Welcome, Cardinal readers, to another edition of Tech Briefs, our weekly batch of items covering the digital and life sciences landscapes. It goes live every Wednesday in Cardinal News.
Got tips, questions or suggestions? Let me know via tad@cardinalnews.org.
AI biases can extend to the place where you live
Can you imagine Virginia Tech with no Hokie stone? Artificial intelligence can.
A Virginia Tech professor recently asked an AI image generator to create a picture of the university’s hometown, Blacksburg.
The image was generic with none of the dolomite blocks that characterize the place, Junghwan Kim said in a university news release. Other landmarks were missing in a seemingly generic display.
“It didn’t capture what makes Blacksburg unique,” said Kim, a geospatial data scientist at the Virginia Tech College of Natural Resources and Environment.
He prompted the generator to make pictures of larger cities: Richmond, Virginia Beach and Washington, D.C. It produced images with familiar landmarks, waterfronts and other features.
Turns out, AI is biased toward large cities when it comes to making accurate location images, according to a study from Kim and colleagues at Virginia Tech, Hong Kong University of Science and Technology and the University of Alabama.
A study they published in the journal Technology in Society raises questions about how generative artificial intelligence tools portray places and whose communities are most visible online, according to the news release.
AI tools are growing more common in travel planning, urban design, marketing and public communication, so such representation gaps matter, Kim said.
“People are increasingly relying on AI-generated content to learn about places,” Kim said. “If smaller cities are not well represented in the data used to train these systems, then the images people see may not reflect the real identity of those communities.”
The team used OpenAI’s DALL·E 2 image generator to generate images of the cities mentioned above. Researchers then asked 129 survey subjects to evaluate how realistic and recognizable the images were.
AI struggled most with landmarks and culturally significant features. Long-time residents noticed the omissions more readily than newcomers, the researchers found, an indication of their stronger local knowledge.
As with many AI biases and so-called hallucinations, the less information the large language model has, the worse it will be at providing accurate information.
“AI systems learn from enormous amounts of online data,” Kim said in the news release. “Larger cities tend to have far more images, media coverage, and digital documentation available online. Smaller towns often do not have the same level of representation.”
Geographically comprehensive datasets and local perspectives are important in AI development, he said.
Kim said the research highlights the importance of building more geographically comprehensive datasets and incorporating local perspectives into AI development. The team’s work is part of a larger conversation about ethical use of AI in planning and design.
“Generative AI can be a powerful tool,” Kim said. “But we also need to understand where it falls short and who may be left out.”
Researchers modify classic combat video game to train Marines
Talk about your multiplayer shooter.
A research team led by a Virginia Tech professor has modified the video game “Call of Duty 4: Modern Warfare” for students in the Marine Corps University Sergeants School. Students are using the program to hone leadership, critical thinking, decision-making and communication in real-time scenarios, according to a Virginia Tech news release.
The platform is collecting player data and feeding it into a large language model for analysis on ways to improve after-action combat reviews.
Brig. Gen. Matthew Tracy, president of the Quantico-based Marine Corps University, gave the game a good review.
“The actual skills translate directly into making good decisions under high stress,” Tracy said in the news release.
The project’s principal investigator is Louis Hickman, assistant professor of industrial-organizational psychology at Virginia Tech. He recruited Ryan McMahan, director at the university’s human-computer interaction center, and Brandon Booth, an assistant professor in the University of Memphis’ computer science department.
They modified the “Call of Duty” game to include 14 scenarios specific to the Marines’ school. Further modifications included collecting player data that will be used in the project’s second phase, scaling up after-action reviews.
“The military considers after-action review one of the most important parts of training,” Hickman said in the news release. “After you’ve gone through a training simulation, you sit down as a group, you discuss what happened, what was supposed to happen, what went well, what didn’t, and what you will do differently next time.”
Also added: player telemetry collection. That means the researchers can “tap into player positions, orientations, where the enemies are, shots fired” and other in-game data, said McMahan, who is also a professor in the Department of Computer Science.
Hickman, a gamer, had the acumen to design the scenarios and was a part of weekly performance testing to make sure the game and modified software were coordinated.
“When Louis is there, it’s like having Rambo on your team, and when he’s not there, it’s like not having Rambo on your team,” McMahan said.
Marine Corps University plans to install the software at the Marine base in Twentynine Palms, California. Hickman said the students at the Sergeants School also support the idea of expansion.
“Many of them asked if there were ways to bring it back to their home unit because they thought it was so valuable that their units would benefit from it as well,” Hickman said in the news release.
