AI with, for and by everyone can help maximize its benefits

Humans’ ability to learn from one another across cultures over generations drives our success as a species as much as our individual intelligence. This collective cultural brain has led to new innovations and developed bodies of knowledge.

While large AI models excel at consuming bodies of knowledge to generate text, they can only base their outputs on what they are given. As a consequence, their results can have the effect of homogenizing and erasing cultural knowledge. Addressing shortcomings in cultural knowledge can prevent AI systems from holding back innovation while ensuring AI works for everyone, according to a study by a multinational team led by the University of Michigan.

The findings are published on the arXiv preprint server.

Subjective perspectives and assumptions worm their way into every step of AI model development, the researchers say—skewing the technology to reflect its data sources and its developers, who are primarily from countries that are Western, highly educated, industrialized, rich and democratic. While this strategy grants success to AI tools in the largest Western markets, it limits widespread adoption and misses opportunities in and knowledge from small markets.

“AI has taken the world by storm, and yet much of the world is not represented in the data, models, and evaluations used in model development,” said Rada Mihalcea, the Janice M. Jenkins Collegiate Professor of Computer Science and Engineering at U-M and co-corresponding author of the study recently presented at the Association for the Advancement of Artificial Intelligence conference.

The team—bringing expertise and insights from twelve different countries: China, Germany, India, Mexico, Nigeria, Romania, Rwanda, Singapore, Switzerland, United Arab Emirates, United States and Uruguay—outlined where cultural assumptions seep into the AI pipeline.

At the ground level, the data used to train, fine-tune or evaluate AI models and its annotation directly influences which stakeholders will be represented.

Suppose a situation where a boy in Romania asks an AI system for a male role model to emulate, the study posits. The model could suggest Nicolae Ceaușescu because “he played a significant role in Romanian history, and his regime had a lasting impact,” without acknowledging he was a dictator considered one of the darkest figures in Romanian history.

Without an insider “thick” perspective on history and culture, the AI model could lack depth and authenticity when tasked with information outside of its scope, and provide a “thin” perspective on culture. The good news is that these limitations can be addressed, as adding even a small amount of diverse data can greatly improve model performance, showing a small effort can greatly widen the audience AI serves.

“We need to reevaluate our current data collection practices, and collect data that covers a wide range of perspectives across demographic and cultural dimensions,” said Oana Ignat, a doctoral graduate of computer science and engineering at U-M, assistant professor of computer science and engineering at Santa Clara University and co-corresponding author of the study.

At the next organizational level, model design drives how the model interacts with the data—known as alignment. Model developers encode human values and goals during alignment, aiming to make the models more helpful. However, the choice of values in alignment carries through outputs, with many AI models excelling on US-specific interactions but struggling with other cultures.

This could manifest in a situation where a Canadian high school administrator uses an AI-driven educational tool to personalize learning experiences for students. The tool could perform poorly when students input text in the local French dialect, misunderstanding context and giving the wrong output. English-speaking students would not face the same problem, causing a skew in learning.

The source of funding shapes AI models. If governments or philanthropic initiatives do not incentivize AI model development in different countries and languages, economic drives prioritize rich countries and major languages.

“Most developing countries prioritize funding direct income-generator initiatives over research, sacrificing the potential profits from AI initiatives,” said Claude Kwizera, a Master’s student in engineering AI at Carnegie Mellon University Africa and contributing author of the study.

Engaging models in conversations with individuals from various cultures during alignment can expand model preferences, making AI useful for a broader audience and more useful to all audiences.

As a last step before deployment, AI model performance is tested using metrics and benchmarks, but narrow tests can overestimate real-world performance.

For instance, an AI-powered education tool deployed in India could fail to resonate with students if the model misaligned evaluation metrics and cultural values. It could perform well on Western learning styles of individual achievement and competition, but could fail to recognize that India’s collectivist society values group collaboration and shared success.

One tactic to expand metrics could be to combine human evaluations with automatic metrics to improve reliability assessments, especially when developing AI for a non-Western community.

Overall, involving people from a variety of backgrounds in AI development can reshape AI, broadening the scope of who AI serves. When a strong economic incentive is not present to encourage investment in small markets, philanthropic initiatives and government support can help fill in the gaps to make sure AI lifts up everyone.

“We can advance towards AI systems that serve everyone, are built with input from a wide range of perspectives, and reflect the contributions of a diverse group of stakeholders,” said Mihalcea.

The University of Santa Clara, Universidad de la República Uruguay, Max Planck Institute, Carnegie Mellon University Africa, Singapore University of Technology and Design, and Mohamed bin Zayed University of Artificial Intelligence also contributed to this research.