¿Qué pasaría si los gobiernos municipales crearan sus propios modelos de IA?

In Boston, Massachusetts, more than 37 percent of residents speak a language other than English at home. To help ensure that city services reach all residents, local law requires the municipal government to provide language access for the languages most commonly spoken by local residents: English, Spanish, Haitian Creole, Vietnamese, Cantonese, Mandarin, Portuguese, Russian, French, Arabic, Somali, and Cape Verdean Creole. Key government functions like voting and the basic city services application (311) also must be available in all of these languages.

New technology offerings, especially large language models (LLMs) with translation capabilities, offer an appealing way to meet these needs, as well as to provide translation resources that aren’t strictly mandated. However, as the City of Boston sought to use the technology to meet locally mandated standards, it identified a significant gap in coverage: there are more than 40,000 Cape Verdean Creole speakers in the greater Boston metropolitan area, but Cape Verdean Creole is not a language supported by Google, which provides free tools that many City employees and constituents rely on for language access. 

This reality invites bigger questions for city staff across the country: how compatible are large commercial tools like LLMs with the needs of local governments? Might other language technologies be more appropriate? What duties of care do public agencies have that their vendors must be prepared to meet? And where might it make sense for cities to step in and build a solution of their own? 

The needs of local governments are often not aligned with the incentives that shape private sector decision-making. In Boston, we see that private sector tools may fail to meet the needs of certain language groups if their language is not an industry priority. While cities can request locally-specific changes to these privately-owned resources, such as the addition of Cape Verdean Creole language support, such requests are unlikely to be addressed given the sheer scale of commercial LLM companies’ client bases. For example, in 2023 the City of San Jose requested information from Google about privacy and data usage in their language translation services, only to be ignored — in part due to the massive number of customers Google serves. This experience led the City to found the Coalición GovAI, banding together with other jurisdictions to advance the same information requests to vendors. 

What if, rather than relying on services from the private sector, municipal governments built their own AI models, or collectively stewarded open source alternatives? This could help ensure language access for constituents who are not represented or who are undersupported in commercially available LLMs and related offerings. And cities may be in a position to develop uniquely valuable AI-powered language services because of who they are and where they are located. In addition to being locally specific, cities have an opportunity to implement developmental approaches that are better aligned with their responsibilities. Imagine if city governments banded together to gather and collectively steward public information — with consent — that could form a dataset for model training. Imagine bottom-up efforts to augment public datasets with local expertise, such as linguistic minority language communities creating translation pairs for improving access to public services. Processes like this could be coordinated cross-jurisdictionally through ambitious national organizations like the Coalición GovAI or New America’s Rethink AI program. Together, partner organizations could collectively create and adapt open source models, and set new standards for transparency, accountability, and equity in system design and evaluation. An LLM stewarded by the public sector represents an opportunity to more deeply democratize AI — using existing jurisdictions and organizational infrastructure to allow for decisions over system data, model development, and allowable uses.

Strategies to Support Municipally Owned Language Technologies

What would it take to realize this vision? Today, local governments face an uphill battle in marshalling the necessary resources: access to compute and engineering talent are both costly. However, the critical path begins with collective action. The following steps can help cities bridge the service coverage gap by staking a path toward a “public AI” alternative to commercial LLMs, while being thoughtful about limited resources. These steps were charted in a workshop with Data & Society’s Colaboración para el Liderazgo en Tecnología Pública network.

Band Together

Working collectively and at scale, local governments have the power to pool data and resources in ways that are beneficial to everyone. An organization like GovAI Coalition could hire developers to experiment with open source models, using data collected and coordinated by coalition partners.

Focus on the Outcome, Not the Technical Solution

When determining what to build, a public sector LLM effort should begin with a vision of high priority use cases. In some cases the best solution might be training a new model, but there are many others in which fine-tuning an existing model could be the best approach. For instance, Google’s platform requires upwards of 30,000 language pairs before there is meaningful improvement to its model. Together, local governments might only be able to provide 1,000 language pairs, in which case that data may be more valuable for improving an existing translation system than for training a new model from scratch. A public sector LLM effort should be agnostic about deciding whether to train a new model or fine-tune an existing open source model.

Resource Your Project for Implementation y Maintenance

State and local governments should pool resources for sharing infrastructure and compute, technical capacity, and community engagement leads. Building, deploying, and maintaining a bespoke model can be energy- and cost-intensive, with third-party AI systems being especially costly due to cloud compute. Initiatives like the Frugal AI Hub and its corresponding principles offer guidelines for how to “do more with less” across compute, energy, data, and capital. Instead of relying on massive datasets, for instance, a public sector LLM effort might be able to achieve good results with smaller ones, reducing the cost of data processing and storage. 

Building and maintaining a model also requires significant technical and engagement capacity. Cities will need to hire or contract vendors who can build the actual model, staff who can deploy it, and staff who can perform ongoing evaluation. Staff should also be dedicated to leading engagement with the language community the model is meant to serve. In many cases, members of the language community or representative community-based organizations should play a role in validating the model and its output.

Work in the Open

Consider building in coalition with civil society and community-based organizations. Where commercial LLMs are often critiqued for being extractive systems built on data that was obtained without consent, working closely with organizations that represent different language communities can model consentful and participatory practices from the start. There are many different models for how participatory governance over datasets could look. ICPSR’s SOMAR or Mozilla Data Collective, for example, take different approaches to consentful data agreements through features like tiered access levels, time-bounded data use agreements, and contributor-controlled licensing. There are also examples of Indigenous and data sovereignty frameworks from the Maori language community to draw from: Te Hiku Media’s Kaitiakitanga License uses mechanisms like opt-in consent, joint decision-making, and a “living” license structure to keep governance authority within the language community.

Manage Risks with Strong Governance Practices

To ensure risks are managed in a way that is aligned with the public interest, the public should be engaged in transparent, structured, and accessible decision-making about how, where, and when these types of tools are used. Municipalities could consider working with community leaders to establish a language group governing board. Board members could receive targeted training and education to help them understand the risks and opportunities these tools present, and to address key implementation questions in partnership with municipal employees. This could include identifying which municipal services might be strong candidates for testing, advising on how feedback on model performance could be channeled most appropriately, or determining how additional data sharing opportunities should be handled. Requiring explicit board authorization to continue the pilot, on a regular basis, could also help ensure that implementation remains responsive to community concerns.

The goal of enabling every constituent to engage with their local government in their preferred language is a worthy one. Although commercially developed AI language models offer one path to realizing that goal, cities working together have an opportunity to build technologies that are both more effective and more accountable: expanding support for underserved languages, improving access to public services, and giving communities a meaningful voice in how these systems are governed and maintained.

Photo by todd kent on Unsplash