Unraveling the Hidden Biases of Artificial Intelligence
For years, Boston University computer scientist and CISE faculty affiliate Professor Mark Crovella (CDS, CS, ECE, SE) has studied the invisible forces shaping what we see online.
The stakes are high: when YouTube’s recommendation system steers users toward more extreme videos, it can fuel polarization and misinformation. Now, with large language models generating instant answers, a similar risk emerges — these systems can reproduce or even amplify hidden biases in their data, subtly shaping opinions and decisions at massive scale.
As biases and misinformation in AI systems raise urgent concerns, Crovella is at the forefront of efforts to uncover the hidden mechanics behind these powerful technologies, work that could shape the future of the digital world.
An early interest for Crovella was recommendation systems, which are AI-powered software that use machine learning and data analysis to provide personalized suggestions for products, services, and content for users. These softwares,used in platforms like Amazon, Spotify, and Netflix, are susceptible to providing biased recommendations.
“I started to be aware of how impactful these things were in people’s lives,” Crovella said. “I started looking at trying to understand the impacts of recommendation systems, specifically on people’s tendency to be exposed to extreme content.”
Crovella’s focus soon transitioned to large language models (LLMs) and their feedback loops.
He approached the problem using the same process he used to evaluate recommendation systems: feed the system an input, check the result it produces, and then judge whether that result shows bias or not. However, it became clear that just looking at inputs and outputs of an LLM was insufficient, and that it was necessary to look at the model’s internals — a field called “interpretability”.
This interpretability research — looking at what is inside these LLMs, the computations that are taking place inside them — and turning them into descriptions that humans can understand earned Crovella a spot as one of 35 projects announced as a National Artificial Intelligence Research Resource (NAIRR) Pilot Project.
Much of the cutting-edge work on LLMs has been concentrated in industry, where private companies like OpenAI, Google, and Anthropic control the data and computing power needed to train and probe these massive systems. Academia, by contrast, has historically faced steep barriers, including the costs of accessing training data and high-performance hardware.
NAIRR aims to rebalance that equation by giving researchers like Crovella the tools and infrastructure they need to study these models independently.
By advancing interpretability research in an open, academic setting, Crovella’s work has the potential to make LLMs more transparent, reveal biases or flaws, and ultimately inform public policy on how AI should be governed.
The goal of Crovella’s research is to identify any biases within LLMs, like OpenAI’s GPT series. Once biases are identified, the next step is to measure their impact and trace where they come from, all while testing the models for objectivity and sharing findings to guide responsible use.
“We’re not there yet, but the idea would be to be able to say that the model is actually manipulating some protected category, for example looking at making a decision based on the race of the person involved, or the gender of the person involved,” Crovella said. “And that would be explicit in the model that you could actually see it happening.”
Finding those intricacies within the LLMs will not be easy, Crovella added.
“They don’t give up their secrets easily,” he said. “We have to think carefully about the algorithms that are being used inside the models and ask how we can tease apart the pieces to understand things and talk about what the model is doing in human-understandable terms.”
Crovella likened the task to working with building blocks, noting that researchers are only beginning to grasp the parts, not the bigger picture of how they fit together.
“We can now sort of see the Lego pieces, but we don’t have a good strategy for explaining why the model puts those Lego pieces together in a certain way, and that’s a huge intellectual challenge, but it’s very exciting,” he said. “It’s an opportunity for great creativity and imagination.”
In February, Crovella was named one of two inaugural Duan Family Faculty Fellows (DFF) in BU’s Faculty of Computing & Data Sciences. Backed by a gift from the Duan Family, the program is designed to bring together leading senior scholars who can drive forward the mission of CDS.
“It’s very meaningful for me personally, because I think it reflects both the work that I’ve done in data science, but also the work that I’ve done in supporting the creation of the data science academic programs,” Crovella, the academic affairs chair at CDS, added.
“There’s been enormous progress. We have put in place in the past five years a PhD program, a master’s degree, an undergraduate degree, a data science minor that anyone in the university can take, an online masters of data science, and there’s more to come.”
Mark Crovella is a Department of Computer Science professor, a founding member of the Faculty of Computing & Data Sciences, and a Data Science Faculty Fellow. He is also Affiliate Faculty in the Department of Electrical and Computer Engineering, Graduate Program in Bioinformatics and Division of Systems Engineering.
His research interests span both computer networking and network science. Much of his work has been on improving the understanding, design, and performance of parallel and networked computer systems, mainly through the application of data mining, statistics, and performance evaluation. In the networking arena, he has worked on characterizing the Internet and the World Wide Web. He has explored the presence and implications of self-similarity and heavy-tailed distributions in network traffic and Web workloads. He has also investigated the implications of Web workloads for the design of scalable and cost-effective Web servers. In addition, he has made numerous contributions to Internet measurement and modeling, and he has examined the impact of network properties on the design of protocols and the construction of statistical models. In the network science arena, he has focused on the analysis of social, biological, and data networks.