Unmasking Bias in AI: How Large Language Models Perpetuate Racial Discrimination

Activities

Divisions

Performances

Activities

Divisions

Performances

Racial Bias in AI: Findings from a Cornell study indicate that AI tools like ChatGPT, Copilot are more prone to sentencing African-American defendants to death.

Individuals pursuing LLMs often think they've overcome racial prejudice. Yet, recent research indicates that the original bias persists, albeit with minor changes. It remains prejudiced against specific ethnic groups.

A recent investigation by Cornell University implies that extensive language models (LLMs) are prone to show prejudice against individuals who use African American English. The study suggests that the particular dialect spoken can impact how AI algorithms perceive people, thereby influencing decisions about their personality, job suitability, and possible criminal behavior.

The research centered on extensive language models such as OpenAI's ChatGPT and GPT-4, Meta's LLaMA2, and French Mistral 7B. These are advanced learning algorithms programmed to create text that resembles human writing.

Scientists carried out a "matched guise probing" method, where they presented cues in both African American English and Standardized American English to the LLMs. They then examined how these models identified different traits of people based on the language they used.

Valentin Hofmann, a researcher at the Allen Institute for AI, suggests that the study's findings show GPT-4 technology tends to deliver death sentences more frequently to defendants who employ English typically linked with African Americans. This is despite no explicit disclosure of their racial identity.

In a post on the social media network X (previously known as Twitter), Hofmann accentuated the issues, stressing the immediate need to address the prejudices found in AI systems that employ large language models (LLMs). This is especially important in areas like business and law, where these systems are being used more and more.

The research also showed that LLMs often presume that individuals speaking African American English have less esteemed occupations compared to those using Standard English, even without knowledge of the speakers' racial backgrounds.

The study intriguingly discovered that a larger LLM demonstrated a better comprehension of African American English and tended to refrain from using openly racist language. Nonetheless, the magnitude of the LLM did not influence its implicit, hidden prejudices.

Hofmann warned not to view the decline in blatant racism in LLMs as an indication that racial prejudice has been eradicated. He emphasized that the research shows a change in how racial bias is displayed in LLMs.

The conventional approach of educating large language models (LLMs) through human feedback doesn't adequately tackle hidden racial prejudice, as the research suggests.

Instead of reducing prejudice, this method may inadvertently teach LLMs to subtly hide their inherent racial biases, while still preserving them at a more profound level.

Search for us on YouTube

Highlighted Shows

Associated News

NVIDIA's Jensen Huang suggests that AI hallucinations can be fixed and anticipates that complete artificial intelligence will be achievable in around 5 years

Apple has at last introduced its multimodal AI model, MM1, which is designed for generating text and images

Mustafa Suleyman, the co-founder of DeepMind, has been recruited by Microsoft to head their new consumer AI team

Samsung and Rebellions, two South Korean chip-making companies, aim to outcompete NVIDIA

AI hallucinations can be addressed, and complete artificial intelligence is expected to be realized in roughly 5 years, according to NVIDIA’s Jensen Huang

Apple has recently unveiled its multimodal AI model, MM1, which is capable of generating text and images

Microsoft brings onboard DeepMind's co-founder, Mustafa Suleyman, to manage their newly formed consumer AI team

NVIDIA faces competition from South Korean chip producers Samsung and Rebellions

This news can be found on YouTube.

Firstpost holds all rights and is protected by copyright © 2024.