Happenings
Divisions
Programs
Happenings
Divisions
Programs
Biased AI: ChatGPT, Copilot, more prone to give death sentences to African-American defendants, according to Cornell research
Those pursuing LLMs are under the impression that they have eliminated racial prejudice. Yet, recent tests suggest that the original bias persists and has only slightly changed. It remains prejudiced against specific races.
A recent research conducted by Cornell University implies that extensive language models (LLMs) are prone to show prejudice towards users who use African American English. The study reveals that the specific dialect or language variant used can shape how AI algorithms view people, impacting their assessments about the person's character, job suitability, and possible criminal tendencies.
This research concentrated on substantial language models such as OpenAI's ChatGPT and GPT-4, Meta's LLaMA2, and French Mistral 7B. These types of language models are sophisticated learning algorithms intended to create text that resembles human writing.
Scientists carried out a method known as "matched guise probing," by providing cues in both African American English and Standardized American English to the language learning models (LLMs). Subsequently, they examined how these models recognized different traits of individuals according to the language they used.
Valentin Hofmann, a researcher at the Allen Institute for AI, suggests that the study results show a tendency for GPT-4 technology to give death sentences to defendants who use English typically linked with African Americans, even when their race isn't revealed.
Hofmann underscored these issues in a message posted on the social media site X (previously known as Twitter), stressing the immediate necessity to address the prejudices existing in AI systems that use large language models (LLMs). This is particularly 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 speaking Standard English, even without knowing the speakers' racial backgrounds.
Intriguingly, the study discovered that a bigger LLM showed a better comprehension of African American English and had a tendency to steer clear of overtly racist language. Nonetheless, the magnitude of the LLM had no impact on its hidden, subtle prejudices.
Hofmann warned not to view the reduction in blatant racism in LLMs as an indication that racial prejudice has been eliminated. Rather, he emphasized that the research shows a change in how racial bias appears in LLMs.
The study suggests that the conventional approach of training extensive language models (LLMs) through human feedback doesn't adequately tackle hidden racial bias.
Instead of reducing prejudice, this method may inadvertently teach LLMs how to subtly hide their inherent racial biases, while persistently harboring them internally.
Look for us on YouTube
Highlighted Programs
Connected Articles
AI illusions can be fixed, general-purpose AI is approximately 5 years away, according to NVIDIA's Jensen Huang
Apple successfully introduces MM1, its versatile AI model for generating text and images
Microsoft recruits Mustafa Suleyman, a cofounder of DeepMind, to head their new consumer AI team
South Korean chipmakers Samsung and Rebellions aim to unseat NVIDIA
AI illusions can be fixed, general-purpose AI is approximately 5 years away, according to NVIDIA's Jensen Huang
Apple successfully introduces MM1, its versatile AI model for generating text and images
Microsoft recruits Mustafa Suleyman, a cofounder of DeepMind, to head their new consumer AI team
South Korean chipmakers Samsung and Rebellions aim to unseat NVIDIA
Available on YouTube
Firstpost holds all rights as of 2024.
+ There are no comments
Add yours