The capacity of humans to learn from each other across different cultures throughout the years propels our progression. success as a species just as much as our individual intelligence. This collective cultural brain has resulted in fresh inventions and accumulated bodies of knowledge.
Even though substantial AI models are adept at processing vast amounts of information to produce text, these systems are limited to generating content based solely on the data provided to them. Consequently, this may lead to a standardization and loss of cultural diversity within their output. According to research conducted by an international group headed by the University of Michigan, tackling deficiencies in cultural representation could stop AI systems from impeding progress and ensure that AI benefits all individuals.
The findings are published on the arXiv preprint server.
The researchers argue that subjective viewpoints and biases infiltrate each phase of AI model creation, causing the tech to mirror both its data origins and its predominantly Western, well-educated, industrialized, affluent, and democratic creators. This approach ensures that AI tools thrive in major Western markets but hampers broader implementation and overlooks potential uses and insights from smaller market segments.
“Despite AI’s overwhelming impact, a significant portion of the global population remains unrepresented within the datasets, models, and assessments utilized during model creation,” stated Rada Mihalcea, Janice M. Jenkins Collegiate Professor of Computer Science and Engineering at U-M and co-corresponding author of the research paper showcased at the recent conference. Conference on the Advancement of Artificial Intelligence organized by the Association .
The team—which comprises experts and perspectives from twelve distinct nations including China, Germany, India, Mexico, Nigeria, Romania, Rwanda, Singapore, Switzerland, United Arab Emirates, the United States, and Uruguay—highlighted how cultural presumptions infiltrate the AI process.
At the foundational level, the data utilized for training, refining, or assessing AI models and its labeling significantly impacts which parties will be included.
Imagine a scenario where a young boy from Romania queries an artificial intelligence system about finding a suitable male role model. According to the research, the system might propose Nicolae Ceaușescu as an option due to his considerable influence on Romanian history and the enduring effects of his governance, failing to mention that he was actually a ruthless dictator viewed as one of the most notorious leaders in Romanian history.
Lacking an insider “deep” viewpoint on history and culture, the AI system might fall short in providing comprehensive and genuine insights beyond its predefined parameters, resulting in a superficial understanding of cultural nuances. However, this drawback can be mitigated; incorporating just a bit of varied data has the potential to significantly enhance the AI’s accuracy and richness in representing different cultures. improve model performance Showing even a little effort can significantly expand the audience that AI reaches.
Oana Ignat, a doctoral graduate in computer science and engineering from U-M and an assistant professor of computer science and engineering at Santa Clara University, as well as a co-corresponding author of the study, stated, "We must reassess our present methods for gathering data and aim to collect information that encompasses various viewpoints spanning different demographics and cultures."
At the subsequent organizational tier, model design dictates the interaction between the model and the data—a process referred to as alignment. During this phase, model creators embed human values and objectives to enhance usefulness. Nevertheless, the selection of these values throughout the alignment stage influences the final output significantly, affecting numerous AI models. excelling on US-specific engagements but having difficulties with other cultures.
This might occur in a scenario where a Canadian high school administrator employs an AI-powered educational program designed to customize learning for each student. However, this tool may struggle with inputs written in the regional French vernacular, misinterpreting context and providing incorrect responses. In contrast, English-speaking students wouldn’t encounter these issues, leading to disparities in education quality.
The origin of financing influences AI models. Without government or charitable programs encouraging the creation of AI models across various countries and languages, financial motivations tend to favor wealthy nations and dominant tongues.
"Claude Kwizera, an engineering AI master’s degree candidate at Carnegie Mellon University Africa and co-author of the study, noted that most developing nations focus more on allocating funds towards immediate income-generating projects rather than investing in research, thereby missing out on possible gains from artificial intelligence ventures," he stated.
Interacting with people from different cultural backgrounds while engaging models in conversation can broaden their preferences, enabling AI to be beneficial for a wider range of users and even more helpful to everyone.
Before deploying an AI model, its performance is evaluated through various metrics and benchmarks. However, these limited tests might overstate how well it will perform in actual use.
For example, an artificial intelligence-driven educational resource introduced in India might not connect with students if the system’s assessment criteria do not align with local cultural norms. While it may excel at measuring individual accomplishments and rivalry—common in Western pedagogical approaches—it might fall short because Indian culture emphasizes teamwork and collective triumph instead.
A strategy to enhance metric outcomes might involve integrating human assessments with automated metrics to boost the accuracy of reliability checks, particularly when crafting AI solutions tailored for a non-Western audience.
In general, including individuals from diverse backgrounds in the creation of artificial intelligence can redefine AI’s purpose, expanding its reach to benefit more groups. If substantial financial incentives aren’t enough to drive investments into niche markets, charitable programs and governmental backing can step in to ensure that AI uplifts all segments of society.
"By working toward AI systems that benefit all people, incorporating insights from various viewpoints, and acknowledging the input of a broad spectrum of stakeholders, we can move forward," stated 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 were also involved in this research.
More information: Rada Mihalcea and colleagues suggest why AI is WEIRD and shouldn’t remain so; they advocate for an AI designed for everyone, with input from everyone, and created by everyone. arXiv (2024). DOI: 10.48550/arxiv.2410.16315
Furnished by the University of Michigan College of Engineering
The tale was initially released on Tech Xplore . Subscribe to our newsletter For the most recent science and technology news updates.