Demystifying AI

Ask ChatGPT how to change guitar strings, and it answers flawlessly, providing a step-by-step tutorial it was never explicitly trained to give. This “emergent capability” is why we have shifted the paradigm from Machine Learning to Artificial Intelligence. It seems like intelligence, but true intelligence must be dependable and understandable. Just as neuroscience seeks to map the cognitive capacities and failure modes of the human brain, AI research requires a similar pedagogy. If we are to call these machines “intelligent,” we must understand the conditions under which that intelligence breaks down.

This falls under the umbrella of trust, a critical point of contention for modern AI systems. One aspect of this trust is interpretability: demystifying the “black box” to see how a model reaches a conclusion. Another is robustness: knowing for certain that a system isn’t susceptible to spontaneous, untraceable corruption. So, can we trust current AI systems? At our current level of progress, the answer is no. These risks are inherent in modern architectures, making them, in their current form, deeply untrustworthy.

The task of the AI researcher is to translate these sociological queries into mathematical systems that emulate our ideals. My current research with Professor Aalok Thakkar focuses on the robustness of transformers which is the machinery behind modern LLMs.

We are developing modifications to ensure these models act predictably in safety-critical domains, such as autonomous vehicles or medical imaging, where even a 0.0001% error is catastrophic. Our goal is to ensure that when a model encounters the unexpected, it doesn’t “hallucinate” or fail unpredictably. But as we work on mitigating these failure modes, we must face a reality: the world has already decided to run on AI.

Responsible AI, then, is no longer about deciding if we use these tools but about how we circumvent their worst-case possibilities.

My work with Professor Kasturi Mitra on cancer prediction via mitochondrial analysis showcased this. We encountered one of AI’s classic cognitive hurdles called “generalisation”: a model might recognise cancer in one population but fail on one from a different area. By utilising “meta-learning”, a technique that teaches the AI how to learn, we worked to ensure the model’s reliability across diverse populations.

Ultimately, the goal of my research is to bridge the gap between inherent trustworthiness and responsible deployment. We are moving toward a future where we don’t just hope a machine is right but also have the mathematical proof to verify its competence.

By mapping the limits of these digital brains, we ensure that as AI moves into our most vital institutions, we are not just following an oracle but are managing a tool we finally understand.

– Prabhpreet Sethia, UG’2025 student