yet learned to see.

The AI Dialogue should build on existing governance work rather than duplicate it. That includes regional regulatory efforts, standards and risk frameworks, UN processes, technical standards bodies, human rights mechanisms, academic research networks, and civil society documentation of real-world system harms. It should also connect with independent practice-based research, because many important governance gaps are first noticed outside formal institutions. Its added value should be coordination, translation, and inclusion. Coordination means creating a space where policy, technical, legal, cultural, and lived-experience perspectives can meet without being treated as separate conversations. Translation means helping different communities understand how their concerns relate to one another, so that governance language is not limited to technical or legal actors alone. Most importantly, the AI Dialogue can add value by identifying what existing mechanisms still miss. Many frameworks already address safety, transparency, accountability, and human oversight, but there are still gaps around how systems alter instructions, meaning, identity, and cultural specificity before outputs are produced. Those gaps may not fit neatly into current categories, yet they are deeply relevant to trust, fairness, and human rights. The AI Dialogue should therefore not only connect established initiatives. It should also act as a bridge for emerging evidence and under-recognised harms to enter international governance discussions before weak assumptions become standard practice.

Different stakeholders should not be invited only to endorse pre-existing agendas. They should be able to contribute different kinds of evidence, language, and lived experience. Governments can contribute policy priorities, implementation challenges, and public-interest safeguards. Technical experts can help clarify what systems are actually doing and where governance assumptions break down in practice. Civil society can identify harms, exclusions, and blind spots that formal frameworks often miss. Independent researchers, artists, educators, and affected communities can contribute practice-based evidence of how AI systems behave in everyday life, especially where harms are subtle, cumulative, or not yet formally named. Industry should contribute operational knowledge, but not define the full terms of the discussion. For the format, the Dialogue should combine open submissions with structured thematic sessions and smaller cross-sector working groups. It should allow written evidence in accessible language, not only formal policy language, and it should include pathways for emerging issues to be raised even if they do not yet fit existing governance categories. The structure should prioritise three things: accessibility, balance, and traceability. Accessibility means multilingual participation, clear formats, and routes for smaller organisations and independent voices to contribute. Balance means no single stakeholder group dominating the framing. Traceability means participants should be able to see how submissions were grouped, interpreted, and reflected in outcomes. A useful Dialogue is not one where everyone says the same thing. It is one where different forms of knowledge can meet without being flattened. That is especially important in AI governance, where some of the most important gaps first appear in practice before they appear in law or standards.

Global AI governance discussions still underrepresent people who experience AI systems directly without having formal power over how those systems are built, deployed, or governed. This includes artists, educators, disabled people, small civil society groups, independent researchers, low-resource language communities, culturally specific communities, and people whose identities are frequently misrepresented, flattened, or altered by AI systems. It also includes practice-based researchers whose evidence may not come from large institutions but is still rigorous, documented, and highly relevant. Many governance conversations still privilege technical, governmental, and corporate voices while overlooking those who can show how systems behave in lived contexts. These voices should not be added symbolically. They should be included through concrete structural measures: open calls that welcome non-institutional evidence, multilingual submissions, financial support for participation, accessible meeting formats, and dedicated space for community-based and independent research. There should also be specific mechanisms for documenting harms that are cumulative, representational, cultural, or instruction-level, because these often fall outside standard reporting categories. Inclusion also requires expanding what counts as valid evidence. If governance only recognises formal technical audits or institutional reports, it will continue to miss important realities. Documented observation, comparative testing, artistic research, classroom experience, and community-led investigation can all reveal system behaviour that policy has not yet learned to name. A stronger global dialogue would not only ask who has expertise. It would ask who is repeatedly seeing the problem first, and why those voices are still treated as secondary.

A strong AI Dialogue should not rely only on panels and prepared statements. It should include formats that allow evidence, disagreement, and emerging issues to surface clearly. One effective format would be short evidence-led interventions, where participants present a specific governance gap, documented case, or practice-based finding in a fixed time and plain language. This would help bring in perspectives that are often lost in high-level discussion. Another useful format would be moderated cross-sector roundtables that mix governments, technical experts, civil society, independent researchers, and affected communities around one concrete question. The aim should not be consensus too quickly, but sharper understanding of where assumptions differ. The Dialogue would also benefit from open issue hearings for cross-cutting or under-recognised problems that do not yet fit neatly into existing governance categories. This is important because many AI harms appear in practice before they are formally named in policy. Structured response sessions could also help: after submissions are made, facilitators could identify recurring themes, tensions, and unresolved questions, then feed those back to participants for refinement. That would make participation more dynamic and traceable. To support meaningful engagement, all formats should allow written, spoken, and multilingual contributions, with strong moderation and equal speaking rules so smaller voices are not crowded out by institutions with more resources. The most useful format is one that does not flatten difference. AI governance needs spaces where technical evidence, lived experience, artistic research, and policy language can meet without one automatically overriding the others. That is how the dialogue can become more than a consultation exercise and instead function as a place where new governance gaps, including instruction-level harms and drft, can actually be recognised.

Effective AI governance needs more than high-level principles. It needs practices that can be applied before, during, and after system deployment. Useful examples include risk and impact assessments before release; documentation practices such as model cards, system cards, and data statements; audit trails and provenance records; independent red-teaming; and clear mechanisms for human review, appeal, and correction when systems cause harm. For generative AI specifically, governance should also include instruction-level testing, not only output evaluation. In my work through Drft Literacy, one of the clearest gaps is that current governance often asks whether outputs are safe, but not whether the system has remained faithful to the user's instruction before the output is formed. That means systems can alter identity, culture, or meaning at the prompt-processing stage without disclosure, while still appearing acceptable at the output level. A stronger practice would be to include prompt integrity checks as part of testing and accountability. This means examining whether a system preserves key instructed attributes, whether it alters protected or culturally specific elements, and whether those alterations are disclosed to the user. For affected communities, this is not a minor quality issue. It is a governance issue. Other promising approaches include participatory evaluation with affected users, multilingual testing, transparency reporting on known model limitations, and mechanisms for reporting repeated patterns of system alteration across platforms. In short, effective AI governance should combine technical safeguards, documentation, human rights protections, and lived evidence from practice. It should govern not only what systems output, but also what they do to an instruction before that output appears. That is where important harms can begin, and where current governance still has a blind spot.