Student group studying in University of Padua

Some international and regional frameworks already recognise that AI must uphold human rights, gender equality, non-discrimination, transparency, and accountability throughout the full AI lifecycle, yet current governance remains too general, too fragmented, and too weakly operationalised when it comes to language-specific harms. A successful Global Dialogue on AI Governance would move beyond reaffirming general principles and instead focus on translating them into more concrete and actionable commitments. While there is already a broad international agreement on the importance of human rights, equality, and accountability in AI, these values are often not sufficiently operationalised in practice. In our view, meaningful success would include the development of shared standards that explicitly address less visible but highly consequential risks, such as language-based harms in AI systems, as well as AI-facilitated gender-based violence. This includes issues such as gender bias, stereotyping, misgendering, and unsafe responses in conversational AI. The Dialogue could also help establish clearer expectations for multilingual and gender-sensitive risk assessment throughout the AI lifecycle. Another important outcome would be strengthening participatory approaches to governance. Ensuring that affected communities are involved in testing and evaluating AI systems would make it easier to identify harms that are often overlooked by purely technical assessments. Finally, the Dialogue could contribute to better international coordination, particularly in areas such as auditing, red-teaming, and accountability mechanisms. Since AI systems operate across borders, more aligned approaches would help reduce fragmentation and improve overall effectiveness. Thus, success would not necessarily mean creating entirely new frameworks, but rather making existing commitments more practical, inclusive, and enforceable.

• Safe, secure and trustworthy AI
• AI capacity-building
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Protection and promotion of human rights

Our selection is based on the observation that current AI governance frameworks already provide strong ethical foundations, but often remain too general to effectively address specific and everyday forms of harm. Through our research, we found that gender bias in AI systems, particularly in language, tends to persist in subtle ways that are not always captured by existing regulatory categories. Safe, secure and trustworthy AI and AI capacity-building: Because these forms of bias are often normalised or difficult to classify as clear violations, they risk remaining largely invisible or unaddressed. This is why we emphasise the importance of moving toward more concrete standards and practices. Translating broad principles such as non-discrimination and inclusivity into practical tools, such as testing procedures, dataset requirements, and accountability measures, would make governance more effective. Furthermore, by integrating AI literacy modules into educational programmes and professional training, the cycle of moving towards a safer and trustworthy AI governance and usage will be complete. Social, economic, ethical, cultural, linguistic and technical implications of AI: We also highlight the importance of participatory approaches, as many of these harms are best identified by those who experience them directly. Without their input, governance risks overlooking important dimensions of inequality and intersectionality. Lastly, we see Protection and promotion of human rights as essential. The gaps within AI systems governance clearly point to the fact that this is not merely a technical or one-time issue, but a rather systemic concern which encompasses access to justice, human rights, dignity, equality, and participation. A greater commitment to actors' obligations to protect rights, rather than optional ethical improvements would help prevent gaps and ensure that protections apply more consistently as they would be rooted in universally recognised ethical principles.

One important issue that remains underexplored is the role of language itself as a site of AI governance. Language is often treated as a secondary or surface-level aspect of AI systems, yet it is one of the main ways users interact with these technologies. Consequently, it is also where many forms of bias and exclusion are reproduced in everyday use. Another key and interrelated gap concerns the multilingual and cultural dimensions of AI bias. Much of the current discussion is shaped by Western-centric and English-language contexts, which can obscure how bias appears differently across languages and cultural settings. Our findings suggest that even when grammatical structures differ, stereotypes can persist through examples, tone, or assumptions embedded in responses. What is more, the increasing use of conversational AI raises concerns about anthropomorphic design. Systems that simulate empathy or human-like understanding may lead users to place unwarranted trust in them, particularly in sensitive situations such as seeking help for abuse or emotional distress. Finally, intersectionality remains insufficiently addressed. Gender bias is often approached in binary terms, without fully considering how it interacts with other factors such as race, class, disability, or geography. Addressing these issues requires more context-sensitive and inclusive, bottom-up approaches to governance.

As a group of students from diverse regions, cultural backgrounds, and linguistic contexts, we bring perspectives reflecting both youth and gender diversity. We are committed to promoting gender equality, human rights, inclusive participation, and access to justice for women in all their diversity, and for gender minorities. From this standpoint, we have examined the below-mentioned challenges. In the EU, there is already a relatively strong regulatory landscape for digital technologies and AI. However, governance gaps remain, particularly when it comes to language-based harms in AI systems. These issues often fall between different legislative frameworks, making them difficult to address in a coherent way. One of the main challenges is fragmentation. Different aspects of AI governance, such as data protection or platform regulation, are handled through separate instruments and institutions. This can make it difficult to respond effectively to everyday harms, such as biased or harmful outputs in conversational AI. In the global context, there are currently little concrete regulatory measures or safeguards for women and gender minorities that ensure the safe usage of AI systems. Further commitments, both at state and supranational levels, are needed in order to address the issue in its entirety and to apply an inclusive, intersectional and multi-level governance approach. Additionally, current approaches tend to be reactive. They rely heavily on transparency and complaint mechanisms, which place the burden on individuals to recognise and report harm. This is particularly problematic in cases where harmful outputs are subtle, short-lived, or difficult to categorise. The existing regulatory framework provides a foundation that could be expanded or adopted by different regions to better address these issues. There is also increasing awareness of systemic bias and the need for rights-based approaches to AI governance, creating space for more targeted and proactive measures, particularly in multilingual testing and participatory oversight.

The AI Dialogue has the potential to serve as an important international platform for coordination and exchange between different regions and actors. Given the global nature of AI systems, cooperation is essential to ensure that governance approaches are not only effective but also consistent across contexts. One of its key roles could be to facilitate the development of shared standards, particularly in areas where current approaches are still evolving, such as bias testing, gender and context-sensitive risk assessment, and auditing practices. By encouraging alignment, the Dialogue could help reduce fragmentation and make it easier to address cross-border challenges. It could also support knowledge and best practice exchange, especially between regions with different levels of resources and expertise. This includes highlighting perspectives that are often underrepresented, such as those from less dominant linguistic or cultural contexts. Importantly, the Dialogue can help bring greater attention to issues that are sometimes overlooked in technical discussions, such as the social, economic, ethical, cultural, and linguistic dimensions of AI bias. By framing these as core governance concerns, it can contribute to more inclusive and comprehensive approaches, which are based on respect for human rights, gender equality, and diversity.

The AI Dialogue should build on initiatives that already combine norms, tools, and participation. At the normative level, UNESCO's Recommendation on the Ethics of AI and the OECD AI Principles already provide widely recognised foundations on human rights, transparency, accountability, fairness, and human oversight, while the Global Digital Compact offers a broader UN framework for linking AI governance to inclusion and international cooperation. But the Dialogue should also connect with more practical and community-facing efforts. UNESCO's Red Teaming AI for Social Good: The PLAYBOOK is especially valuable because it treats testing as a participatory process and helps uncover bias, stereotypes, and harms that standard benchmarks may miss. UNESCO's Women4Ethical AI is another important initiative because it already brings together experts to advance gender equality in AI design and governance. Outside intergovernmental spaces, organisations such as the Algorithmic Justice League show why advocacy and lived experience matter: their work is built around amplifying the voices of communities most affected by AI harms. Similarly, NetHope demonstrates how responsible AI adoption can be developed in practice for humanitarian and development settings, where questions of access, capacity, and trust are especially important. The added value of the AI Dialogue would be to connect these efforts instead of leaving them fragmented. It could help translate broad principles into shared practices, connect technical testing with human rights and gender equality, and give more visibility and support to civil society and grassroots actors whose expertise often remains outside formal policy spaces. In that sense, the Dialogue should not replace existing work, but make it more coherent, more inclusive, and easier to learn from across regions.

Different stakeholders should contribute in ways that match both their responsibilities and their proximity to harm. Governments should set rights-based standards, support implementation, and ensure that AI governance is linked to existing commitments on human rights, gender equality, and inclusion. Companies should not only provide technical expertise, but also take responsibility for safer design, multilingual testing, dataset quality, and transparency when harms occur. Civil society, especially women's rights groups, LGBTQI+ organisations, youth groups, digital rights advocates, and survivor-support services, is essential because it often identifies harms earlier and more clearly than formal institutions do. Academia and the technical community can contribute independent research, evaluation methods, and evidence on how AI systems affect different languages, cultures, and social groups. The Dialogue should also recognise affected communities as contributors, not just end users. As our submission argues, women, gender minorities, speakers of underrepresented languages, and communities from the Global South often experience AI harms in ways that remain invisible in purely technical or policy-led discussions. Their lived experience is therefore essential to meaningful governance. In terms of structure, the Dialogue should not rely only on formal statements and high-level panels. It would benefit from a mix of plenary sessions, thematic workshops, multilingual consultations, and participatory formats where communities can directly surface harms and priorities. A multistakeholder advisory mechanism, similar to a MAG-style process, could help make participation continuous rather than one-off. Just as importantly, inclusion has to begin before the event itself: language access, cost, digital access, and institutional barriers should be addressed early, and the final summary should visibly reflect the input of underrepresented groups rather than only state and corporate positions.

Many of the people most affected by AI are still underrepresented in global AI governance discussions. This includes women and gender minorities who experience online abuse, misgendering, exclusion, or biased automated decisions directly; speakers of underrepresented and low-resource languages; communities from the Global South; and people whose realities fall outside English-dominant technical and policy spaces. Our own multilingual work showed that even in a grammatically gender-neutral language like Uzbek, stereotypes can persist because models absorb assumptions from dominant datasets. This shows that linguistic exclusion is not a side issue: it directly shapes whose experiences are visible and whose harms remain ignored. Other missing voices include Indigenous communities, local cultural actors, youth, community-based organisations, and survivor-support groups. These groups often recognise harms earlier and more clearly than formal institutions do, especially when those harms are subtle, cumulative, or culturally specific. Yet global discussions still tend to privilege highly technical, institutional, and English-speaking perspectives over lived experience. Including these voices requires more than simply inviting them to participate. The Dialogue should provide multilingual access and interpretation, hybrid and low-bandwidth participation options, travel and participation support, and space for community-led interventions. It should also create safer ways to contribute for people who may face online backlash, stigma, or institutional exclusion if they speak publicly. Most importantly, participation should follow a "nothing about us without us" approach. Affected communities should not only be consulted at the margins, but meaningfully involved in setting priorities, shaping agendas, and reviewing outcomes. That is essential if AI governance is to be inclusive, culturally aware, and responsive to real-world harms.

The Dialogue would be much more meaningful if it moved beyond speeches and formal panels and created spaces where people can actually work through AI governance problems together. One useful format would be multilingual testing and red-teaming sessions, inspired by UNESCO's Red Teaming Artificial Intelligence for Social Good: The PLAYBOOK. These sessions could bring together technical experts, civil society, youth groups, and affected communities to test AI systems in real time and identify harms that standard benchmarks often miss, especially around language, bias, and exclusion. It would also be valuable to establish a Multistakeholder Advisory Group (MAG), similar to the model used in the Internet Governance Forum. This would allow governments, companies, academia, civil society, and underrepresented communities to shape the agenda and outcomes together, instead of participating in isolated or one-off ways. The Dialogue should combine cross-cutting sessions on human rights, accountability, transparency, and linguistic justice with sector-specific deep dives on issues like education, healthcare, labour, gender-based violence, and platform governance. This would help connect broad principles to concrete policy problems. Another useful format would be scenario-based workshops. Instead of only discussing risks in the abstract, participants could work through realistic cases, for example how an AI chatbot responds to someone reporting abuse, or how a model behaves across different languages and cultural contexts. These kinds of exercises make governance discussions more grounded and easier to connect to real-world harm. Finally, participation should not end when the event closes. The Dialogue should include follow-up mechanisms, open feedback channels, and a transparent process for shaping the final summary. That would make engagement more continuous, less symbolic, and more responsive to emerging challenges.

Some of the most effective approaches are the ones that already work for people in practice, not only on paper. Australia's eSafety Commissioner is a strong example because it gives victims of image-based abuse a clear place to report harm and links reporting to concrete action, including removal orders and support. That kind of survivor-oriented model matters because AI harms often move fast, cross borders, and leave victims navigating fragmented systems on their own. Nordic digital-inclusion programmes also offer a useful lesson: sustained investment in digital skills and online safety, especially for women, young people, and marginalised groups, can strengthen resilience before harm escalates. A second promising approach is public red-teaming. UNESCO's Red Teaming Artificial Intelligence for Social Good: The PLAYBOOK is valuable because it does not treat testing as a task for engineers alone. It shows how communities, civil society, and affected users can help uncover bias, stereotypes, and safety failures that standard benchmarks often miss. This is especially relevant for multilingual AI, where harms may appear differently depending on grammar, culture, and context. Our own prompt audit reached a similar conclusion: bias persisted across English, Italian, Polish, and Uzbek, even if it surfaced in different ways. A third useful model is gender-responsive AI governance from the start. UNESCO's 2021 Recommendation on the Ethics of Artificial Intelligence and UN Women's work on gender-responsive AI both point in the same direction: bias prevention, inclusion, and accountability should be built into the whole AI lifecycle, not added only after harm occurs. For us, the added value of the AI Dialogue would be to connect these examples, amplify them internationally, and help turn isolated good practices into shared standards that are multilingual, inclusive, and grounded in lived experience.