Teaching AI to Speak our Language: A Showcase of Global Efforts to Bridge the ‘Last Mile’ of AI Inclusion for Local Impact

James D'Ercole highlights that communication has been a persistent challenge in UN peacekeeping for over 20 years, with field reports consistently identifying it as a problem. With approximately 50,000 uniformed personnel from around 120 troop and police contributing countries deployed across 11 missions, the scale of the communication gap is significant. The diversity of peacekeepers' home countries means that inter-peacekeeper communication is also a major challenge, compounded by the fact that many come from low-resource language backgrounds.

D'Ercole noted that there are about 50,000 or more uniformed personnel in 11 missions, representing approximately 120 troop and police contributing countries . He stated that since being in peacekeeping for over 20 years, communication has always been a problem, with countless field reports from governments, military and police advisors, and human rights bodies consistently identifying it . He also pointed out that peacekeepers themselves come from low-resource language backgrounds, making inter-peacekeeper communication a major challenge .

The project focuses on developing a real-time translation tool to close the communication gap between peacekeepers and local communities in Juba, South Sudan. A critical design requirement is that the tool must function completely offline, as the environments where peacekeepers operate have little to no connectivity. An edge device functioning as a mini-server with battery pack capability is being used to enable this offline functionality.

D'Ercole explained that the solution focuses on real-time translation as a tool to help close the communication gap , and that the device must work completely offline in austere environments with no signal . He described an edge device on loan that functions like a mini-server, capable of operating remotely without any connectivity as long as it has a battery pack .

D'Ercole emphasises that the AI translation tool is designed to assist human interpreters and language assistants rather than replace them, with a person always present to validate outputs. The human-in-the-loop principle is central to the project's AI safety approach, ensuring that human judgment is never removed from the process. Rather than eliminating jobs, the tool is intended to amplify the reach and influence of language assistants.

D'Ercole stated that the human is always in the loop, AI assists the process, a person will always validate, and it never replaces human judgment or removes jobs, but instead amplifies and elevates language assistants . He described the proof of concept as a shadow mode where the tool listens to an interpreter and the interpreter then validates outputs to help build guardrails .

The project is using a physical edge device on loan that acts as a portable mini-server, enabling offline translation capabilities in remote field environments. The proof of concept involves deploying this device in shadow mode, where it observes and listens to human interpreters working in real situations. The interpreter then reviews the device's outputs, providing yes/no feedback to help build guardrails and iteratively improve the model.

D'Ercole described a sample edge device on loan that functions like a mini-server, operable remotely without connectivity as long as it has a battery pack . He outlined the proof of concept plan to deploy it in shadow mode, where it listens to an interpreter and the interpreter validates outputs to build guardrails through an improvement cycle .

D'Ercole argues that the project is building a cultural corpus rather than merely a language dataset, recognising that language cannot be separated from its cultural context. The tool must incorporate local meaning, register, and context specific to Juba Arabic to avoid producing outputs that are culturally inappropriate or misleading. National colleagues in the mission are being engaged to validate existing data and collect additional data to make the corpus more robust and culturally grounded.

D'Ercole stated that the project is building a cultural corpus, not just a language dataset, and that national colleagues in mission are being used to validate data and collect more to add cultural depth . He noted that an expert has been engaged to integrate culture into the AI language models in order to make the model less Western-centric and more specific to Juba .

D'Ercole argues that digitising languages such as Juba Arabic has a preservation benefit that extends beyond immediate communication needs, as preserving the language also preserves the culture it carries. Once a language is digitised and its data made open source, it could eventually attract commercial partners who might develop further tools, potentially enabling speakers to access the internet in their own language. This cultural preservation dimension is described as one of the most important aspects of the project for D'Ercole personally.

D'Ercole stated that by digitising languages, the project is preserving the language and, in doing so, preserving the culture, which he described as really important . He added that once digitised and open source, commercial partners could one day pick up the language, potentially enabling communities to surf the web in that language .

D'Ercole argues that AI safety cannot be an afterthought but must be embedded from the very beginning of the project's design. Community ownership is central to this approach, with guardrails and acceptable use boundaries being mutually defined with the community rather than imposed externally. The project also seeks to ensure cultural and ethical alignment with the community through engagement with national universities and government.

D'Ercole stated that the project wants to do AI safety by design, integrating it from the beginning and amplifying reach . He described the goal of mutually defining guardrails with the community, building local ownership, and working with the government through national universities to ensure cultural and ethical alignment .

D'Ercole highlights that working with voice data introduces specific and complex data protection challenges, particularly around protecting individual speakers' identities while still making the data useful. The project is guided by the UN's Office of Data Protection and Privacy, a new office that has been both helpful and challenging to work with given its own nascent status. The do-no-harm approach and community ownership principles are applied throughout to manage these risks.

D'Ercole noted that voice data is quite challenging to work with and that the project aims for safe public text and protected speech without exposing individual voices . He stated that all of this is done under the guidance of the Office of Data Protection and Privacy, a new UN office that has been fabulous but also difficult to move through because of its newness .

D'Ercole describes a multi-layered partnership approach for the Juba project, involving local institutions, international universities, civil society organisations, and NGOs. A proposed language cooperative would allow the communities contributing language data to eventually benefit from that data, particularly if it is made open source with a paywall for commercial entities. National colleagues have been especially supportive, with language assistants voluntarily contributing their own time because they believe in the project.

D'Ercole described exploring partnerships with local institutions, international universities, CSOs, and NGOs, and the idea of forming a language cooperative so that contributing communities can benefit from the open-source data . He noted that language assistants contributed on their own time during the online volunteer phase because they believed in digitising Juba Arabic .

Ansari argues that the distribution of text data available online, which underpins the performance of large language models, is deeply skewed towards languages spoken in wealthier countries. Languages with tens or hundreds of millions of speakers in lower-income countries have very limited data and resources, meaning they are poorly served or entirely absent from frontier AI models. This structural inequality in data availability directly translates into inequality in AI performance and accessibility.

Ansari presented research showing how much text data exists online across languages, indicating that languages spoken in wealthier countries dominate while others, including those with hundreds of millions of speakers, have limited resources . She cited the example of Breton, spoken by about 200,000 people in France, which has over 50 AI models, compared to Nigerian Pidgin with about 85 million speakers, which has only 10 models .

Ansari argues that the gap in language coverage is even more severe for speech models than for text models, with only a fraction of global languages meaningfully supported by automatic speech recognition. In northeast Nigeria, of the ten most common primary languages spoken in Borno, Adamawa, and Yobe states, only Hausa has any ASR support, and even that is only tested in laboratory conditions. This means that ASR-assisted workflows relying on Hausa or English risk excluding approximately 70% of the population of northeast Nigeria.

Ansari presented a map showing the most common primary languages in northeast Nigeria, noting that of the ten languages identified, only Hausa has any ASR recognition including a commercial API . She stated that this means roughly 70% of the population of northeast Nigeria risks being excluded from AI-assisted communication workflows .

Ansari highlights that many low-resource languages present unique challenges for AI model development because they are primarily oral, with no standardised written form. This makes the transcription process required to build ASR models extremely difficult, as terminology and expression vary widely between speakers, and code-switching between languages is common. The absence of a standard alphabet or writing system means that even the act of writing down a language requires co-design with communities to achieve sufficient consistency.

Ansari noted that some languages are primarily oral, making manual transcription for ASR development a really hard and complex process . She described wide variation in terminology, code-switching between languages, and the challenge this poses for building digital language technologies . She gave the example from Canary where one recording had ten different equally accepted ways to write it down, with contributors unable to agree on the correct version .

Ansari draws attention to the stark disparity between the number of speakers a language has and the number of AI models developed for it, arguing that this reflects structural inequalities rather than linguistic need. Breton, spoken by only about 200,000 people in France, has over 50 AI models, while Nigerian Pidgin, with approximately 85 million speakers, has only 10. This disparity illustrates how commercial and geopolitical factors, rather than speaker population size, drive investment in language AI.

Ansari cited a study she had read that day from Cambridge, noting that Breton, spoken by about 200,000 people, has over 50 AI models, while Nigerian Pidgin with about 85 million speakers has only 10 models . She noted the same pattern applies to languages like Seraki in northeast Nigeria .

Ansari argues that the quality of ASR models depends fundamentally on the quality and diversity of the voice data used to train them, and that achieving this requires a systematic and resource-intensive data collection process. Hundreds of hours of recorded data must be gathered from speakers who are diverse across multiple dimensions, including age, gender, educational level, and dialect. Without this breadth of data, ASR models will not perform reliably in real-world field conditions.

Ansari stated that you need hundreds of hours of recorded data done systematically, collected from across diverse speakers including by age, gender, educational level, and dialect, if you really want to make an ASR model that works well and widely in the field .

Ansari describes Clear Global's methodology as deeply community-centred, involving community members at every stage of the language technology development pipeline rather than treating them merely as data sources. The organisation trains community members as linguists, equipping them with the skills to conduct recordings and navigate the complexities of writing down primarily oral languages. With 100,000 linguists working across 300 languages, Clear Global brings significant linguistic expertise to this work.

Ansari described a community-centred approach to data collection, collaborating with community members to understand how models will be used and to understand linguistic complexities before any data collection begins . She noted that Clear Global trains over 100,000 linguists working in 300 languages, and that their background as linguists and translators informs the work . She also described co-design approaches to manage the complexity of writing down primarily oral languages .

Ansari argues that informed consent is a non-negotiable guiding principle for all of Clear Global's data collection work, particularly given the sensitive nature of voice data. Community workshops and engagement processes are used to ensure that contributors genuinely understand how their recordings will be used and stored, what the risks are, and how they can withdraw consent at any time. All datasets are ultimately published openly under non-commercial licences to ensure broad sector access.

Ansari described informed consent as a big pillar and guiding principle for their platform, implemented through workshops and community engagement to ensure people understand how recordings will be used and stored, what the risks are, and how they can withdraw consent at any time . She noted that all datasets are published openly under non-commercial licences so they can be used across the sector .

Ansari argues that language data collected from communities should be made openly available under non-commercial licences to maximise its benefit across the humanitarian and social impact sector. At the same time, she emphasises the ethical imperative of ensuring that communities are not exploited through the extraction of their language data, which is then used to build commercial models sold back to them. Fair wages for data contributors, rather than reliance solely on volunteers, are part of this ethical framework.

Ansari stated that all of Clear Global's datasets are published openly under non-commercial licences so they can be used across the sector . She also emphasised the importance of paying people fair wages to collect data and not exploiting people or extracting language data to build a model that is then sold back to communities .

Ansari raises the concern that when governments build sovereign AI models, they risk embedding the political and ethnic biases of those in power, potentially excluding minority languages and cultures from the national digital infrastructure. She uses South Sudan as a concrete example, noting that a government composed primarily of Dinka people may have little incentive to build a sovereign model that fairly represents the Nuer language and culture. She argues that the UN has a role in ensuring that global public infrastructure is built in a fair and equitable manner.

Ansari used the example of South Sudan, where she noted that if a government is composed only of Dinka, it is unlikely to want to build a sovereign model that reflects Nuer culture and language without bias . She argued that one of the roles of the UN must be to ensure that when building global public infrastructure, it is built in a way that is fair and equitable .

Ansari identifies a critical coordination gap in the language AI ecosystem, where multiple organisations are independently trying to develop similar technologies without collaborating, resulting in duplicated effort and prohibitive costs for each. She argues that international organisations are well-placed to play a convening role, bringing these fragmented actors together so they can pool resources and share the cost of data collection. Individual grassroots organisations typically cannot afford the $50,000–$100,000 required, but five or six of them contributing $10,000 each could make it feasible.

Ansari described the challenge of knowing that multiple organisations are all trying to develop similar technology but not coming together, and that if they worked together they could build relevant data at reasonable cost for each . She noted that data collection costs $50,000-$100,000, which most grassroots organisations cannot afford alone, but that five or six organisations contributing $10,000 each could manage it . She stated that this coordination work is very difficult and that Clear Global does not have the overview that international organisations do .

Ansari argues that the private sector should not be treated as a monolithic entity, and that a meaningful distinction exists between large technology companies such as Meta, OpenAI, and Anthropic, and smaller mission-aligned private sector actors such as Viamo and Lilapa AI. She notes that Clear Global has worked with Meta to test models and assess their safety, and that this kind of collaboration is appropriate and valuable. The core concern is not about the private sector per se, but about data governance and ensuring that communities are not commercially exploited.

Ansari distinguished between large companies like Meta, OpenAI, and Anthropic, and smaller private sector actors like Viamo, GSMA, and Lilapa AI, noting that Clear Global works a lot with the latter . She noted that Clear Global has also worked with Meta to test models and assess whether they are working well enough to be safe, and that Meta is not the enemy . She clarified that the key concern is data governance and how it is built to communities, noting that speakers of languages like Dinka are not commercially interesting to large companies .

Bromová argues that Serbian presents a distinctive digitisation challenge because, unlike many low-resource languages, it has a significant body of historical written material, yet this material is trapped on paper and unavailable for digital use. Serbian sits at the bottom of language data distributions despite this historical richness, partly because standard OCR models struggle with the language's use of both Cyrillic and Latin alphabets interchangeably. Historical scans present additional difficulties for OCR processing beyond those encountered with contemporary text.

Bromová noted that Serbian is not necessarily the lowest-resource language in the world and has quite a lot of historical sources, but that this material is on paper and not available for datasets or data scientists . She stated that Serbian sits at the very bottom of the data distribution and that standard OCR models had a very hard time with Serbian partly because the language uses both Cyrillic and Latin alphabets interchangeably, with historical scans presenting additional challenges .

Bromová describes the Librarify project, developed in collaboration with the National Library of Serbia, which created a workflow to digitise historical documents using image enhancement, segmentation, OCR, and post-OCR correction. The success of this project led to the development of Loria, a more generalisable and reusable open tool published on GitHub, designed to be adapted across different languages and institutions. Both projects place archivists at the centre of the human-in-the-loop process, recognising that institutional knowledge is as important as technical expertise.

Bromová described the Librarify project workflow, which prepares images from scans, segments them to distinguish text from images or headers, enhances segments, applies OCR, and performs post-OCR adjustment in a recursive improvement cycle . She noted that in the initial phase, over 400 gigabytes and over 16,000 documents from the library's archives were digitised . She described Loria as an open tool published on GitHub for reuse and adaptation, built with archivists at the centre of the human-in-the-loop process .

Bromová acknowledges that while UNDP is open to private sector partnerships in language digitisation work, there is relatively low interest from private sector actors in very low-resource languages because the speaker communities are not commercially significant online audiences. This makes it difficult for private sector partners to justify the investment required, creating a structural gap in funding for these languages. Some private sector partners are open-minded about data ownership and governance questions, but this is not universal.

Bromová noted that in approaching private sector partners, there is relatively low interest in doing work specifically on very low-resource languages because of the cost involved . She also noted that data ownership and data governance questions are open considerations, with some private sector partners being very open-minded but not all of them .

Bromová argues that the Loria tool represents a scalable and reusable solution to the problem of documents and data trapped in paper legacy formats, which is a challenge faced by many languages and public institutions around the world. By publishing the tool openly on GitHub, UNDP enables other institutions to download, adapt, and deploy it for their own languages and document types. The design places archivists at the centre of the process, recognising that understanding institutional processes and priorities is as important as technical expertise.

Bromová described Loria as an open tool published on GitHub for download and adjustment, built for reuse and adaptation across different languages . She emphasised that it was developed with archivists at the centre, and that being fluent not only in the technology and data but also in the processes and priorities of the institutions is critical for successful deployment .

Feinholz argues that meaningful linguistic and cultural inclusion in AI requires communities to be involved throughout the entire life cycle of a tool, from initial design through data collection, development, and deployment. A common failure mode is to include communities only at late stages of the process, which undermines genuine inclusion and ownership. This principle is embedded in UNESCO's approach to AI ethics and is reflected in the work of the Coalition for Linguistic Diversity in AI.

Feinholz stated that linguistic and cultural inclusion in AI, in order to be meaningful, must be done together with the respective linguistic and cultural communities across the entire life cycle, noting that communities are sometimes included very late in the process . She cited the example of speakers two presentations before as a model of including communities from the very beginning and throughout .

Feinholz describes the Coalition for Linguistic Diversity in Artificial Intelligence, established by UNESCO in partnership with the Icelandic Ministry of Culture, Innovation and Higher Education and the Icelandic Centre for Language and Technology. The coalition, launched in June 2025, brings together over 30 experts from almost all world regions who are working on community-led digital inclusion of languages. A key motivation is to enable sharing of experiences and good practices across communities rather than each working in isolation.

Feinholz described the establishment of an agreement with the Icelandic Ministry of Culture, Innovation and Higher Education and the Icelandic Centre for Language and Technology to create the Coalition for Linguistic Diversity in Artificial Intelligence . She noted that the coalition was launched in Bangkok and currently comprises more than 30 experts working on community-led digital inclusion of languages from almost all world regions .

Feinholz argues that the Coalition for Linguistic Diversity in AI is distinctive in its genuinely multi-stakeholder composition, bringing together governments, academia, communities, technical experts, international organisations, and the private sector in the same space. This diversity of perspectives is seen as essential for finding solutions that are technically sound, culturally appropriate, and politically viable. The coalition has found that what works for one community can often work for another, making cross-community sharing particularly valuable.

Feinholz described the coalition as bringing together governments, academia, communities, technical experts, international organisations, and the private sector, all sitting together in the same group . She noted that this has proven very useful because there is a lot of sharing of experiences, and that what works for one community can often work for another .

Feinholz describes the coalition's plan to build a repository of good practices in community-led linguistic and cultural inclusion in AI, making this knowledge accessible to practitioners who may not know where to find it. A co-created questionnaire, developed with all stakeholders, is being designed to systematise the collection of relevant information from projects in a methodologically coherent way. The repository is conceived as a living resource that will be continuously enriched and will eventually inform policy guidance on linguistic diversity in AI.

Feinholz described the plan to build a repository of good practices, noting that the idea is to include practices that have proven successful and to make this knowledge accessible . She explained that a questionnaire is being co-created between all stakeholders and UNESCO to identify what relevant information should be collected from each project, with stakeholders involved because they know their own projects best . She noted that findings will help inform policy guidance on linguistic diversity and AI, which has been identified as a priority area by many member states .

The audience member, who works with Meta on building large language models with a focus on internationalisation, identifies a dual cost problem in language AI development for low-resource languages. The first cost is the financial expense of building models in these languages, and the second is the challenge of finding people with sufficient linguistic expertise to contribute meaningfully, particularly when there is no canonical written form. This is compounded by the fact that even well-resourced languages face challenges around cultural nuance, such as the 80-plus pronouns in Vietnamese that carry precise relational meaning.

The audience member described the two-cost problem as either fighting to get a model in a language or fighting to get people who know the language well enough, noting that this is a privileged question . They gave the example of Vietnamese having 80-plus pronouns that carry precise relational meaning, illustrating the depth of cultural knowledge required . They also noted the challenge of Romanised scripts for languages like Telugu or Tamil, where there is no canonical version that is totally correct every time .

The speaker argues that the foundational approach to building language models for low-resource languages must start with the community and with the data, rather than with technology or commercial considerations. The goal should be to democratise access to knowledge, not merely to data, ensuring that the benefits of language AI flow back to the communities that contributed to it. Building a standardised orthographic structure with linguists and community members, potentially endorsed by bodies such as IEEE or UNESCO, is proposed as a way to preserve nuance and context in transcription.

The speaker described the approach of building a standardised orthographic structure with linguists who are experts in the language, potentially standardised through IEEE or UNESCO, to ensure nuance and context are not lost in transcription . They stated that the way to build it is with the community, and that the goal is to start with the data, start with the people, build it with the people for the people, and democratise access to knowledge and not just data .