AI Alliance’s Tapestry initiative is a potential solution for building culturally diverse open models amid cultural misalignment of AI models in the region.
In an interview with FutureCIO, Vincent Caldeira, chief technology officer at Red Hat, says that many AI models remain culturally misaligned because of the dominance of Western training data.
Citing a recent study examining the cultural alignment of AI models, he said: “What they saw was very shocking. They found, not surprisingly, clustering where the majority of models fell into North American alignment, irrespective of the country that tuned the model.”
According to Caldeira, the issue lies less in who develops the model and more in the datasets used to train it.
Cultural misalignment
For enterprises operating across the Asia Pacific, culturally misaligned models could undermine regional trust, impact customer interactions, and challenge the reliability of AI-generated outputs in local contexts.
Caldeira posits that addressing cultural misalignment is essential for defining ‘responsible AI’ across different cultures and regions.
“We already have Japan, India, and a number of governments that have manifested their interest,” he said. “I think it’s helping solve one of the main challenges I see with AI, that we risk losing cultural diversity because people don’t realise these systems may not reflect their own cultural models.”
Project Tapestry is an initiative by AI Alliance that aims to build a new open-source platform to enable distributed, globally federated training of frontier open models.
