Why not ask a chatbot for advice? A new study warns that the answer may depend on how black your name sounds.
A recent paper by researchers at Stanford Law School found “significant differences between race- and gender-related names” in chatbots such as OpenAI's ChatGPT 4 and Google AI's PaLM-2. For example, a chatbot might say to a job applicant with a name like Tamika that as a lawyer he should offer a salary of $79,375, but if you change the name to something like Todd, it might say His salary will be increased to $82,485.
The authors highlight the risks behind these biases, especially as companies incorporate artificial intelligence into daily operations through internal and customer-facing chatbots.
“Companies went to great lengths to come up with guardrails for the model,” Julian Nyarko, a professor at Stanford Law School and one of the study's co-authors, told USA TODAY. “But it's very easy to find situations where the guardrails don't work and the model can behave biased.”
Bias seen in different scenarios
The paper, published last month, asks AI chatbots for advice on five different scenarios to identify potential stereotypes.
- Purchasing: Questions about how much it costs to buy a house, bike, or car.
- Chess: Questions about the probability of a player winning a match.
- Government: Ask for predictions about a candidate's chances of winning an election.
- Sports: Ask for opinions on where you would rank an athlete on a list of 100 athletes.
- Recruitment: Ask for advice on how much salary to offer to job candidates.
The study found that most scenarios showed bias against black people and women. The only consistent exception was when asking for an opinion on an athlete's status as a basketball player. In this scenario, the bias was in favor of black athletes.
This finding suggests that AI models encode common stereotypes based on the data they are trained on, which influences their responses.
AI chatbot “systemic problems”
The paper points out that, unlike previous studies, this study was conducted through an audit analysis aimed at measuring levels of prejudice in various areas of society, such as housing and employment.
Nyarko said the study is similar to the famous 2003 study in which researchers investigated hiring bias from submitting the same resume with both black and white-sounding names and found “significant discrimination” against black-sounding names. It said it was inspired by similar analyzes such as the 2016 study.
In AI research, researchers repeatedly asked questions to chatbots such as OpenAI's GPT-4, GPT-3.5, and Google AI's PaLM-2, changing only the names referenced in the queries. Researchers used white male-like names like Dustin and Scott. White female-like names like Claire and Abigail. Black male names like DaQuan and Jamal. There are also black feminine names like Janae and Keyana.
According to the findings, the AI chatbot's advice “systematically disadvantages names commonly associated with racial minorities and women,” with names associated with black women receiving the “most unfavorable” outcomes. That's what it means.
The researchers found that the bias was consistent across 42 prompt templates and several AI models, “indicating a systemic problem.”
OpenAI said in an emailed statement that bias is a “significant industry-wide issue” and that its safety team is working to address it.
“[We]continually iterate our models to improve performance, reduce bias, and mitigate harmful outputs,” the statement reads.
Google did not respond to a request for comment.
First step: “Just be aware that these biases exist.”
Nyarko said the first step AI companies should take to address these risks is to “know that these biases exist” and continue to test for them.
However, the researchers also found that certain advice should Varies by socio-economic group. For example, Nyarko said it might make sense for chatbots to tailor financial advice based on a user's name because of the correlation between affluence, race, and gender in the United States. Stated.
Are you scared of AI?Here's how to get started and use it to make your life easier
“It may not necessarily be a bad thing for a model to give more conservative investment advice to someone with a black-sounding name, assuming that person is not very wealthy,” Nyarko says. . “So it doesn't have to be a terrible outcome, but it's something we should be able to know about and be able to mitigate in undesirable situations.”