Is AI Biased against Some Groups and Spreading Misinformation and Extreme Views? – Boston University

Photo by Lidiia Moor/iStock
Millions of us have played with artificial intelligence tools like ChatGPT, testing their ability to write essays, create art, make films, and improve internet searches. Some have even explored if they can provide friendship and companionship—perhaps a dash of romance.
But can we, should we, trust AI? Is it safe, or is it perpetuating biases and spreading hate and misinformation?
Those are questions that Boston University computer scientist Mark Crovella will investigate with a new project backed by a first-of-its-kind National Science Foundation (NSF) and Department of Energy program. The National Artificial Intelligence Research Resource (NAIRR) Pilot aims to bring a new level of scrutiny to AI’s peril and promise by giving 35 projects, including Crovella’s, access to advanced supercomputing resources and data.
A BU College of Arts & Sciences professor of computer science and a Faculty of Computing & Data Sciences professor and chair of academic affairs, Crovella will use the high-powered assist to examine a type of AI known as large language models, or LLMs. His goal is to audit LLMs—AI programs trained to study and summarize data, produce text and speech, and make predictions—for “socially undesirable behavior.” LLMs help drive everything from ChatGPT to automated chatbots to your smart speaker assistant. Crovella will be joined on the project by Evimaria Terzi, a CAS professor of computer science.
According to the NSF, the research resource pilot grew out of President Joe Biden’s October 2023 executive order calling for a federally coordinated approach to “governing the development and use of AI safely and responsibly.”
The Brink asked Crovella about the rapid expansion of AI, how it’s already part of our everyday lives, and how the NAIRR award will help his team figure out if it’s trustworthy and safe.
Crovella: Large language models are software tools like ChatGPT that are very rapidly becoming widely used. LLMs are quickly finding uses in education—both students and teachers use LLMs to accelerate their work; in social settings—many companies now are selling LLMs as online companions or assistants; and in science—researchers use LLMs to find and summarize important developments from the flood of research results published every day. Apple, Microsoft, and Meta have all announced integrations of LLMs into their product lines. In fact, ChatGPT had the fastest uptake of any new software, reaching 100 million users in just two months—much faster than did TikTok or Facebook.
Crovella: Given that millions of people will soon be interacting with LLMs on a daily basis, we think it’s important to ask questions about what social values these models incorporate. We’d like to know whether such models incorporate biases against protected groups, tendencies to propagate extreme or hateful views, or conversational patterns that steer users toward unreliable information.
The question is how to assess these tendencies in a system as complex as ChatGPT. In previous research, we’ve studied simpler systems from the outside. That is, we gave those systems inputs and observed whether their outputs were biased. However, when we look at LLMs, this strategy starts to break down. For example, we’ve found cases where an LLM will correctly refuse to answer a question on a sensitive topic when the question is posed in English, but simply by asking the same question in a different language (Portuguese), the model provides an answer that it shouldn’t. 

So, looking at an LLM just from the outside is not reliable. The genesis of this new project is to ask whether we can look inside an LLM, observe the representations of concepts that it is using, observe the logic that it is following, and, from that, detect whether the system is likely to incorporate undesirable social behaviors. In essence, we want to study an LLM the way a neuroscientist studies a brain in an fMRI machine.
I take Senator Markey’s views as emphasizing the need for independent research on AI systems. These fantastically capable systems wouldn’t exist without Big Tech—there’s just no way to marshal the resources needed elsewhere. At the same time, we need a lot of eyes on the problem of whether these new systems are serving us well. I want to point out that our work depends on the fact that some large tech companies have actually “open-sourced” their models—given their code and data away for free. This has been a socially beneficial act on their part and it has been crucial for the kind of work we are doing.
An LLM is large in two ways: it contains a huge amount of knowledge, obtained from vast training data, and it has an enormously complex system for generating output that makes use of billions of parameters. As a result, the internal representations used in LLMs have been referred to as giant and inscrutable. Just processing the huge amount of information inside a modern LLM is an enormous computational task. Our NAIRR grant gives us access to supercomputing facilities at top national laboratories that will enable us to efficiently analyze the internals of modern LLMs.
There are currently over half a million different LLMs available for the public to use. No doubt in the near future, we will each have our own personalized LLM that will know a lot about us and will help us with many tasks on a minute-to-minute basis. How will we know that these “giant and inscrutable” systems are trustworthy and safe? Our research is intended to provide methods for answering those questions.
Is AI Biased against Some Groups and Spreading Misinformation and Extreme Views?
Andrew Thurston is originally from England, but has grown to appreciate the serial comma and the Red Sox, while keeping his accent (mostly) and love of West Ham United. He joined BU in 2007, and is the editor of the University’s research news site, The Brink; he was formerly director of alumni publications. Before joining BU, he edited consumer and business magazines, including for corporations, nonprofits, and the UK government. His work has won awards from the Council for Advancement and Support of Education, the In-House Agency Forum, Folio:, and the British Association of Communicators in Business. Andrew has a bachelor’s degree in English and related literature from the University of York. Profile
Boston University moderates comments to facilitate an informed, substantive, civil conversation. Abusive, profane, self-promotional, misleading, incoherent or off-topic comments will be rejected. Moderators are staffed during regular business hours (EST) and can only accept comments written in English. Statistics or facts must include a citation or a link to the citation.
Your email address will not be published. Required fields are marked *

Boston University’s Alumni Magazine
News, Opinion, Community
Pioneering Research from Boston University


Leave a Reply

The Future Is A.I. !
To top