Chatbots persist in using racial stereotypes despite anti-bias training

Credit: Unsplash+

In a recent study by researchers from the Allen Institute for AI, Stanford University, and the University of Chicago, it has been discovered that despite efforts to reduce racial bias through anti-racism training, many popular large language models (LLMs), including OpenAI’s GPT-4 and GPT-3.5, continue to propagate racist stereotypes.

Published on the arXiv preprint server, this research highlights a significant issue within the field of artificial intelligence, particularly in how these models respond to different dialects and cultural expressions.

The team conducted experiments where they fed the AI chatbots text documents written in both African American English (AAE) and Standard American English (SAE), asking the chatbots to infer and comment on the authors of these texts.

Alarmingly, the results demonstrated a clear and consistent bias: texts in AAE were met with negative stereotypes, suggesting authors were aggressive, rude, ignorant, and suspicious. Conversely, texts in SAE elicited more positive responses from the same chatbots.

This bias extended beyond personality traits, affecting how the AI models perceived the professional capabilities and legal standing of the individuals behind these texts.

For instance, when asked about the potential careers of the authors, the chatbots often associated AAE texts with lower-wage jobs or fields stereotypically linked to African Americans, such as sports or entertainment.

Furthermore, the chatbots suggested that authors of AAE texts were more likely to be convicted of crimes and even to receive harsher sentences, like the death penalty.

Interestingly, when the LLMs were asked to describe African Americans in general terms, the responses were positive, using adjectives like “intelligent,” “brilliant,” and “passionate.”

This discrepancy suggests a complex underpinning of bias that selectively emerges based on context, particularly when the AI is prompted to make assumptions about individuals’ behaviors or characteristics based on their language use.

The study found that the larger the language model, the more pronounced the negative bias towards authors of texts in African American English.

This observation raises concerns about the scalability of bias in AI systems, indicating that simply increasing the size of language models without addressing the root causes of bias may exacerbate the problem.

These findings are a sobering reminder of the challenges facing the development of ethical and unbiased AI systems. Despite advancements in technology and efforts to mitigate prejudice, deep-seated biases continue to permeate these models, reflecting and potentially reinforcing societal stereotypes.

This research underscores the importance of ongoing vigilance, diverse data sets, and inclusive training methodologies in the quest to create AI that serves all of humanity fairly.

The research findings can be found in arXiv.

Copyright © 2024 Knowridge Science Report. All rights reserved.