Every community has a language of its own. Most AI has never learned it.
Large language models are trained on the internet.
The internet reflects the world at scale — its dominant cultures, its most-published perspectives, its highest-volume sources of text. What it does not reflect, at any meaningful depth, is St. Petersburg. Or any specific city.
The decisions a city makes, the way its institutions communicate, the patterns in its civic participation, the texture of its neighborhoods, the language of its zoning boards and city commissions and local foundations — these exist in records, archives, meeting minutes, community media, and civic data that never made it into the training sets of general AI models. The local signal is present. It is just too small to register.
The Civic Language Model is AICOE's long-horizon research program to change that.
The decisions a city makes, the way its institutions communicate, the patterns in its civic participation, the texture of its neighborhoods, the language of its zoning boards and city commissions and local foundations — these exist in records, archives, meeting minutes, community media, and civic data that never made it into the training sets of general AI models. The local signal is present. It is just too small to register.
The Civic Language Model is AICOE's long-horizon research program to change that.
The biome principle
In nature, biomes are not just geographic regions. They are distinct ecosystems with their own species, their own feedback loops, their own logic. A forest in the Pacific Northwest and a wetland in coastal Florida share the same planet and many of the same underlying biological principles. But the organisms that thrive in one cannot be transplanted to the other and expected to behave the same way.
Communities are biomes. The civic patterns of St. Petersburg — its political culture, its development history, its institutional relationships, its neighborhood identities, its economic rhythms — are not generic. They are specific. And that specificity is exactly what makes a language model trained on local data fundamentally different from one trained on everything, everywhere, at once.
The CLM is trained on discrete, hyper-local data sources. The goal is a model that understands St. Petersburg the way a long-tenured city editor understands it. Not as an instance of a city in general, but as this place, with this history, these relationships, and these stakes.
Communities are biomes. The civic patterns of St. Petersburg — its political culture, its development history, its institutional relationships, its neighborhood identities, its economic rhythms — are not generic. They are specific. And that specificity is exactly what makes a language model trained on local data fundamentally different from one trained on everything, everywhere, at once.
The CLM is trained on discrete, hyper-local data sources. The goal is a model that understands St. Petersburg the way a long-tenured city editor understands it. Not as an instance of a city in general, but as this place, with this history, these relationships, and these stakes.
What the CLM trains on
Local data takes many forms. Civic records and government documents. Local news archives across decades. Public meeting transcripts. Community planning documents and neighborhood studies. Local economic and demographic data. The behavioral signals generated by Cityverse across every interaction — what residents engage with, what creators publish, what civic moments generate participation, what questions go unanswered.
The CLM is the intelligence layer that sits across all of that. As Cityverse expands to additional cities, each city's CLM trains on its own discrete data set, building a model specific to that community. The result is not one general civic AI. It is a system of city-specific models, each one tuned to its own biome.
The CLM is the intelligence layer that sits across all of that. As Cityverse expands to additional cities, each city's CLM trains on its own discrete data set, building a model specific to that community. The result is not one general civic AI. It is a system of city-specific models, each one tuned to its own biome.
What it enables
A CLM trained on St. Petersburg can answer questions that general models cannot. It can surface patterns in local civic participation. It can identify relationships between neighborhoods, institutions, and public decisions that only emerge from local data. It can power AI tools — from civic simulations to community intelligence surfaces to public-facing products — with the specificity that makes those tools genuinely useful to the people they serve.
Inside Cityverse, the CLM powers DASH, the platform's civic AI. When a resident asks a question about their city, DASH draws from the CLM first. The answer is grounded in local reality, not in the averaged-out knowledge of the internet at large.
Inside Cityverse, the CLM powers DASH, the platform's civic AI. When a resident asks a question about their city, DASH draws from the CLM first. The answer is grounded in local reality, not in the averaged-out knowledge of the internet at large.
A long-horizon program
Language model development takes time. The CLM is not a product to be launched in a quarter. It is a research program with a multi-year arc, designed to get more accurate, more useful, and more complete as the data it trains on grows.AICOE is building the infrastructure now.
The data collection, the training architecture, the evaluation frameworks, and the civic partnerships that will make the CLM a genuine civic resource over time. The first generation of the model will be imperfect. It will also be the first model of its kind — trained specifically, intentionally, and openly on the life of a city.
The data collection, the training architecture, the evaluation frameworks, and the civic partnerships that will make the CLM a genuine civic resource over time. The first generation of the model will be imperfect. It will also be the first model of its kind — trained specifically, intentionally, and openly on the life of a city.
get aicoe news to your inbox
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
%2FNegative.svg){kind=small}
© 2026 AICOE - ALL RIGHTS RESERVED