Are AI Companies Working to Improve Cultural Representation in Their Models?

Steps Companies Have Taken

Some AI companies have made public efforts specifically aimed at improving cultural representation within their models. This has included sourcing more diverse and multilingual training data intended to reduce the imbalance toward heavily represented cultures and languages, working with regional experts, cultural consultants, or affected communities to identify specific representational inaccuracies, and incorporating dedicated cultural bias evaluations into pre-release testing processes for new models or significant updates. These efforts reflect a recognition within parts of the AI industry that cultural representation is a legitimate quality and fairness concern that merits deliberate attention, rather than something that will resolve on its own as models generally improve.

Uneven Adoption and Ongoing Challenges

Despite these efforts, the scope, consistency, and depth of investment in cultural representation work varies considerably across the AI industry. There is no single, universally adopted industry standard specifically requiring companies to meet particular benchmarks for cultural representation, meaning the degree of attention paid to this issue can differ significantly from one company or product to another. Even among companies that have made public statements or commitments regarding cultural representation, documented examples of misrepresentation or inaccuracy have continued to surface, sometimes even in newer model versions, indicating that this challenge has not been fully or permanently resolved by current efforts.

This pattern is consistent with how researchers describe algorithmic bias more broadly: addressing it well generally requires sustained, ongoing attention across each new model version and update, rather than being something achieved with a single fix that then remains permanently solved.

Why This Remains an Active, Evolving Area

Because AI models are frequently retrained, fine-tuned, and updated, cultural representation efforts generally need to be revisited with each significant model change rather than being treated as a one-time accomplishment. Additionally, given the sheer diversity and number of distinct cultures, subcultures, and languages around the world, comprehensively addressing representation for all of them is an enormous and ongoing undertaking that likely exceeds what any single round of data sourcing or evaluation could fully achieve. Researchers and advocates in this space generally frame progress on cultural representation as incremental and continuing, rather than something that has reached, or is likely to soon reach, a definitive endpoint.

Bottom Line

Yes, some AI companies have taken meaningful steps to improve cultural representation in their models, including more diverse training data sourcing, partnerships with cultural experts, and dedicated bias evaluations. However, these efforts vary considerably across the industry, lack a unified standard, and documented representational issues continue to surface even in actively updated models, indicating this remains an ongoing area of work rather than a fully solved problem.

Go deeper

• Can AI Translate Idioms and Cultural Nuance Correctly?