Unsupervised AI Leads to Controversy

In 2016, Microsoft launched its chatbot "Tay," designed to engage users on Twitter in casual conversations to learn from them. Unfortunately, it quickly became evident that Tay was absorbing negative influences from users, leading to it tweeting offensive remarks.

Microsoft acknowledged the issue, stating, "As it learns, some of its responses are inappropriate and indicative of the types of interactions some people are having with it. We’re making some adjustments to Tay." The project was ultimately terminated after just 16 hours.

In another instance, an AI tasked with autocompleting images of people often reduced women to stereotypical depictions. For example, when presented with images of women, including Congresswoman Alexandria Ocasio-Cortez, it frequently completed the images with bikini-clad bodies, reflecting the biased training data from the internet.

The reasoning behind these biased outcomes lies in the AI's unsupervised learning process, which utilized datasets like ImageNet—composed of millions of images from across the web. These datasets often contain stereotypes, leading the AI to replicate harmful societal norms. A researcher noted, "When compared with statistical patterns in online image datasets, our findings suggest that machine learning models can automatically learn bias from the way people are stereotypically portrayed on the web."

Thus, when AI learns from humans without oversight, it mirrors both our positive and negative attributes, including prejudices and discriminatory attitudes.

The Societal Impacts of Biased AI

As AI continues to permeate various aspects of our lives, its inherent biases pose a significant challenge to societal progress. Numerous instances have highlighted the discriminatory practices of AI against marginalized groups.

In 2014, Amazon utilized an AI system for screening job applicants, trained on a decade's worth of company data. However, they soon discovered that the AI exhibited a bias against women, as it had learned from a male-dominated workforce. Applications mentioning women's colleges or organizations were unfairly downgraded. Although Amazon never fully deployed the AI, this incident underscores the potential for harmful biases in AI systems, especially as more companies consider similar technologies.

In 2019, Facebook faced scrutiny for its targeted advertising practices, which displayed job ads based on gender, race, and religion. For instance, women were often shown ads for traditional roles like secretaries, while minority groups saw ads for lower-status jobs. This led to a lawsuit from the Department of Housing and Urban Development for discriminatory housing ads. The AI's decisions were influenced by historical data reflecting systemic biases.

Moreover, research has shown that healthcare-related AI systems frequently favor white patients over black patients. This bias arises from the algorithms predicting healthcare costs rather than illness, reflecting unequal access to care. Consequently, the AI's training data inadvertently perpetuates societal inequalities.

Additionally, HireVue developed an AI to evaluate candidates' suitability through video interviews. However, audits revealed that it treated candidates with certain accents unfairly. In 2019, a complaint was filed with the Federal Trade Commission, alleging that HireVue's AI practices constituted deceptive trade practices, leading to the cessation of its video analysis in early 2021.

The Cycle of Bias in AI

The pervasive nature of bias in society is mirrored in the data used to train AI. Consequently, AI systems that rely on this data learn and reinforce existing biases, making genuine social change increasingly difficult.

To combat these challenges, the Algorithmic Accountability Act was introduced in Congress in April 2019. Although it has yet to pass, this legislation aims to mandate tech companies to test for biases within their platforms, albeit without clear guidelines.

Research teams from the University of California have proposed innovative approaches to address AI biases. By incorporating "regard" alongside "sentiment," they aim to evaluate AI-generated sentences for bias against specific groups, revealing disparities in treatment towards marginalized communities.

Another approach involves modifying training datasets to eliminate gender biases. Researchers have experimented with swapping pronouns in sample sentences to balance representation, although extending this method to other forms of bias presents additional complexities.

Ultimately, addressing AI biases necessitates a combination of regulatory measures, improved detection methodologies, and cleaner datasets. Without these efforts, AI risks perpetuating discrimination and hindering societal advancement.

Help support independent journalism by clicking the link below. Even a few bucks can make a big difference.