Algorithmic Divergence in Artificial Intelligence
Abstract
With the evolution of 4th and 5th generational warfare, grokking local culture is an essential component to effective military operations. Lack of cultural intelligence and failing to adapt to new cultural settings to make accurate cultural inferences has exacerbated tensions as well as cognitive load on the soldier. Sociocultural failures have dire strategic, operational, and tactical consequences, such as when locals misinterpret a soldierÕs gestures at checkpoints. As AI is increasingly applied to military problems, AI solutions must be both accurate and provide culturally-relevant interpretations and insight. Achieving this capability however is compounded by cultural bias influencing the very design of current state-of-the-art AI systems. Culture encompasses oneÕs language, belief system, perceptions, and social-cognitive systems of organization. It permeates the ways individuals reason, solve problems, categorize, judge, and predict future events. Culturally-driven worldviews shape the ways humans create products. AI development is a creative process led by human designers, and at each decision point, designers draw from their cultural perspective. Cultural differences in social orientation and cognitive style can subtly influence AI designer decisions, leading to unintentionally biased solutions. Significant work today is focused on addressing this bias in measurement, aggregation, and evaluation. Cultural divergence in AI however has been overlooked. Cultural divergence occurs when cultures significantly deviate in approaches to development, application, and interaction with AI and it remains unexplored for implications to AI design, function, robustness, and vulnerability, posing significant consequences to the long-term development of AI. Historically, cultural divergence can result in far-reaching technological advancements for different cultures. With AI, algorithmic designs may be viewed as weak or ineffective by one culture and cognitive style, but another group pursues the designs, creatively applying algorithms to advance development. Characterizing and recognizing divergence is critically important to optimal AI design, workforce diversity optimization, and for anticipating vulnerabilities that may be exploited. This proposal is for seedling funding to characterize cultural divergence in AI algorithm development. This seedling will attempt to prove the presence (or absence) of algorithmic divergence in AI development. Deliverables will include three reports outlining the data utilized, the methods, and evidence regarding divergent trends in algorithm design. The risk of this effort is minimal; the greatest risk is being able to confidently draw conclusions regarding divergence. We will mitigate this risk with comparison to random, control baselines and stringent analysis requirements. All documents and analysis scripts created as part of this effort will serve as deliverables with unlimited government purpose rights. If divergence exists, the results of this seedling will have pertinent repercussions for future DoD and IC AI as they will need to both protect and exploit implications of divergence. This seedling will guide a future program on divergent AI, including the development of predictive algorithms to estimate how divergence may comprise an AI solution, the design of systems to protect against the long-term effects, and the utilization of divergence to guide culturally sensitive and context-aware AI applications.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 20, 2020
- Source ID
- W911NF2010293
Entities
People
- Nichola Lubold
Organizations
- Army Contracting Command
- Defense Advanced Research Projects Agency
- Honeywell International, Inc.