Safe artificial intelligence (AI) isn't an oxymoron to computer engineering Ph.D. student Bilgehan Sel MS'24 - it's his future.

That future is one step closer thanks to his recent acceptance into the Anthropic Fellows program.

"I am very proud of Bilgehan. It’s a significant, well-deserved honor," said Ming Jin, assistant professor, and Sel's advisor. "What makes him a strong fit for Anthropic is his ability to take a complex, theoretical idea, like the abstract AI safety concepts we discuss, and iteratively deepen his understanding using empirical insights, ultimately turning it into a real, working system. That ability to gain clarity and uncover what's possible through empirical observation is an essential trait for making progress in modern AI safety research."

Anthropic, the developer of the large language model and AI chatbot, Claude, is dedicated to the safety and social impact of AI. Through the Fellows program, students work on projects like

  • adversarial robustness and AI control
  • scalable oversight
  • model organisms of misalignment
  • interpretability

Sel is one of only 32 students out of 2,000 global applicants to be accepted into the second cohort of the Fellows program. 

"This fellowship is a logical and exciting next step for him," said Jin, the Shirish S. Sathaye Junior Faculty Fellow in ECE. "He is deeply committed to building safe and reliable AI, and this role places him at the center of that mission. I am confident he will make meaningful contributions to AI safety and will be a future leader in this field."

We asked Sel a few questions about his research background, what he hopes to learn as a Fellow, and what safe AI means to him.

What's your area of research, and why are you drawn to it?

My research focuses on safe decision-making in large language models, particularly how to ensure that AI systems reason in ways that are aligned with human values and remain dependable in complex real-world settings. I develop training and evaluation techniques that help models reflect on their reasoning, assess uncertainty, and avoid harmful or unreliable behaviors.

I was drawn into this area through my work with my advisor, Prof. Ming Jin, whose guidance has been central to shaping both the technical and conceptual direction of my dissertation. As AI systems become more integrated into decision processes across society, ensuring that they act responsibly and reliably is not just a research challenge, but an urgent one. That importance is what motivates my work.

What drew you to the Anthropic fellowship? What are you excited to learn?

Anthropic is one of the leading organizations focused on developing highly capable AI systems that are designed to be safe and aligned from the ground up. Their research culture emphasizes careful reasoning, interpretability, scalable oversight, and long-term system reliability, all of which closely connect to my research interests.

I am excited to collaborate with researchers working on alignment techniques that can operate at the scale of state-of-the-art models. Being able to work directly in that environment will help me both contribute to and learn from ongoing advances in safe AI development.

What does safe AI mean to you?

Safe AI is not simply AI that avoids obvious mistakes. To me, safe AI means systems that understand context, communicate uncertainty, and make decisions in ways that respect human values and priorities. A safe AI system should be transparent in how it reasons, robust when facing new situations, and aligned with the intentions of the people using it.

Ultimately, safe AI is about trust. We want systems that people can collaborate with confidently. Building that kind of trust is a long-term effort and a shared responsibility across research, policy, and industry.