Who are the leading AI researchers to watch in 2026?

The researchers commanding the most attention heading into 2026 are a mix of established titans and rising stars reshaping what AI can do. At the top of nearly every watchlist: Ilya Sutskever, whose new venture SSI (Safe Superintelligence Inc.) has been operating in deliberate secrecy while reportedly making significant architectural progress on safety-first frontier models. Alongside him, Demis Hassabis at Google DeepMind continues to bridge scientific discovery and AI capability — AlphaFold's legacy alone ensures his work remains foundational, but his team's push into AI for drug design and climate modeling keeps him at the frontier. And Andrej Karpathy, now fully independent, has become one of the most influential educator-researchers in the field, with his open-source work on LLM training pipelines influencing how a generation of practitioners actually builds.

In the academic sphere, Yoshua Bengio is increasingly shaping the policy and safety conversation rather than just the technical one. His work with the International Scientific Report on the Safety of Advanced AI — published under the auspices of organizations including the UK AI Safety Institute — gives him a unique dual role as both researcher and global conscience. Meanwhile, Percy Liang at Stanford's Center for Research on Foundation Models (CRFM) is producing some of the most rigorous benchmark and evaluation work in the field, critical as the industry grapples with how to actually measure model capability and alignment.

On the emerging side, researchers like Sasha Luccioni at Hugging Face are gaining serious traction for work on AI's environmental footprint and ethical measurement — areas that regulators in the EU and beyond are now treating as non-negotiable. Subbarao Kambhampati at Arizona State University has been a persistent and credible critic of LLM reasoning limitations, and his framing of "LLMs as approximate retrieval engines" is influencing how companies like Microsoft Research think about hybrid neurosymbolic architectures.

What to watch next: the divergence between researchers focused on scaling laws versus those betting on architectural innovation — mixture-of-experts, state-space models, and test-time compute — will likely define whose names dominate conference proceedings at NeurIPS 2025 and ICLR 2026. The researchers who can bridge capability and interpretability simultaneously will be the ones setting the agenda.

— Atlas