From Idea to Market to Scale: UNC AI Community Launches New Learning Series

A new series hosted by Innovate Carolina brings together AI enthusiasts and entrepreneurial experts from the startup world to navigate new ways to create, collaborate and scale.

AI-Community-Networking
September 24, 2025
By Brock Pierce, Innovate Carolina
Photography by Brock Pierce, Innovate Carolina

What are some of the intriguing possibilities—and dangerous pitfalls—to consider when using artificial intelligence to transform research discoveries into new innovations? The UNC AI Community, convened by Innovate Carolina, launched its 2025–26 event series with a panel discussion exploring how AI is shaping research translation. Held in September at the Innovate Carolina Junction in downtown Chapel Hill, the kickoff session titled From Discovery to Direction: How AI Powers Research Translation highlighted practical examples of AI in action and the challenges, ethics, and future opportunities that researchers face.

This panel was just the first in the 2025-26 UNC AI Community series—How AI Is Reshaping the Innovation Journey: From Idea to Market to Scale. Future events, held regularly through March, will address topics from early-stage idea validation to fundraising, company creation, and sustainable growth. If you missed the September session, read on for a rewind of the highlights. Or if you want to catch the next event, you can learn more about the session and register: Before the Startup: How AI Supports Idea Validation and Market Strategy (Oct. 7).

Session 1 Rewind: How AI Powers Research Translation

Moderated by Richard Marks, PhD, Professor of the Practice in the UNC School of Data Science and Society, the panel brought together researchers from across the Carolina campus who use AI for translational research projects:

  • Roni Sengupta, PhD, Assistant Professor, Department of Computer Science, UNC College of Arts and Sciences.
  • Xianming Tan, PhD, Research Professor, Department of Biostatistics, UNC Gillings School of Global Public Health.
  • David Borland, PhD, Assistant Director of Analytics & Data Science, Renaissance Computing Institute (RENCI), UNC-Chapel Hill.

These panelists are four of nearly 60 experts featured in the community’s AI Expert Directory, a resource designed to encourage collaboration, support research initiatives, and connect the campus community with individuals advancing the frontiers of AI. During the session, they discussed how AI tools and methods advance their work, the challenges of applying rapidly evolving technologies, and the responsibilities researchers hold in ensuring ethical and effective AI use.

AI Applications in Research

Panelists shared examples of how they’re using AI to address a range of complex research problems.

  • Transportation and safety. David Borland described a collaboration with the North Carolina Department of Transportation to use machine learning to analyze video data collected from state highways. By automatically detecting roadside hazards from vast volumes of images, AI has the potential to improve safety and efficiency in maintaining infrastructure.
  • Cultural and historical analysis. Borland also highlighted a project with an art history scholar using AI to analyze 18th- and 19th-century French satirical cartoons. Generative AI tools demonstrated surprising abilities, such as recognizing mythological references, along with limitations, like failing to identify common objects like guitars.
  • Medical and clinical research. Xianming Tan emphasized the potential of AI to accelerate labor-intensive systematic reviews and analysis of clinical notes recorded by health care providers. By helping researchers process thousands of papers or extract structured insights from unstructured records, AI offers time savings, though quality and privacy remain key concerns.
  • Image and video editing. Roni Sengupta’s lab develops AI algorithms for understanding physical properties from images and videos, with applications ranging from augmented reality to robotics. One area her team is focused on is building training-free approaches, which adapt large models for new tasks without requiring additional datasets or retraining AI models.

Challenges: Rapid Evolution of AI Technology and Governance

The panelists noted that the speed of AI’s evolution makes it difficult for researchers to keep up.

  • Learning curve. Sengupta advised fellow researchers working with AI to start with a solid baseline code rather than coding from scratch. This approach reduces errors and allows teams to build incrementally on proven foundations.
  • Governance. Panelists and audience members raised concerns about managing data privacy, responsible use of AI in student work, and the risks of proprietary information entering external models. Marks noted that universities can help guide researchers toward sanctioned tools that protect confidentiality.
  • Knowledge graphs. Borland discussed the importance of combining large language models with structured representations of knowledge. Knowledge graphs—databases that capture relationships between entities (e.g., genes linked to diseases, or drugs linked to outcomes)—help ground AI outputs in validated knowledge. Participants commented that knowledge graphs are increasingly essential because they add context and reduce hallucinations, enabling AI systems to provide more reliable reasoning in scientific domains.

Ethical Considerations: Bias, Hallucinations, and Trust

Panelists and discussion participants also talked about ethical dimensions of AI in research.

  • Bias and misinterpretation. Borland described studies showing how human biases influence data visualization interpretation, and how large language models replicate these biases. He emphasized the importance of building systems that detect and guard against such distortions.
  • Hallucinations. Marks and others highlighted examples of AI inventing false information—sometimes pulling from spoof websites or fabricating nonexistent legal cases. The risk is not just overt error but subtle misinformation entering research.
  • Deepfakes and detection. Sengupta shared findings on AI-generated talking-head videos, noting that while such deepfakes can be nearly indistinguishable to humans, algorithms can detect pixel-level inconsistencies with high accuracy. She raised concerns about how humans can be trained to recognize manipulated media when algorithms alone cannot scale to the open internet.
  • Evaluating AI vs. traditional methods. Tan recommended that researchers compare AI-generated outputs against proven statistical methods to ensure validity. He advocated for sampling and quality checks before trusting AI results in systematic reviews.
  • Privacy in clinical contexts. Tan also stressed the risk of patient data exposure when clinical notes are processed with external AI tools. To prevent breaches, his team works on developing methods that can run locally while maintaining accuracy.

Future Directions and Advice for Researchers

Looking ahead, panelists emphasized both opportunities and cautions for the role of AI in research.

  • Physical AI and robotics. Sengupta predicted a resurgence of robotics research, particularly as AI systems are tasked with safely interacting in physical spaces. She described this as an exciting frontier where AI must integrate with disciplines like mechanical engineering to achieve breakthroughs.
  • Creative problem selection. Sengupta also recommended that researchers consider tackling creative problems that large companies aren’t currently working to solve themselves. For example, while big tech invests heavily in standard image-editing tools, her lab explores niche challenges such as developing AI models that will adhere to strict definitions of specific elements that need to be changed or corrected in an image without changing other characteristics—work that industry is interested in but less likely to prioritize right now.
  • Workflow automation. Marks encouraged using AI to offload repetitive or disliked tasks, such as generating test code or form letters. He framed AI as a “second pair of eyes” that can boost productivity without replacing human judgment.
  • Quality and evaluation. Borland emphasized the need for continuous evaluation of AI outputs, especially as models evolve rapidly. Establishing testing protocols and feedback loops, he noted, will be key for maintaining reliability.
  • NIH policy on AI in grant proposals. The panel closed with discussion of new guidance from the National Institutes of Health (NIH), which restricts the use of AI tools in preparing grant submissions. The NIH now requires applicants to disclose if generative AI is used for editing and prohibits relying on AI for idea generation or substantive content creation. Panelists agreed that while AI can be helpful for copyediting or polishing language, it cannot replace the intellectual contribution of researchers. This policy highlights the need for transparency and reinforces that quality control and originality remain central to academic integrity.

Looking for more of these kinds of insights and recommendations? Sign up to join Innovate Carolina and the UNC AI Community for the next session in October—Before the Startup: How AI Supports Idea Validation and Market Strategy.