Even though he wanted to bike commute from his Capitol Hill home to the 天美影视传媒, Jared Hwang often took transit because he struggled to find a good bike route. Apps like Google Maps and Strava might suggest hilly, busy streets simply because they have bike lanes. He even headed to Reddit to crowdsource ideas.听

鈥淚 was like, surely, this cannot be the best way to do things,鈥� said , a UW doctoral student in the Paul G. Allen School of Computer Science & Engineering. 鈥淭his data is out there. We know where bike lanes are, what the roads are like, what the speed limits are. We should be able to easily access all this information at once.鈥�

So Hwang and a team of UW researchers built , a demo web app that lets users find personalized bike routes in Seattle. Cyclists plug in their origin and destination 鈥� just like in other mapping apps 鈥� and can then create personalized routes by adjusting eight sliders.听聽

Researchers initially worked with four participants to understand how cyclists tend to plan their routes. Based on that, they built a prototype of BikeButler. For the basic street layout and other info, they pulled data from OpenStreetMap and government data sets. But those didn鈥檛 have information on more subjective qualities.听

For those, researchers turned to Google Street View. They used a visual language model, or VLM 鈥� a type of artificial intelligence 鈥� to analyze street images and rate subjective attributes like greenery and pavement quality. The team had the VLM rate the level of greenery on streets and then compared this with two researchers鈥� ratings. The humans agreed with each other about as much as they agreed with the VLM 鈥� about 60% of the time. Future research might try to gather individual users鈥� greenery preferences to offset this discrepancy.听

Once they鈥檇 mapped most of Seattle, the team tested the prototype with 16 participants.听

鈥淥verall the response was really positive,鈥� Hwang said. 鈥淲e found that people do, in fact, have contextual preferences. A cyclist riding for fun on a Saturday might want a safer, greener route compared with their fast work commute. People intuitively know this, but it hadn鈥檛 been established through research.鈥澛�

Researchers say future work might integrate feedback from the user study, such as the ability to drag routes to change them slightly and an option to take fewer turns. The team is currently studying how to quantify cyclists鈥� preferences around intersections and turns.

The researchers note that the quality of BikeButler鈥檚 recommendations is constrained by the recency and accuracy of the data it uses. For instance, a new bike lane might not yet appear on a map, or it could appear in OpenStreetMap but not Google Street View. Also, since the team planned this as a proof of concept, BikeButler is limited to Seattle, though it could be expanded to other areas.听

鈥淚鈥檓 a lifelong biker and bike commuter,鈥� said senior author , a UW professor in the Allen School. 鈥淲hat excites me most about Jared鈥檚 work is how it points to a future where we receive route choices individualized to our preferences. So whether I鈥檓 biking with my two young children, or riding for groceries, I can find a route for that context.鈥�

Co-authors include , a student at Issaquah High School and intern in the Allen School; , a UW doctoral student in urban design and planning; and , a UW student in the Allen School. This study was supported by the National Science Foundation.

For more information, contact Hwang at jaredhwa@cs.washington.edu.

天美影视传媒 researchers developed the first system that incorporates tiny cameras in off-the-shelf wireless earbuds to allow users to talk with an AI model about the scene in front of them. For instance, a user might turn to a Korean food package and say, 鈥淗ey Vue, translate this for me.鈥� They鈥檇 then hear an AI voice say, 鈥淭he visible text translates to 鈥楥old Noodles鈥� in English.鈥�

The prototype system called VueBuds takes low-resolution, black-and-white images, which it transmits over Bluetooth to a phone or other nearby device. A small artificial intelligence model on the device then answers questions about the images within around a second. For privacy, all of the processing happens on the device, a small light turns on when the system is recording, and users can immediately delete images.听

The team will April 14 at the Association for Computing Machinery Conference on Human Factors in Computing Systems in Barcelona.听

鈥淲e haven鈥檛 seen most people adopt smart glasses or VR headsets, in part because a lot of people don鈥檛 like wearing glasses, and they often come with , such as recording high-resolution video and processing it in the cloud,鈥� said senior author , a UW professor in the Paul G. Allen School of Computer Science & Engineering. 鈥淏ut almost everyone wears earbuds already, so we wanted to see if we could put visual intelligence into tiny, low-power earbuds, and also address privacy concerns in the process.鈥�

Cameras use far more power than the microphones already in earbuds, so using the same sort of high-res cameras as those in smart glasses wouldn鈥檛 work. Also, large amounts of information can鈥檛 stream continuously over Bluetooth, so the system can鈥檛 run continuous video.听

The team found that using a low-power camera 鈥� roughly the size of a grain of rice 鈥� to shoot low-resolution, black-and-white still images limited battery drain and allowed for Bluetooth transmission while preserving performance.

There was also the matter of placement.听

鈥淥ne big question we had was: Will your face obscure the view too much? Can earbud cameras capture the user鈥檚 view of the world reliably?鈥� said lead author , who completed this work as a UW doctoral student in the Allen School.听

The team found that angling each camera 5-10 degrees outward provides a 98-108 degree field of view. While this creates a small blind spot when objects are held closer than 20 centimeters from the user, people rarely hold things that close to examine them 鈥� making it a non-issue for typical interactions.

Sixteen participants also wore VueBuds and tested the system鈥檚 ability to translate and answer basic questions about objects. VueBuds achieved 83-84% accuracy when translating or identifying objects and 93% when identifying the author and title of a book.

This study was designed to gauge the feasibility of integrating cameras in wireless earbuds. Since the system only takes grayscale images, it can鈥檛 answer questions that involve color in the scene.听

The team wants to add color to the system 鈥� color cameras require more power 鈥� and to train specialized AI models for specific use cases, such as translation.听聽

鈥淭his study lets us glimpse what鈥檚 possible just using a general purpose language model and our wireless earbuds with cameras,鈥� Kim said. 鈥淏ut we鈥檇 like to study the system more rigorously for applications like reading a book 鈥� for people who have low vision or are blind, for instance 鈥� or translating text for travelers.鈥澛�

Co-authors include , a UW master鈥檚 student in the Allen School, and , , , and , all UW students in electrical and computer engineering.听

For more information, contact vuebuds@cs.washington.edu.

Heart rate is an important sign of fetal health, yet few technologies exist to easily and inexpensively track fetal heart rates outside of doctors鈥� offices. This can create risks for pregnancies in low-resource regions where doctors are far away or inaccessible.听

A team led by 天美影视传媒 researchers has created DopFone, a system that uses an off-the-shelf smartphone鈥檚 existing speaker and microphone to accurately estimate fetal heart rate. The phone mimics a Doppler ultrasound, emitting a tone and listening for the subtle variations in its echo caused by fetal heart beats. A machine learning model then estimates the heart rate. In a clinical test with 23 pregnant women, DopFone estimated heart rate with an average error of 2 beats per minute, or bpm. The accepted clinical range is within 8 bpm.听

The team Dec. 2 in the Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies.听

鈥淓ventually DopFone could let people test fetal heart rate regularly, rather than relying on the intermittent tests at a doctor鈥檚 office, or not getting tested at all,鈥� said lead author , a UW doctoral student in the Paul G. Allen School of Computer Science & Engineering. 鈥淧atients might then send this data to doctors so that they can better judge patients鈥� health when they鈥檙e not in a clinic.鈥�

Traditional Doppler ultrasounds, the clinical standard for fetal heart rate monitoring, work by sending high-frequency sound into a person鈥檚 body and tracking how the echo changes in frequency. They鈥檙e very accurate at measuring fetal heart rate but require costly equipment and a skilled technician to operate it.

To use DopFone, a user places the phone鈥檚 microphone against their abdomen for one minute. The phone emits a subaudible 18 kilohertz tone. The team chose this low frequency because 鈥� unlike a Doppler鈥檚 high frequencies, above 2,000 kilohertz 鈥斅� it sits within the range smartphone microphones can record while still traveling well through tissue. As the tone is reflected through the user鈥檚 abdomen, the fetus鈥檚 heartbeat creates small shifts in the sound.听

A machine learning model then estimates the heart rate using the audio and the patient鈥檚 demographic information

The team tested DopFone in UW Medicine鈥檚 maternal-fetal medicine division on 23 pregnant patients between 19 and 39 weeks of pregnancy. On average its readings were within 2.1 bpm of the medical Doppler ultrasound. Its accuracy was slightly diminished for patients with high body mass indexes, though those readings were still within normal limits. Because an irregular fetal heartbeat is often an emergency, DopFone was not tested on patients with irregularities.听

Next, the team plans to gather more data outside a lab to better train the model. Eventually they want to deploy it as a publicly available app.

鈥淭his women鈥檚 health space is often overlooked,鈥� Garg said. 鈥淪o I want to focus on accessible alternatives that can be available to people in low resource areas, whether that鈥檚 here in the U.S. or in other countries. Because health belongs to everyone.鈥�

Co-authors include , a UW graduate student in electrical and computer engineering; and , both OB/GYNs in UW Medicine鈥檚聽 maternal-fetal medicine division; and , a UW assistant professor in the Allen School. , a UW professor in the Allen School and in electrical and computer engineering, and of the Georgia Institute of Technology, were senior authors. This research was funded by the UW Gift Fund.听

For more information, contact Garg at pgarg70@uw.edu.

Keeping up with the latest research is vital for scientists, but given that are published every year, that can prove difficult. Artificial intelligence systems show promise for quickly synthesizing seas of information, but they still tend to make things up, or 鈥渉allucinate.鈥澛�

For instance, when a team led by researchers at the 天美影视传媒 and , or Ai2, studied a recent OpenAI model, , they found it fabricated 78-90% of its research citations. And general-purpose AI models like ChatGPT often can鈥檛 access papers that were published after their training data was collected.听

So the UW and Ai2 research team built OpenScholar, an open-source AI model designed specifically to synthesize current scientific research. The team also created the first large, multi-domain for evaluating how well models can synthesize and cite scientific research. In tests, OpenScholar cited sources as accurately as human experts, and 16 scientists preferred its response to those written by subject experts 51% of the time.听

The team Feb. 4 in Nature. The project鈥檚 are publicly available and free to use.

鈥淎fter we started this work, we put the demo online and quickly, we got a lot of queries, far more than we鈥檇 expected,鈥� said senior author , a UW associate professor in the Paul G. Allen School of Computer Science & Engineering and senior director at Ai2. 鈥淲hen we started looking through the responses we realized our colleagues and other scientists were actively using OpenScholar. It really speaks to the need for this sort of open-source, transparent system that can synthesize research.鈥�

Researchers trained the model and then created a set of 45 million scientific papers for OpenScholar to pull from to ground its answers in established research. They coupled this with a technique called “,鈥� which lets the model search for new sources, incorporate them and cite them after it鈥檚 been trained.听

鈥淓arly on we experimented with using an AI model with Google鈥檚 search data, but we found it wasn鈥檛 very good on its own,鈥� said lead author , a research scientist at Ai2 who completed this research as a UW doctoral student in the Allen School. 鈥淚t might cite some research papers that weren鈥檛 the most relevant, or cite just one paper, or pull from a blog post randomly. We realized we needed to ground this in scientific papers. We then made the system flexible so that it could incorporate emerging research through results.鈥澛�

To test their system, the team created ScholarQABench, a benchmark against which to test systems on scientific search. They gathered 3,000 queries and 250 longform answers written by experts in computer science, physics, biomedicine and neuroscience.听

鈥淎I is getting better and better at real world tasks,鈥� Hajishirzi said. 鈥淏ut the big question ultimately is whether we can trust that its answers are correct.鈥�

The team compared OpenScholar against other state-of-the-art AI models, such as OpenAI鈥檚 GPT-4o and two models from Meta. ScholarQABench automatically evaluated AI models鈥� answers on metrics such as their accuracy, writing quality and relevance.听

OpenScholar outperformed all the systems it was tested against. The team had 16 scientists review answers from the models and compare them with human-written responses. The scientists preferred OpenScholar answers to human answers 51% of the time, but when they combined OpenScholar citation methods and pipelines with GPT-4o (a much bigger model), the scientists preferred the AI written answers to human answers 70% of the time. They picked answers from GPT-4o on its own only 32% of the time.

鈥淪cientists see so many papers coming out every day that it鈥檚 impossible to keep up,鈥� Asai said. 鈥淏ut the existing AI systems weren鈥檛 designed for scientists鈥� specific needs. We鈥檝e already seen a lot of scientists using OpenScholar and because it鈥檚 open-source, others are building on this research and already improving on our results. We鈥檙e working on a followup model, , which builds on OpenScholar鈥檚 findings and performs multi-step search and information gathering to produce more comprehensive responses.鈥澛�

Other co-authors include , , , all UW doctoral students in the Allen School; , a UW professor emeritus in the Allen School and general manager and chief scientist at Ai2; , a UW postdoc in the Allen School and postdoc at Ai2; , a UW professor in the Allen School; , a UW assistant professor in

the Allen School; Amanpreet Singh, Joseph Chee Chang, Kyle Lo, Luca Soldaini, Sergey Feldman, Mike D鈥橝rcy, David Wadden, Matt Latzke, Jenna Sparks and Jena D. Hwang of Ai2; Wen-tau Yih of Meta; Minyang Tian, Shengyan Liu, Hao Tong and Bohao Wu of University of Illinois Urbana-Champaign; Pan Ji of University of North Carolina; Yanyu Xiong of Stanford University; and Graham Neubig of Carnegie Mellon University.

For more information, contact Asai at akaria@allenai.org and Hajishirzi at hannaneh@cs.washington.edu.

Seattle Public Library, or SPL, is the only U.S. library system that makes its anonymized, granular checkout data public. Want to find out how many times people borrowed the e-book version of Toni Morrison鈥檚 鈥淏eloved鈥� in May 2018? That data is available.听

The hitch is that the library鈥檚 data set contains nearly 50 million rows, and a single title can appear variously. Morrison鈥檚 鈥淏eloved,鈥� for instance, is listed as 鈥淏eloved,鈥� 鈥淏eloved (unabridged),鈥� 鈥淏eloved : a novel / by Toni Morrison鈥� and so on.听

To track trends in the catalogue over the last 20 years, 天美影视传媒 researchers analyzed the checkout data of the 93 authors included in the post-1945 volume of 鈥淭he Norton Anthology of American Literature.鈥� It鈥檚 assigned in U.S. English classes more than virtually any other anthology, so what鈥檚 thought of as the contemporary American 鈥� the books and writers we鈥檝e deemed culturally important.听

The team found that among these vaunted writers 鈥� including Morrison, Viet Thanh Nguyen, David Foster Wallace and Joan Didion 鈥� science fiction was particularly popular. Ursula K. Le Guin and Octavia E. Butler topped the list.听

The team Nov. 21 in Computational Humanities Research 2025, and created .听

鈥淚t鈥檚 kind of mind-boggling and ironic that in this age of abundant data, we have so little data about what people are reading,鈥� said senior author , a UW assistant professor in the Information School. 鈥�, particularly for researchers, so I鈥檝e been obsessed with SPL鈥檚 data for years now. But extracting insights from it is actually a really hard computational and bibliographic modeling problem.鈥�

To organize the data, the team used computational methods, such as stripping away subtitles and standardizing punctuation. They also manually identified things like translations of a work.听

鈥淲e worked with the Norton anthology in part because it’s a small enough scale for us to handle,鈥� said lead author , a UW doctoral student in the Information School. 鈥淚t allows us to have a ground truth to work off of. We can still put a human eye on things.鈥澛�

In all the team looked at 1,603 works by the 93 authors, which were checked out a total of 980,620 times since 2005.

The 10 top authors were:

1. Ursula K. Le Guin
2. Octavia E. Butler
3. Louise Erdrich
4. N.K. Jemisin
5. Toni Morrison
6. Kurt Vonnegut
7. George Saunders
8. Philip K. Dick
9. Sherman Alexie
10. James Baldwin

The 10 top books were:聽

1. 鈥淧arable of the Sower鈥� by Octavia E. Butler
2. 鈥淟incoln in the Bardo鈥� by George Saunders
3. 鈥淭he Fifth Season鈥� by N.K. Jemisin
4. 鈥淭he Sympathizer鈥� by Viet Thanh Nguyen
5. 鈥淜indred鈥� by Octavia E. Butler
6. 鈥淏eloved鈥� by Toni Morrison
7. 鈥淭he Left Hand of Darkness鈥� by Ursula K. Le Guin
8. 鈥淭he Absolutely True Diary of a Part-Time Indian鈥� by Sherman Alexie
9. 鈥淭he Year of Magical Thinking鈥� by Joan Didion
10. 鈥淭he Sentence鈥� by Louise Erdrich

Researchers noted several trends that may have driven checkouts. In general, books with genre and sci-fi elements were some of the most popular.听

鈥淚 found the prevalence of sci-fi books and writers really interesting,鈥� Gupta said. 鈥淭hese are recent additions to the anthology, since sci-fi and genre fiction haven鈥檛 always been seen as important literature. So while it鈥檚 a bit unsurprising, it鈥檚 also striking to see that despite comprising a small portion of the anthology, these are the authors people are actually reading the most.鈥�

News events also drove spikes in readership, such as film adaptations of James Baldwin鈥檚 鈥淚f Beale Street Could Talk鈥� and Don DeLillo鈥檚 鈥淲hite Noise,鈥� or the deaths of authors such as Didion, Wallace, Morrison and Philip Roth.听

The top book, 鈥淧arable of the Sower,鈥� saw a huge spike in readership in 2024 鈥� the year the futuristic novel is set, and the year SPL selected the novel for its program.听

鈥淲e鈥檝e deemed these canonical authors important enough to continue reading, to continue teaching, to continue studying and talking about, so it鈥檚 fascinating to see who we鈥檙e actually reading and when,鈥� Walsh said. 鈥淚 find it very beautiful that after years of these big debates about diversifying the canon, the works that people are turning to the most are by women and Black and Native writers, who previously were not even included in these anthologies.鈥�

Co-authors include Daniella Maor, Karalee Harris, Emily Backstrom and Hongyuan Dong, all students at the UW. This research was supported in part by the .

For more information, contact Walsh at melwalsh@uw.edu and Gupta at ngupta1@uw.edu.

Once the domain of buttons and knobs, car dashboards are increasingly home to large touch screens. While that makes following a mapping app easier, it also means drivers can鈥檛 feel their way to a control; they have to look. But how does that visual component affect driving?

New research from the 天美影视传媒 and Toyota Research Institute, or TRI, explores how drivers balance driving and using touch screens while distracted. In the study, participants drove in a vehicle simulator, interacted with a touch screen and completed memory tests that mimic the mental effort demanded by traffic conditions and other distractions. The team found that when people multitasked, their driving and touch screen use both suffered. The car drifted more in the lane while people used touch screens, and their speed and accuracy with the screen declined when driving. The effects increased further when they added the memory task.听

These results could help auto manufacturers design safer, more responsive touch screens and in-car interfaces.

The team Sept. 30 at the ACM Symposium on User Interface Software and Technology in Busan, Korea.听

鈥淲e all know ,鈥� said co-senior author , a UW professor in the Paul G. Allen School of Computer Science & Engineering. 鈥淏ut what about the car鈥檚 touch screen? We wanted to understand that interaction so we can design interfaces specifically for drivers.鈥�

As the study鈥檚 16 participants drove the simulator, sensors tracked their gaze, finger movements, pupil diameter and electrodermal activity. The last two are common ways to measure mental effort, or 鈥渃ognitive load.鈥� For instance, pupils tend to grow when people are concentrating.听

While driving, participants had to touch specific targets on a 12-inch touch screen, similar to how they would interact with apps and widgets. They did this while completing three levels of an 鈥淣-back task,鈥� a memory test in which the participants hear a series of numbers, 2.5 seconds apart, and have to repeat specific digits.听

The participants鈥� performance changed significantly under different conditions:

• When interacting with the touch screen, participants drifted side to side in their lane 42% more often. Increasing cognitive load had no effect on the results.
• Touch screen accuracy and speed decreased 58% when driving, then another 17% under high cognitive load.
• Each glance at the touchscreen was 26.3% shorter under high cognitive load.
• A 鈥渉and-before-eye鈥� phenomenon, in which drivers鈥� reached for a control before looking at it, increased from 63% to 71% as memory tasks were introduced.

The team also found that increasing the size of the target areas participants were trying to touch did not improve their performance.听

鈥淚f people struggle with accuracy on a screen, usually you want to make bigger buttons,鈥� said , a UW doctoral student in the Allen School. 鈥淏ut in this case, since people move their hand to the screen before touching, the thing that takes time is the visual search.鈥�

Based on these findings, the researchers suggest future in-car touch screen systems might use simple sensors in the car 鈥� eye tracking, or touch sensors on the steering wheel 鈥� to monitor drivers鈥� attention and cognitive load. Based on these readings, the car鈥檚 system might adjust the touch screen鈥檚 interface to make important controls more prominent and safer to access.

鈥淭ouch screens are widespread today in automobile dashboards, so it is vital to understand how interacting with touch screens affects drivers and driving,鈥� said co-senior author , a UW professor in the Information School. 鈥淥ur research is some of the first that scientifically examines this issue, suggesting ways for making these interfaces safer and more effective.鈥�

, a UW doctoral student in the Information School, is co-lead author. Other co-authors include , , and of TRI. This research was funded in part by TRI.

For more information, contact Wobbrock at wobbrock@uw.edu and Fogarty at jfogarty@cs.washington.edu.

Artificial intelligence systems absorb values from their training data. The trouble is that values differ across cultures. So an AI system trained on data from the entire internet won鈥檛 work equally well for people from different cultures.

But a new 天美影视传媒 study suggests that AI could learn cultural values by observing human behavior. Researchers had AI systems observe people from two cultural groups playing a video game. On average, participants in one group behaved more altruistically. The AI assigned to each group learned that group鈥檚 degree of altruism, and was able to apply that value to a novel scenario beyond the one they were trained on.

The team Dec. 9 in PLOS One.听

鈥淲e shouldn鈥檛 hard code a universal set of values into AI systems, because many cultures have their own values,鈥� said senior author , a UW professor in the Paul G. Allen School of Computer Science & Engineering and co-director of the Center for Neurotechnology. 鈥淪o we wanted to find out if an AI system can learn values the way children do, by observing people in their culture and absorbing their values.鈥�

As inspiration, the team looked to showing that 19-month-old children raised in Latino and Asian households were more than those from other cultures.听

In the AI study, the team recruited 190 adults who identified as white and 110 who identified as Latino. Each group was assigned an AI agent, a system that can function autonomously.听

These agents were trained with a method called inverse reinforcement learning, or IRL. In the more common AI training method, reinforcement learning, or RL, a system is given a goal and gets rewarded based on how well it works toward that goal. In IRL, the AI system observes the behavior of a human or another AI agent, and infers the goal and underlying rewards. So a robot trained to play tennis with RL would be rewarded when it scores points, while a robot trained with IRL would watch professionals playing tennis and learn to emulate them by inferring goals such as scoring points.

This IRL approach more closely aligns with how humans develop.听

鈥淧arents don鈥檛 simply train children to do a specific task over and over. Rather, they model or act in the general way they want their children to act. For example, they model sharing and caring towards others,鈥� said co-author , a UW professor of psychology and co-director of Institute for Learning & Brain Sciences (I-LABS). 鈥淜ids learn almost by osmosis how people act in a community or culture. The human values they learn are more 鈥榗aught鈥� than 鈥榯aught.鈥欌��

In the study, the AI agents were given the data of the participants playing a modified version of the video game Overcooked, in which players work to cook and deliver as much onion soup as possible. Players could see into another kitchen where a second player had to walk further to accomplish the same tasks, putting them at an obvious disadvantage. Participants didn鈥檛 know that the second player was a bot programmed to ask the human players for help. Participants could choose to give away onions to help the bot but at the personal cost of delivering less soup.听

Researchers found that overall the people in the Latino group chose to help more than those in the white group, and the AI agents learned the altruistic values of the group they were trained on. When playing the game, the agent trained on Latino data gave away more onions than the other agent.听

To see if the AI agents had learned a general set of values for altruism, the team conducted a second experiment. In a separate scenario, the agents had to decide whether to donate a portion of their money to someone in need. Again, the agents trained on Latino data from Overcooked were more altruistic.听

鈥淲e think that our proof-of-concept demonstrations would scale as you increase the amount and variety of culture-specific data you feed to the AI agent. Using such an approach, an AI company could potentially fine-tune their model to learn a specific culture鈥檚 values before deploying their AI system in that culture,鈥� Rao said.听

Additional research is needed to know how this type of IRL training would perform in real-world scenarios, with more cultural groups, competing sets of values, and more complicated problems.

鈥淐reating culturally attuned AI is an essential question for society,鈥� Meltzoff said. 鈥淗ow do we create systems that can take the perspectives of others into account and become civic minded?鈥�

, a UW research engineer in the Allen School, and , a software engineer at Microsoft who completed this research as a UW student, were co-lead authors. Other co-authors include , a scientist at the Allen Institute who completed this research as a UW doctoral student; , an assistant professor at San Diego State University, who completed this research as a post-doctoral scholar at UW; and , a professor in the Allen School and director of the at UW.听

For more information, contact Rao at rao@cs.washington.edu.

An organization drafts a job listing with artificial intelligence. Droves of with chatbots. Another AI system sifts through those applications, passing recommendations to hiring managers. Perhaps AI avatars conduct screening interviews. This is increasingly the state of hiring, as people seek to streamline the stressful, tedious process with AI.

Yet research is finding that hiring bias 鈥� against people with disabilities, or certain races and genders 鈥� permeates large language models, or LLMs, such as ChatGPT and Gemini. We know less, though, about how biased LLM recommendations influence the people making hiring decisions.听

In a new 天美影视传媒 study, 528 people worked with simulated LLMs to pick candidates for 16 different jobs, from computer systems analyst to nurse practitioner to housekeeper. The researchers simulated different levels of racial biases in LLM recommendations for resumes from equally qualified white, Black, Hispanic and Asian men.听

When picking candidates without AI or with neutral AI, participants picked white and non-white applicants at equal rates. But when they worked with a moderately biased AI, if the AI preferred non-white candidates, participants did too. If it preferred white candidates, participants did too. In cases of severe bias, people made only slightly less biased decisions than the recommendations.

The team Oct. 22 at the AAAI/ACM Conference on Artificial Intelligence, Ethics, and Society in Madrid.听

鈥淚n one survey, 80% of organizations using AI hiring tools said they don鈥檛 reject applicants without human review,鈥� said lead author , a UW doctoral student in the Information School. 鈥淪o this human-AI interaction is the dominant model right now. Our goal was to take a critical look at this model and see how human reviewers鈥� decisions are being affected. Our findings were stark: Unless bias is obvious, people were perfectly willing to accept the AI鈥檚 biases.鈥�

The team recruited 528 online participants from the U.S. through surveying platform , who were then asked to screen job applicants. They were given a job description and the names and resumes of five candidates: two white men and two men who were either Asian, Black or Hispanic. These four were equally qualified. To obscure the purpose of the study, the final candidate was of a race not being compared and lacked qualifications for the job. Candidates鈥� names implied their races 鈥� for example, Gary O鈥橞rien for a white candidate. Affinity groups, such as Asian Student Union Treasurer, also signaled race.

In four trials, the participants picked three of the five candidates to interview. In the first trial, the AI provided no recommendation. In the next trials, the AI recommendations were neutral (one candidate of each race), severely biased (candidates from only one race), or moderately biased, meaning candidates were recommended at rates similar to rates of bias in real AI models. The team derived rates of moderate bias using the same methods as in their 2024 study that looked at bias in three common AI systems.听

Rather than having participants interact directly with the AI system, the team simulated the AI interactions so they could hew to rates of bias from their large-scale study. Researchers also used AI generated resumes, rather than real resumes, which they validated. This allowed greater control, and AI-written resumes are increasingly common in hiring.

鈥淕etting access to real-world hiring data is almost impossible, given the sensitivity and privacy concerns,鈥� said senior author , a UW associate professor in the Information School. 鈥淏ut this lab experiment allowed us to carefully control the study and learn new things about bias in human-AI interaction.鈥�

Without suggestions, participants鈥� choices exhibited little bias. But when provided with recommendations, participants mirrored the AI. In the case of severe bias, choices followed the AI picks around 90% of the time, rather than nearly all the time, indicating that even if people are able to recognize AI bias, that awareness isn鈥檛 strong enough to negate it.

鈥淭here is a bright side here,鈥� Wilson said. 鈥淚f we can tune these models appropriately, then it’s more likely that people are going to make unbiased decisions themselves. Our work highlights a few possible paths forward.鈥�

In the study, bias dropped 13% when participants began with an , intended to detect subconscious bias. So companies including such tests in hiring trainings may mitigate biases. Educating people about AI can also improve awareness of its limitations.

鈥淧eople have agency, and that has huge impact and consequences, and we shouldn’t lose our critical thinking abilities when interacting with AI,鈥� Caliskan said. 鈥淏ut I don鈥檛 want to place all the responsibility on people using AI. The scientists building these systems know the risks and need to work to reduce systems鈥� biases. And we need policy, obviously, so that models can be aligned with societal and organizational values.鈥�

, a UW doctoral student in the Information School, and , a postdoctoral scholar at Indiana University, are also co-authors on this paper. This research was funded by The U.S. National Institute of Standards and Technology.

For more information, contact Wilson at kywi@uw.edu and Caliskan at aylin@uw.edu.

Even though video games have grown as an artistic medium , they are still often written off as mindless entertainment. Research is increasingly exploring meaningful gaming experiences. Less studied, though, are the ways such experiences can alter people鈥檚聽 lives long term.听

In a new study, 天美影视传媒 researchers surveyed gamers about video games鈥� effects. Of 166 respondents researchers asked about meaningful experiences, 78% said such experiences had altered their lives. Researchers then pulled recurring themes from the responses 鈥� such as the power of聽 rich storytelling 鈥� so that developers, gamers and even parents or teachers might focus on those elements.听

The team will Oct. 14 at the Annual Symposium on Computer-Human Interaction in Play in Pittsburgh.听

To learn more about the paper, UW News spoke with lead author , a UW doctoral student in human centered design and engineering; co-senior author , a UW professor and chair in human centered design and engineering; and co-senior author , a UW professor in the Information School.听

What are the most significant findings in the study?

Nisha Devasia: We highlighted three conclusions drawn from modeling the data. The first is that playing games during stressful times was strongly correlated with positive outcomes for physical and mental health. For example, during COVID, people played聽 games they felt strongly improved their mental health, such as Stardew Valley. Others mentioned that games that required movement, or games that had characters with interesting physical abilities, inspired them to get outside or try new sports. Many participants also said that they gained a lot of insight from the game narrative. Story-based games often tell a sort of hero’s journey, for instance. People reported that the insight they gained from those stories correlated to their own self-reflection and identity building.

Finally, most people had these meaningful experiences in very early adulthood or younger, when they’re still trying to figure out who they are and what they want to be in the world. Playing as a character and seeing your choices change the course of events is pretty unique to games, compared with other narrative media like novels or movies.

Do any individual stories really stand out to you from the survey you took?

ND: All the stories about Final Fantasy VII, because that’s the game that I love. I鈥檓 actually sitting in my childhood bedroom right now and the wall behind me is covered in Final Fantasy VII posters. The quote we used in the title also really resonated with me: 鈥淚 would not be this version of myself today without these experiences.鈥� I definitely cannot imagine what I would be doing in my life if I had not played Final Fantasy VII when I did.听

People also said things like, 鈥淭his helped me build the skills that ended up being my career. I learned how to program because I wanted to make games.鈥� I worked in the gaming industry and can verify that鈥檚 true for many people in the industry.听

How should these findings fit into how we view games as a society?

Julie Kientz: People have a tendency to treat technology as a monolith, as if video games are either good or bad, but there’s so much more nuance. The design matters. This study hopefully helps us untangle the positive elements. Certainly, there are bad elements 鈥� toxicity and addictiveness, for example. But we also see opportunities for growth and connection. Some people in the study met their spouses through games.

Jin Ha Lee: What Nisha studies is essentially what I live. I鈥檓 a gamer, and I have definitely started playing certain games with my two children specifically because I wanted to have more conversations with them. When my daughter plays games with interesting stories, we have the opportunity to talk about our lives as we analyze the story. What were these people thinking? Why did they make certain decisions?聽

As researchers, we develop games for learning, for instance, for teaching people about misinformation or AI, or promote digital civic engagement, because we want to foster meaningful experiences. But a lot of the existing research just focuses on the short-term effects of games. This study really helps us understand what actually caused a game to make a difference in someone鈥檚 life.

What societal changes could we make in our approach to gaming?

JK: Because people have a tendency to oversimplify things, some of the proposed solutions can be counterproductive. For instance, limiting kids鈥� screen time can actually interfere with positive experiences, especially if someone is immersed in the storyline and identifies with the characters. If 30 minutes into a game, a kid鈥檚 Nintendo Switch turns off because of parental controls, that might hinder the ability to have a positive experience. If we aren鈥檛 using these tools consciously, it might actually lead to kids playing more casual, junk games, because those can be played in 30 minutes.

ND: You see this with discourse around game addiction, too. Sometimes excessive gaming is because of dark patterns in a game鈥檚 design. But it is often a symptom of someone going through something difficult in their life, and the game happens to be a way to cope. As our study shows, there鈥檚 the potential for growth in that coping.听

JHL: There鈥檚 also a place for games and media that we consider 鈥渂ad.鈥� You might play a game that鈥檚 so horrible that you make a meme out of it, and the jokes you share become a way to build community. Online communities can grow into offline events and friendships. But that isn鈥檛 necessarily obvious if you just view gaming as something you need to protect your children from.

What technological changes might accentuate the meaningful effects of games?

JHL: Games are naturally interactive and complex, so there鈥檚 a lot of opportunity for critical engagement beyond just the gameplay. There鈥檚 music, there鈥檚 art, there鈥檚 storytelling. All of these offer space for meaningful interaction. Designers can skillfully incorporate these elements to prompt reflection, evoke emotions, or challenge players鈥� perspectives.听

ND: We鈥檙e calling our next study Video Game Book Club. Right now I’m building a tool to allow people to annotate their gameplay as if they were writing in the margins of a book. While you play, a little pop-up lets you make a note. At the end, an interface pops up showing your gameplay stream and all the notes you made, which should allow them to reflect on what they were thinking as they were playing.

We鈥檙e also working on a reflection chatbot. Every time after you play a session that’s 30 minutes to an hour long, you’ll interact with this bot that prompts you to think critically about the experience, much like we鈥檙e taught to relate to literature. What was really memorable? How is this connected to your life?聽

Co-authors include , a UW doctoral student in human centered design and engineering, and , a UW doctoral student in the Information School. This research was funded by the .听

For more information, contact Devasia at ndevasia@uw.edu, Kientz at jkientz@uw.edu and Lee at jinhalee@uw.edu.

While innovation is core to American identity, women inventors were named on only 13% of 2019 U.S. patents. In part, that鈥檚 because .听

Research by the 天美影视传媒 and the United States Patent and Trademark Office, or USPTO, found that some simple interventions increased the probability that female inventors would get patents by 12%. For first-time applicants, that probability increased to 17%. The study, the first randomized controlled trial of inventors at the USPTO, followed inventors who applied 鈥減ro se,鈥� meaning without the help of a lawyer. Researchers randomly assigned some senior patent examiners to provide extra help and encouragement navigating the complicated examination process.听

The paper was published in the of the American Economic Journal: Economic Policy.听

The study began in 2014, when USPTO created a unit to help pro se inventors through the patent process. The office selected 15 senior patent examiners, who received 20 hours of training on strategies to better assist pro se inventors. In the span of a year, 2,273 applications were divided between the treatment and control arms. Of those applications,聽16% had more than half women inventors.

In the treatment arm, examiners used more encouraging language and gave more detailed responses in their first written decisions. They also prompted the applicants to call for an interview about the decision. Interviews increased 25% for both genders, but majority-women teams were 8% more likely to work out specific changes in those interviews.听

鈥淭his was a very effective, fairly low-cost program,鈥� said author , a UW assistant professor in the Information School. 鈥淭here鈥檚 this ideal of the garage inventor tinkering with something, coming up with an idea to start a company. That group of people usually doesn’t have access to lawyers, so they apply as individuals. This intervention helped more people find success.鈥澛�

A full list of co-authors is included with the .

For more information, contact Teodorescu at miketeod@uw.edu.