Artificial Intelligence Models Show Bias Against Certain Spoken Accents
Researchers discovered that advanced artificial intelligence systems give lower scores to speakers with certain accents, even when the spoken words are identical. The evidence comes from a single study presented at a scientific conference.
In short: Researchers discovered that advanced artificial intelligence systems give lower scores to speakers with certain accents, even when the spoken words are identical. The evidence comes from a single study presented at a scientific conference.
When you talk to artificial intelligence, your accent might change how fairly the machine treats you.
What happened, in plain words
A team of researchers led by the University of Washington investigated how modern artificial intelligence models evaluate human speech in job interviews, presentations, and English tests. The team tested voices speaking English with five different accents using both computer-generated voices and human participants. They found that the models consistently ranked American and British accents the highest, Indian accents in the middle, and Chinese and Nigerian accents the lowest. The researchers presented these findings at a conference in Budapest, Hungary.
Key points
- Testing multiple accents: The team tested American, British, Chinese, Indian, and Nigerian accents across three different settings: job hiring, presentations, and English proficiency tests.
- Identical text used: To test fairly, the researchers had computer-generated voices read the exact same pieces of text in different accents to the artificial intelligence models.
- Good transcription, bad scores: Although the artificial intelligence models transcribed the words accurately, they still gave lower scores to Chinese and Nigerian speakers and claimed those speakers mumbled.
- Human verification: The team checked their results with fifty-two human participants who also read scripted responses and answered questions.
Terms explained
- Multimodal language models — Advanced computer programs that can understand and process more than one type of information at the same time, such as both text and spoken audio. Example: A computer program that can read a written story and also listen to someone read that same story aloud.
- Proprietary AI models — Artificial intelligence systems built and owned by specific private companies, where the inner workings and training data are kept secret. Example: A brand-new video game made by a private company that you can play, but you are not allowed to look at or change its underlying computer code.
- Acoustic distance — A measurement of how different the sound waves and pronunciation patterns of one accent are from another accent. Example: Comparing the sound of a word spoken with a very different pronunciation versus a standard pronunciation.
Critical analysis
What is encouraging. The study tackles a very real and growing problem as artificial intelligence tools take on roles like screening job applicants and evaluating school presentations. By testing both computer-generated and human voices across multiple popular models, the researchers thoroughly demonstrated how hidden biases can affect automated scoring systems.
What deserves caution:
- Limited accent variety. The study only looked at five national accents and did not test the wide variety of regional variations within those countries. This means the findings do not show how these models treat every possible way of speaking English.
- Selected models tested. Researchers focused on a small group of specific models, including three proprietary systems and one open-source model. Other artificial intelligence systems available today might behave differently.
- Controlled testing environment. Many tests relied on computer-generated artificial voices rather than a massive, diverse group of real human speakers. While human participants were also used, laboratory tests may not fully capture every nuance of real-world conversations.
Bottom line: Current artificial intelligence systems can show clear bias against certain accents even when they understand the words perfectly.
What we still don't know
The source does not explain how to completely fix the training data to remove these biases, nor does it cover how other types of models perform in medical emergency situations yet.
Based on reporting from University of Washington. This is an independent explainer, written in our own words with AI assistance; University of Washington has not reviewed or endorsed it. Read the original for the full details.