Rothko uses a proprietary architecture and training corpus to predict how individuals and groups make decisions.

The model, which is based on three years of research, takes a pre-trained model and changes its architecture so that it can produce decisions as outputs rather than natural language text.

Rothko is then trained on human behavioural data, combining sources such as anonymised social media information, transaction histories, mobility traces, unstructured expert interviews and traditional surveys.

Artificial Societies said that the model is also designed to understand and try to address the ‘say-do-gap’.

Rothko can be used for predicting purchase intent, gauging likely public reactions to corporate strategies, helping with human-AI interaction and also tackling misinformation and polarisation.

Artificial Societies was founded in London in 2024 by James He, Patrick Sharpe and Tom Whittle and is a startup focused on building a foundational model of human societies. Artificial Societies is backed by Point72 Ventures, Kindred Capital and Y Combinator.

James He, co-founder and chief executive at Artificial Societies, said Rothko would help to “provide clarity and conviction for the world’s most consequential decisions” and marks a “next milestone on a journey that started with writing the first peer-reviewed paper on AI societies”. 

Felix Wallis, head of research at Artificial Societies, added: “Creating a foundational model of human behaviour requires rethinking the architectures and data we typically use to train these large models from the ground up.

“It’s extremely difficult, but once you can do it, you unlock huge value across marketing, corporate strategy, cognitive defence and AI safety.”