Occurrences
Divisions
Performances
Occurrences
Divisions
Performances
Prejudiced AI: Study from Cornell discovers that ChatGPT, Copilot, and others are more prone to recommend capital punishment for African-American defendants
Those pursuing LLMs are under the impression that they have eliminated racial prejudice. Yet, recent tests reveal that the original bias persists, albeit in a slightly modified form. This bias still discriminates against particular races.
A recent investigation by Cornell University points out that there's a higher probability of substantial language models (LLMs) showing prejudice towards users who converse in African American English. The study suggests that the variation of language used can sway how artificial intelligence (AI) algorithms perceive people, which in turn affects the AI's assessment of their character, job suitability, and assumed propensity for criminal behavior.
The research concentrated on substantial language models such as OpenAI's ChatGPT and GPT-4, Meta's LLaMA2, and French Mistral 7B. These large language models are sophisticated learning algorithms developed to produce text that resembles human writing.
Scientists performed a study referred to as "matched guise probing," where they provided prompts in both African American English and Standard American English to the Language Learning Models (LLMs). The researchers then examined how these models discerned different traits of individuals from the language they used.
Valentin Hofmann, a scientist at the Allen Institute for AI, suggests that the study's findings show that GPT-4 technology tends to give death verdicts to defendants who use English typically linked with African Americans, even when their racial identity is not revealed.
In a message shared on the social media network X (previously known as Twitter), Hofmann underscored the immediate necessity to address the prejudices found in AI systems that use large language models (LLMs). He stressed that this is particularly crucial in areas like business and legal sectors where the use of such systems is on the rise.
The research also showed that LLMs often presume that individuals speaking African American English have less esteemed occupations compared to those who communicate in Standard English, even without knowledge of the speakers' racial backgrounds.
Remarkably, the study revealed that the bigger the LLM, the superior its comprehension of African American English, and it was more likely to refrain from overtly racist expressions. However, the magnitude of the LLM didn't influence its concealed, subtle prejudices.
Hofmann warned not to interpret the decline in apparent racism in LLMs as an indication that racial prejudice has been eradicated. He emphasized that the study shows a transformation in how racial bias is exhibited in LLMs.
The conventional approach of educating substantial language models (LLMs) through human input does not adequately tackle hidden racial prejudice, as suggested by the research.
Instead of reducing prejudice, this method may inadvertently train LLMs to effectively "hide" their inherent racial biases, continuing to harbor them at a more profound level.
Search for us on YouTube
Highlighted Shows
Associated Articles
NVIDIA's Jensen Huang believes AI illusions can be addressed, with artificial general intelligence expected in about 5 years
Apple finally introduces MM1, its blended AI model for generating text and images
Microsoft recruits Mustafa Suleyman, the cofounder of DeepMind, to head their new consumer AI division
Samsung and Rebellions, both South Korean chip makers, aim to surpass NVIDIA
AI illusions can be addressed, with artificial general intelligence projected to be ready in approximately 5 years, according to NVIDIA's Jensen Huang
Apple has at last presented MM1, its hybrid AI system designed for text and image generation
Microsoft has brought Mustafa Suleyman, one of DeepMind's founding members, onboard to oversee their new AI team for consumers
Samsung and Rebellions, two chip production companies from South Korea, have set their sights on outdoing NVIDIA
This is available on YouTube.
Firstpost holds all rights, protected by copyright, as of 2024
+ There are no comments
Add yours