Q&A: Maria Mahbub on AI trustworthiness

Published: July 23, 2026

Updated: July 23, 2026

Maria Mahbub is a research associate at the Department of Energy's Oak Ridge National Laboratory in emerging cyber technologies. Though she's early in her career, her work is making a major impact in artificial intelligence trustworthiness as she transforms complex, unstructured data into actionable insight for health, safety, and national security.

Mahbub finished her doctorate in 2023 from the University of Tennessee, Knoxville in computer science, and afterward joined the lab as a postdoctoral researcher. She joined the ORNL staff in 2024 as part of the Cyber Resilience and Intelligence Division in the National Security Sciences Directorate and has since been consistently recognized for her work in artificial intelligence research.

Her work in artificial intelligence is a testament to the technology's wide range of applications, including healthcare and national security. She's contributing to a multi-year project with the Department of Veterans Affairs (VA) Office of Mental Health and Office of Suicide Prevention, where she is analyzing how effective large language models are in identifying substance use disorders (SUDs) severity-related information from clinical notes. The VA hopes this work can contribute to improved risk assessment and treatment planning for SUD patients.

She is also contributing to a project funded by the Department of Homeland Security focusing on risk evaluation of AI and machine learning models. Her work currently focuses on assessing the reliability and failure modes of Retrieval Augmented Generation (RAG) evaluation metrics.

Through the Laboratory-Directed Research and Development (LDRD) program, Mahbub recently won a competitive Early Career Competition award for a project on using image technology to detect drones in the sky in adverse weather conditions.

Read on to learn more about her work, what she sees when she gazes into the AI crystal ball, and whether she'd let a robot do her gardening for her.

Q: How did you find yourself working in emerging cyber technologies?

A: My bachelor's degree was in mathematics. When I was looking for a doctoral program, I started looking at data science programs because I found data fascinating. Then I started learning about machine learning and artificial intelligence. I was fascinated by how these models could learn from all kinds of data and be applied to so many different problems. I completed my PhD in natural language processing with biomedical data.

Q: Your early career research has spanned a range of disciplines. What overarching goals connect your work, and how does AI help advance those objectives?

A: I like to be involved in a variety of projects, so I don't limit my exposure to one direction. With the rise of large language models (LLMs), the natural language processing field has changed dramatically. As LLMs have become widely adopted, many of the questions I studied during my PhD have become increasingly important. What's exciting to me is that similar techniques can be used with image data, genetic data, time-series data and many other types of information.

I like to say I'm both an AI person and a data person. I'm interested in developing methods that can learn from different kinds of data and help answer important questions. At the same time, I acknowledge that AI expertise doesn't replace domain expertise. For example, when I work with genetic data, I can build models and analyze the results, but I rely on biologists and other subject-matter experts to help interpret what those results actually mean. The most impactful research often comes from those collaborations.

Another area I'm really interested in is AI reliability. Everybody is using AI systems now, but how do we know when we can trust them? How do we know when they're doing the right thing, and how do we identify when they might fail? Those are the kinds of questions I'm interested in studying.

Q: When you talk about robustness and reliability in AI models, what does that mean?

A: When I talk about robustness and reliability, I'm really asking whether we can trust a model's output.

One aspect is understanding how a model behaves when things don't go as expected. What happens if the data look different from what the model saw during training? What if there are errors, noise, or unusual patterns in the data? We often stress-test models by intentionally challenging them to identify their weaknesses and understand where they might fail.

Another aspect is verifying the accuracy of a model's output. With modern AI systems, especially large language models, it's often impossible for humans to review every output. In some applications, a model may generate millions of predictions or responses. I recently worked on a project that explored how automated evaluation methods can assess clinical outputs at that scale, without requiring clinicians to manually review every single result - they don't have the time.

There's also the question of data fidelity and evaluation quality. We often focus on whether we can trust the model, but we also need to ask whether we can trust the data and the evaluation methods we're using to measure performance. If those are flawed, then our conclusions about the model may be flawed as well.

Ultimately, robustness and reliability are about understanding a model's limitations, identifying where it can fail, and building confidence that it will perform as expected in real-world settings.

Q: You talked about needing somebody with domain expertise to get the takeaway from the data. What's your experience partnering with these experts been like?

A: I've had great experiences working with domain experts. For example, I'm part of a multi-year project with Department of Veterans Affairs, where I collaborate closely with clinical psychologists. We have ongoing discussions about the results, what they mean, and what questions are most important to answer.

A big part of my role is translating between what the data can tell us and what the clinicians need. I help them understand the capabilities and limitations of the models, and they help me interpret the findings and put them into proper clinical context.

I really enjoy that kind of collaboration because, without them, the data often lack meaning.

Even when publishing research, it's important to have a subject-matter expert validate whether the results make sense.

Q: Can you explain the research you are doing with drones?

A: Yes, that's my early career LDRD project. I'm working with image data to detect drones in the sky from the ground. One of the key questions is whether we can reliably detect drones in adverse weather conditions. If it's raining, snowing or foggy, can the system still detect the drones?

There's also a robustness aspect to the project. The idea is to understand how we can make drone detection systems more reliable when conditions change. If the model encounters conditions that are different from what it saw during training, can it still detect drones accurately?

If you think about AI more broadly, many of the underlying concepts are the same. The data may be different, text versus images, but many of the challenges around learning from data, generalization, and robustness remain the same.

Q: What have you learned so far in this project?

A: We are still in the early stages of the project, and right now we're focused on simulating data. One thing we have learned is that data simulation is much more challenging than it might seem. There are a lot of technical considerations, from the simulation software itself to ensuring that the generated data accurately represent real-world conditions. And we decided to use simulation because we're especially interested in adverse weather scenarios. If we relied entirely on real-world data collection, we would have to wait for the right weather conditions to occur, which isn't practical within the timeline of the project.

Q: Have you worked with any other fields of science as you apply AI to different things?

A: Yes. I've worked with genetic data, particularly in applications related to disease prediction, which was also the topic of my Your Science in a Nutshell competition.

In addition to those application areas, I've worked on AI robustness and security, including adversarial attacks on machine learning models, as well as RAG and the evaluation of RAG systems in real-world deployment settings.

What I enjoy most is working across disciplines. The applications may be very different, but I often find myself asking the same kinds of questions: Can we trust the model? Where might it fail? How do we know it's working the way we expect?

Q: What have been your favorite impacts of the research you've done so far?

A: I have learned that meaningful impact takes time. The projects I've been involved with the longest are often the ones where I've been able to see the greatest real-world impact.

For me, that's been the VA project. Over the years, we've developed tools and analyses that have helped inform future considerations regarding screening processes within the VA. We've also been working on concept extraction methods to help identify veterans who may have mental health indicators in their clinical records.

Q: You're doing the technical part of things, but there's also a very human part of these projects. What does it feel like to have real-world impact with some of the work you're doing?

A: That's one of the things I value most about my work. I tend to be drawn to projects where I can see a clear connection between the research and a real-world problem. Whether it's helping improve mental health screening for veterans, supporting cancer prediction research, or working on technologies that have implications for national security, I know that the work has the potential to make a difference beyond the research itself.

As researchers, we spend a lot of time developing models and analyzing data, so it's especially rewarding when you can see how that work might ultimately help people or contribute to solving important challenges. That's a big part of what motivates me.

Q: Where do you hope your research path goes in the future, and what do you hope to continue?

A: In the near future, I hope we achieve strong results from my LDRD project and show capabilities that lead to additional research opportunities and greater visibility for the lab's work in drone detection.

I also hope to continue my work with the VA and contribute to improvements in healthcare through the AI tools and methods we're developing.

Overall, I see myself continuing to work in AI and applying it to important real-world problems.

Q: Where do you see AI as a field moving toward the future?

A: It's always difficult to predict where AI will go next because the field is moving so fast. But one trend I expect to continue is increased automation.

We're already seeing the emergence of AI agents that can perform tasks and make decisions with limited human intervention. Right now, humans are still heavily involved in overseeing and validating those decisions, but I think we'll continue to see progress toward systems that can operate more independently.

I also think that automated evaluation of AI systems will become increasingly important. As we give AI more autonomy, we need reliable ways to assess its performance and make sure its decisions are trustworthy.

I'm also excited about the growth of autonomous laboratories, where AI systems help design experiments and guide robotic platforms. I think they have the potential to significantly accelerate scientific discovery.

Q: Outside of your research, what do you enjoy doing in your free time?

A: I love to play board games with my friends and spend time with my cat. I've also recently started gardening. I'm growing tomatoes, of course and okra. I'm originally from Bangladesh, so I'm also trying to grow some vegetables from there from seed. It's been exciting to watch the seedlings grow.

Q: Do you plan on using AI for gardening?

A: No, I'm not letting AI take that joy away from me!

UT-Battelle manages ORNL for DOE's Office of Science, the single largest supporter of basic research in the physical sciences in the United States. The Office of Science is working to address some of the most pressing challenges of our time. For more information, please visit energy.gov/science.

Media Contact

Eric J Swanson , Communications Coordinator, National Security Sciences , 865.341.1642 | [email protected]