When Snigdha Chaturvedi first began programming as a middle school student in India, she was, admittedly, terrible at it. She approached coding with the mindset of memorizing every line, focusing on recalling each letter and number exactly as it appeared, rather than understanding the underlying logic or principles.
“It’s such a rich field with something for everyone,” she reflects.
Chaturvedi first heard the term artificial intelligence (AI) in a class during her first year of university. She was fascinated with the idea that a machine could take a set of data, imitate human intelligence and behavior, and make predictions.
“Of course, my fascination with AI in general grew as I took more courses, went to grad school, and did my PhD,” she says. “And the technology developed as I developed.”
As Chaturvedi learned more about AI models, she became deeply concerned with how quickly the technology was evolving, leading to potential misuse and a lack of public understanding. More recently, she has noticed how comfortable people seem sharing their personal information with AI agents, making privacy and data sharing a focus of her worries.
One of her biggest concerns is bias in AI systems. AI is increasingly used in critical real-world applications like college admissions, job hiring, and loan approvals. Its widespread use in practical applications means that biases have tangible consequences in people’s lives. For example, biased hiring algorithms can discriminate against qualified candidates, and biased loan approval systems can deny credit to individuals based on unfair factors.
Now a computer scientist at UNC-Chapel Hill, Chaturvedi is working to build better methods to train AI models.
Impact Report
From the health sciences to journalism, sports to social work, UNC-Chapel Hill researchers are applying AI to the world’s problems. Carolina is also home to one of the top 30 computer science graduate programs in the U.S.
Snigdha Chaturvedi’s research is funded by grants from the National Science Foundation. This funding supports vital research, helping the people of North Carolina and beyond.
Finding her specialty
Before finding her place in academia, Chaturvedi was a researcher at IBM-India Research Labs. She worked on their information management team from 2009 to 2011 to improve the processing efficiency for large volumes of unstructured text data — text that doesn’t follow a fixed format.
“This research experience helped me discover my passion for Natural Language Processing (NLP) and solidified my decision to pursue a PhD in the field,” she explains.
In 2020, she left her position as an assistant professor at the University of California, Santa Cruz, to join the faculty at Carolina. Since then, UNC-Chapel Hill has become a leader in AI research and application, and Chaturvedi has been a major contributor to conversations about ethics and education in AI.
“I get to do what I’m passionate about, but at the same time, I’m in close contact with other people who are doing related research, which is an incredible resource,” she says.
Redefining language models
Chaturvedi’s research focuses on how NLP and Large Language Models (LLMs) can understand and respond to human communication with more social awareness.
NLP is a branch of AI that trains machines to understand and interact with human language. Think of it as teaching a computer to read and understand a book — deciphering words, sentences, and their meanings. LLMs, like GPT and BERT, are some of the most advanced tools in NLP.
These models are trained on vast amounts of text, which they then learn from to generate responses. The better the data fed to the system and the better the training procedure, the more human-like and accurate its response. If data with bias, inaccuracies, or misinformation is used to teach the system, that will lead to biased outputs or nonsensical responses.
“Humans are so natural with language,” Chaturvedi explains. “Even a 2-year-old can understand language and to some extent speak. But language is just so challenging for machines.”
These models struggle to grasp the social context, emotions, and deeper meanings behind language. And Chaturvedi is determined to find ways to make the technology more empathetic and in tune with how humans interact — a strategy she hopes will diminish the social bias that these machines sometimes exhibit.
To do this, she is exploring what type of text should be used to train AI models. This iterative process starts with feeding specific information to a system with the goal of having it predict the next word for a given sequence with the social awareness of humans. Depending on the response she gets, Chaturvedi adjusts the system’s internal settings, called weights, based on whether they’re right or wrong and repeats this process millions of times until the AI model understands language, grammar, and meaning on a human level.
In 2024, Chaturvedi created SocialGaze, a framework designed to guide a language model through a thoughtful, multi-step process in which it examines a social scenario from different people’s feelings, attitudes, and actions. She found that when prompted to analyze multiple perspectives, LLMs were more aligned with human judgements.
“There are so many implications of a word, and not just linguistic implications, but also social and pragmatic implications,” she says. “So, it’s very important to explore these possibilities.”
Advocating for compassionate technology
When thinking about the potential impact of her research, Chaturvedi envisions a world where AI agents can provide support in areas where people lack access to essential services like therapy or health care.
“I come from a conservative society in India where people don’t have access to those facilities,” she says. “Or if they have access, there’s just so much social stigma associated with seeking help that people would rather suffer their entire life than make an appointment.”
In these cases, having AI agents that understand the subtle complexities of human communication is critical to providing thoughtful responses that relate to specific human experiences.
“Imagine the potential of an AI agent with deeper social understanding,” she suggests. “If somebody is stressed, they can just take out their phone and chat with someone — an AI agent. I think that’s a very powerful use case.”
She envisions AI as a tool to support professionals like doctors and social workers. These models can excel at routine tasks like transcribing notes and analyzing data, freeing up professionals to focus on more complex and personal interactions with patients or clients.
Chaturvedi also believes AI can act as a “co-expert,” quickly processing data and identifying patterns that might take humans hours to find, enhancing the quality of care.
“The recent progress in NLP has been very invigorating,” she says, excitedly. “Because I see that my research is now being used for so many people in day-to-day life.”