I’m Harish Tayyar Madabushi, Associate Professor (Senior Lecturer) in Artificial Intelligence at the University of Bath. I work at the intersection of Natural Language Processing, reasoning, linguistics and AI safety, asking a simple but stubbornly hard question: how do large language models actually learn to use language and reason, and what does that mean for how we deploy them in the real world?

My research treats usage-based learning as the foundation for how language competence and bounded reasoning emerge in large language models. On the linguistic side, I draw on Construction Grammar (CxG) to understand what kinds of constructions models really acquire; on the modelling side, I develop Context-Directed Extrapolation, a framework for explaining so-called “emergent abilities” not as magic or AGI, but as models extrapolating from statistical priors when guided by context. This work has reshaped parts of the debate around LLM capabilities, influencing research, policy discussions (including the UK AI Safety Summit), and industry practice. 

I organise my work into three strands:

• Emergent reasoning and limits of LLMs: mapping what models can and cannot do, and when scaling stops being a silver bullet. 

• Linguistically grounded evaluation: using insights from theoretical linguistics, specifically construction grammar to probe abstraction, generalisation and comprehension in LLMs.

• Applications in sensitive domains: from debiasing speech systems and supporting public services, to studying how LLMs may nudge users towards extremist content, and designing scaffolds (like divergent chain-of-thought and frame-based retrieval) that make models more reliable, interpretable and standard-aligned.  

To read more about how I think about AI safety, “existential threat”, and the real-world risks of current systems, see the Press & news coverage section of this site. To see my publications, visit the Research page; to learn more about my courses and supervision, see the Teaching page.