As AI advances at an accelerating pace, its promises and risks are top of mind for researchers and the public alike. There is a growing call for “trustworthy AI”—systems that are reliable, explainable, aligned with established knowledge, and unbiased.
But when even developers can’t fully comprehend the complex workings of the models they create, precisely shaping their behavior is an open challenge.
Savannah Thais, an Associate Research Scientist at the Data Science Institute, suggests that physics, which Nobel laureates Geoffrey Hinton and John Hopfield used to help lay the foundations for AI, might also hold the key to building trustworthy AI.
In her recent commentary in Nature Reviews Physics, Thais proposes using physics datasets as a testing ground for AI models, bridging the gap between theoretical models and real-world application to make AI systems more transparent and dependable.
We spoke with Thais, who works at the intersection of science, AI, and ethics, to hear more.
Can you start by explaining what “trustworthy AI” is and why it’s important?
Trustworthy AI is about ensuring that an AI system’s behavior is predictable and transparent across different scenarios. My work focuses on using physics data, which is well-structured and precise, as a testing environment for AI models. By applying these models to physics datasets, we can observe how well they align with established physical laws and detect deviations. This allows us to refine the models, ensuring they behave as expected and reducing biases or errors that might arise in more unpredictable settings, such as healthcare or social sciences. It’s a way to bring AI evaluation closer to the controlled, observational methods used in physics.
You mention an “empirical gap” in AI research. Can you elaborate on why this gap exists and why it’s problematic?
The empirical gap I discuss is the difference between the theoretical foundations of AI and the models that we actually use in real-world applications.
There’s a lot of great research that explains how AI should behave under specific conditions, but this often applies only to simple models or controlled environments. With complex systems like ChatGPT, for example, we might understand the basics of how it works, but we can’t always predict why it gives a particular response in every situation.
This gap becomes a major issue when AI systems are deployed outside controlled research settings. The behavior of these models in practical scenarios doesn’t always match what we expect based on theory. This unpredictability can undermine trust, making it critical for us to bridge this gap if we want AI to be dependable and ethically aligned.
We all have a role in shaping the future of technology, and I believe it takes collaboration from all sectors—academia, industry, and the public—to build a future where AI is safe, fair, and beneficial for everyone.
– Savannah ThaisHow does physics fit into your strategy for addressing these challenges in AI ethics and evaluation?
Physics provides a unique opportunity to develop trustworthy AI because of its structured and precise nature. In my paper, I suggest using physics data as a “sandbox” to test and refine AI models. Physics datasets are mathematical and quantitative, which makes them compatible with the way AI models work—they, too, are built on quantitative data and algorithms.
By testing AI models against the well-established laws of physics, we can see if they behave as expected. If a model deviates from these known laws, it’s easier to identify where it went wrong. This kind of precision is challenging to achieve with less structured data, such as social data, which can be unpredictable. Physics allows us to create a controlled environment where we can systematically evaluate and fine-tune AI systems.
I mentioned three key areas, among others, where physics can help: reducing bias, improving explainability, and evaluating the architecture of a model. Physics gives us the advantage of working within a predictable and controlled framework, which is crucial for building transparent and accountable AI.
Has this approach—using physics principles to evaluate AI—been applied before?
While AI has been used to solve physics problems for some time, leveraging physics principles to evaluate and improve AI systems is less common. That said, the connection between AI and physics isn’t new. The 2024 Nobel Prize in Physics awarded to John Hopfield and Geoffrey Hinton highlights the application of physics concepts to develop neural networks, which are the foundation of many current AI systems. This shows that physics has long been foundational to AI development, not just in building systems but also in understanding and testing them.
The work of Hopfield and Hinton underscores that physics can be a powerful tool for ensuring AI systems are not only effective but also trustworthy. This perspective is just beginning to gain traction, and I believe it’s an area with significant potential for growth.
What inspired you to explore this intersection of AI and physics?
My PhD is in physics, and I initially used machine learning to solve physics problems during my PhD. But I’ve always had an interest in social sciences and policy. As I started learning more about AI ethics, especially issues like bias and misinformation, I noticed a gap between theory and practice that intrigued me.
I realized that my expertise in physics could be instrumental in addressing these challenges. AI has a profound impact on society, influencing everything from healthcare to education, and I wanted to apply my skills to make these systems more reliable and trustworthy. It felt like a meaningful way to contribute to the ongoing development of AI.
How do you see the role of places like the Data Science Institute, in tackling these challenges?
DSI and similar organizations are crucial for promoting interdisciplinary work, because academia is often siloed: different fields have distinct publishing venues, tenure criteria, and academic traditions, which can make working at the intersection of fields difficult. DSI helps bridge these gaps by supporting collaboration across departments through initiatives like seed grants. It’s a very unique support for cross-disciplinary work.
The tech industry is leading much of AI research today, but contributions from academia are still vital. Unlike industry, academia allows for exploration of long-term, foundational questions without the same pressure for profitability. This freedom enables us to tackle ethical and structural issues in AI without the constraints that industry researchers may face.
Is there anything else you would like to share?
AI is something we all need to think critically about, not just researchers. Many of us at Columbia and beyond use AI in our work, and it’s important to understand how these models function to ensure that our research remains accurate and responsible. We all have a role in shaping the future of technology, and I believe it takes collaboration from all sectors—academia, industry, and the public—to build a future where AI is safe, fair, and beneficial for everyone.