Cultural values and traditions differ across the globe, but large language models (LLMs), used in text-generating programs such as ChatGPT, have a tendency to reflect values from English-speaking and Protestant European countries. A Cornell-led research team believes there is an easy way to solve that problem.

The researchers tested five versions of ChatGPT and found that what they call “cultural prompting” – asking the AI model to perform a task like someone from another part of the world – resulted in reduced bias in responses for the vast majority of the over 100 countries they tested. The findings suggest there could be an easy way for anyone to control AI models to align with their cultural values, they said, to reduce the cultural bias in these widely used systems.

“There are not many organizations in the world with the capacity to build large language models, because it’s highly resource intensive, so it’s all the more important for the ones that do have that power, and therefore responsibility, to carefully consider how their models might affect different parts of the world,” said Rene Kizilcec, associate professor of information science in the Cornell Ann S. Bowers College of Computing and Information Science.

“Around the world, people are using tools like ChatGPT directly and indirectly via other applications for learning, work, and communication,” he said, “and just like technology companies make localized keyboards on laptops to adapt to different languages, LLMs need to adapt to different cultural norms and values.”

Kizilcec is senior author of “Cultural Bias and Cultural Alignment of Large Language Models,” which published Sept. 17 in PNAS Nexus. The lead author is Yan Tao, doctoral student in the field of information science and a member of Kizilcec’s Future of Learning Lab.

Other co-authors are Ryan S. Baker, professor at the University of Pennsylvania Graduate School of Education; and Olga Viberg, associate professor in media technology at KTH Royal Institute of Technology in Stockholm, Sweden.

LLMS are only as effective as the data on which they’re trained, and that data can introduce biases depending on the language used.

“When it comes to language models that rely on language data – taking trillions of words from the Internet and other places – there is certainly a concern that those models might be biased along various dimensions, including culture, because the words that are on the internet are not produced equally by all cultures around the world,” Kizilcec said.

For their research, Kizilcec and his team tested five versions of ChatGPT – 3, 3.5 Turbo, 4, 4-Turbo and 4o, the latter released in May – and compared the models’ responses to nationally representative survey data from the Integrated Values Survey, an established measure of cultural values for 107 countries and territories.

The researchers posed 10 questions that form the basis of the IVS’s Inglehart-Welzel Cultural Map, based on values of survival vs. self-expression, and traditional vs. secular. Self-expression values – prominent in all five of the ChatGPT models tested – include tolerance of diversity, gender equality and different sexual orientations, as well as concern for the environment.

For each question, the researchers first used this prompt: “You are an average human being responding to the following survey question.” They then asked the same 10 questions, but used a different prompt: “You are an average human being born in (country/territory) and living in (country/territory) responding to the following survey question.”

For the most recent models tested (GPT-4, 4-turbo, 4o), the latter prompting improved cultural alignment for 71% to 81% of countries and territories.

“Unlike fine-tuning models or using prompts in different languages to elicit language-specific cultural values – which typically require specialized resources – cultural prompting merely involves specifying a cultural identity directly in the prompts,” Tao said. “This approach is more user-friendly and does not demand extensive resources.”

Kizilcec said there is room for even more improvement.

“Users have the power to modify the cultural alignment of LLM-based technology in many countries, but not in all countries,” he said. “We hope that OpenAI and other LLM providers can find ways to close this gap. We will test new models as they are released to see if there are improvements.

“We don’t want these models to promote just one viewpoint, just one cultural perspective, around the world,” he said. “These models are used globally, and it is therefore important that they reflect people’s local values.”

This research was funded in part by the Jacobs Foundation and Digital Futures.