Artificial intelligence models replicate, and sometimes amplify, all sorts of human biases. In the last few years, Ģ����Ƶ researchers have found AI has erased female characters from kids’ stories, favored job applicants of certain races and genders, and swayed human decisions.��

Much bias research has focused on text and visuals. Yet people speak to AI models with increasing regularity, and AI tools are now used to screen job applicants and score oral presentations. Now, a Ģ����Ƶ-led team of researchers found that state-of-the-art AI models were biased against speakers with certain accents.��

The team investigated how state-of-the-art multimodal language models, including two from Google and one from OpenAI, evaluated human speech in job interviews, conference presentations and spoken English proficiency tests. The models scored responses from voices speaking English with five different accents: American, British, Chinese, Indian and Nigerian. The AI models generally ranked Chinese and Nigerian accents the lowest, Indian in the middle, and American and British the highest. Even when the content was exactly the same and the models could accurately transcribe the speech, the systems ranked both the delivery and content higher for the American and British speakers.��

The team Oct. 25 at the Conference on Empirical Methods in Natural Language Processing in Budapest, Hungary.��

“People are now trusting audio AI models as sort of third-party evaluators, for hiring, for oral exams in education, for certain English tests in immigration — contexts where fairness really matters,” said co-lead author , a Ģ����Ƶ doctoral student in the Paul G. Allen School of Computer Science & Engineering. “So I was curious how models perceive speech from people who don’t speak with so-called ‘standard’ American accents.”

The team chose accents from five regions that have large populations of English speakers and where the researchers also had connections. Researchers then tested the accents in three different domains — hiring, presentations and English tests. To study differences at scale, they had AI voices from read the same pieces of text to the AI models being tested. They validated the results with 52 human speakers from the five countries.��

After some testing, the team focused the study on three proprietary AI models  — Google’s Gemini 2.5 Pro and Gemini 2.5 Flash, and OpenAI’s GPT-Audio 1.5 — and one open-source model, Qwen 3 Omni.��

“We wanted to check these state-of-the-art models,” said senior author , a Ģ����Ƶ associate professor in the Allen School, “because these are what companies would deploy in a real-world setting.”

The team had the AI voices read various scripts — interview responses, a class presentation, a conference presentation and English proficiency test responses — with the five different accents in both male and female voices. The four AI models then rated the responses on a scale of one to seven for both delivery and content, and gave a brief explanation for their scores.��

Here are samples of AI voices reading the same hiring passage: 

British

Nigerian

Indian

American

Chinese

Researchers repeated the interview response test with the 52 human participants, having them read four scripted responses each and respond to questions without a script.��

The results favored American and British accents — with a greater disparity between the human accents than between the AI accents.

To better understand the results, researchers analyzed features of the accents. They then measured their “acoustic distance” and found that accents further from Americans’ were often ranked lower.��

The team found this apparent hierarchy manifested in the models’ explanations, too.��

“Even though they repeatedly claimed Chinese and Nigerian speakers ‘mumbled’ or were ‘difficult to understand,’ the models were actually very good at transcribing the accented speech,” Ahia said. “They understood the speakers, but they just had this implicit notion of hierarchy among them based on accent.”

This bias can seep into models during training in various ways — possibly through data that describes accented speakers as hard to understand, through data sets that underrepresent certain populations and accents, or through training methods that reinforce conventional expectations about how a professional speaker should sound. English, for instance, is the official language in Nigeria and is the teaching language in Indian schools, yet both accents were rated lower than British and American.

“One clear direction for future work is to create more representative training and evaluation data sets that could improve the accuracy and fairness of language models,” Tsvetkov said.��

The team plans to expand the research by studying more human accents — including regional variations within countries — and by looking at how AI models are analyzing audio in medical contexts, such as automated emergency response.��

“For all of these, we hope that developers build and publish evaluations of their models that don’t focus only on average performance, but also break it down into subpopulations, dialects and so on,” Tsvetkov said. “We don’t want the average performance to be 97% accurate for the general population, but for a smaller group only 20%.”

Additional co-authors include of Koel Labs, who completed this work as a Ģ����Ƶ undergraduate in the Allen School; of University of Wisconsin–Madison, who completed this research as a Ģ����Ƶ postdoctoral scholar in the Allen School; and of the University of Toronto; , a Ģ����Ƶ doctoral student in the Allen School; of the CISPA Helmholtz Center; of LMU Munich, who completed this research as a Ģ����Ƶ postdoctoral scholar in the Allen School; of Ohio State University; and , Ģ����Ƶ vice provost for AI and professor in the Allen School.��

This research was funded in part by gifts from Google and OpenAI.��

For more information, contact Ahia at oahia@cs.washington.edu and Tsvetkov at yuliats@cs.washington.edu.

Source

An estimated that between a third and two thirds of accounts on the social media site were bots. And many of these automatons flooding social media are dispatched to sow political polarization, hate, misinformation, propaganda and scams. The ability to sift them out of the online crowds is vital for a safer, more humane (or at least more human) internet.

But the recent proliferation of large language models (known as “LLMs”), such as OpenAI’s ChatGPT and Meta’s Llama, stands to complicate the world of social media bots.

A team led by Ģ����Ƶ researchers found that while operators can use customized LLMs to make bots more sophisticated at evading automated detectors, LLMs can also improve systems that detect bots. In the team’s tests, LLM-based bots reduced the performance of existing detectors by 30%. Yet researchers also found that an LLM trained specifically to detect social media bots outperformed state-of-the-art systems by 9%.

The team Aug. 11 at the 62nd Annual Meeting of the Association for Computational Linguistics in Bangkok.

“There’s always been an arms race between bot operators and the researchers trying to stop them,” said lead author , a doctoral student in the Paul G. Allen School of Computer Science & Engineering. “Each advance in bot detection is often met with an advance in bot sophistication, so we explored the opportunities and the risks that large language models present in this arms race.”

Researchers tested LLMs’ potential to detect bots in a few ways. When they fed Twitter data sets (culled before the platform became X) to off-the-shelf LLMs, including ChatGPT and Llama, the systems failed to accurately detect bots more than currently used technologies.

“Analyzing whether a user is a bot or not is much more complex than some of the tasks we’ve seen these general LLMs excel at, like recalling a fact or doing a grade-school math problem,” Feng said.

This complexity comes in part from the need to analyze three types of information for different attributes to detect a bot: the metadata (number of followers, geolocation, etc.), the text posted online and the network properties (such as what accounts a user is following).

When the team fine-tuned the LLMs with instructions on how to detect bots based on these three types of information, the models were able to detect bots with greater accuracy than current state-of-the-art systems.

The team also explored how LLMs might make bots more sophisticated and harder to detect. First the researchers simply gave LLMs prompts such as, “Please rewrite the description of this bot account to sound like a genuine user.”

They also tested more iterative, complicated approaches. In one test, the LLM would rewrite the bot post. The team then ran this through an existing bot-detection system, which would estimate the likelihood that a post was written by a bot. This process would be repeated as the LLM worked to lower that estimate. The team ran a similar test while removing and adding accounts that the bot followed to adjust its likelihood score.

These strategies, particularly rewriting the bots’ posts, reduced the effectiveness of the bot detection systems by as much as 30%. But the LLM-based detectors the team trained saw only a 2.3% drop in effectiveness on these manipulated posts, suggesting that the best way to detect LLM-powered bots might be with LLMs themselves.

“This work is only a scientific prototype,” said senior author , an associate professor in the Allen School. “We aren’t releasing these systems as tools anyone can download, because in addition to developing technology to defend against malicious bots, we are experimenting with threat modeling of how to create an evasive bot, which continues the cat-and-mouse game of building stronger bots that need stronger detectors.”

Researchers note that there are important limitations to using LLMs as bot detectors, such as the systems’ . They also highlight that the data used in the paper is from 2022, before Twitter effectively .

In the future, researchers want to look at bot detection beyond text, such as memes or videos on other platforms such as TikTok, where newer data sets are available. The team also wants to expand the research into other languages.

“Doing this research across different languages is extremely important,” Tsvetkov said. “We are seeing a lot of misinformation, manipulation and the targeting of specific populations as a result of various world conflicts.”

Additional co-authors on this paper are and , both undergraduates at Xi’an Jiaotong University; , an assistant professor at Xi’an Jiaotong University; and , a doctoral student at the University of Notre Dame. This research was funded by an NSF CAREER award.

For more information, contact Feng at shangbin@cs.washington.edu and Tsvetkov at yuliats@cs.washington.edu.

Source

Two faculty members at the Ģ����Ƶ have been awarded early-career fellowships from the Alfred P. Sloan Foundation. The new Sloan Fellows, announced Feb. 15, are , an assistant professor in the Department of Biology, and , an assistant professor in the Paul G. Allen School of Computer Science & Engineering.

Open to scholars in eight scientific and technical fields — chemistry, computer science, Earth system science, economics, mathematics, neuroscience and physics — the fellowships honor those early-career researchers whose achievements mark them among the next generation of scientific leaders.

The 118 were selected in coordination with the research community. Candidates are nominated by their peers, and fellows are selected by independent panels of senior scholars based on each candidate’s research accomplishments, creativity and potential to become a leader in their field. Each fellow will receive $75,000 to apply toward research endeavors.

This year’s fellows come from 51 institutions across the United States and Canada, spanning fields from evolutionary biology to data science.

Abrahms is an assistant professor of biology and holds the inaugural Boersma Endowed Chair in Natural History and Conservation. Her research program integrates animal bio-logging technology, Earth observation and big data analytics to advance understanding of the causes and consequences of wildlife responses to global change across marine and terrestrial systems. In addition to advancing basic ecological theory, Abrahms is passionate about developing data-driven, publicly available tools that bolster capacity to conserve the natural world.

“How do animals make decisions in the face of global change, and what are the consequences of those decisions for individual fitness, populations and interactions with other animals and humans? This is a big question my group is working to answer, which can inform both biodiversity conservation and human sustainability,” Abrahms said. “We’re becoming increasingly interested in understanding how species responses to environmental change can have unanticipated and often negative consequences for social-ecological systems so that we can reduce unwanted outcomes in the future.”

Tsvetkov is an assistant professor in the Allen School. She engages in multidisciplinary research at the nexus of machine learning, computational linguistics and the social sciences to develop practical solutions to natural language processing, or NLP, problems that combine sophisticated learning and modeling methods with insights into human languages and the people who speak them.

“The huge success of contemporary AI-powered NLP technology stems from the fact that it has matured enough to effectively serve and interact with humans. However, there is still a technological divide: The rich ecosystem of language-aware applications — machine translation, question answering, educational applications, summarization — are well-equipped to serve privileged users. But those same applications are systematically biased in ways that render them less useful for millions of other users, including speakers of low-resource languages or representatives of disadvantaged groups discriminated by gender, race, age or ethnicity,” Tsvetkov said. “The long-term goal of my research has been to bridge this divide, and to develop effective, accessible and equitable language technologies, serving all users, across populations, cultures and language boundaries.”

For more information, contact Abrahms at abrahms@uw.edu or Tsvetkov at yuliats@cs.washington.edu.

Source