Matt Pagett

Welcome to my personal website.

Research interests

  • AI and robotics explainability, trustworthiness, and accountability
  • Digital humanities
  • World models

Select projects

Research projects

EMPO - Empowerment-based AI agents (led by Jobst Heitzig; I am contributing in connection with AI Safety Initiative Groningen)

Hackathons

Cross-Border Agentic AI Compliance (CBAAC): Embedding Regulatory and Cultural Risk Compliance into Agentic Communication (with Tomoko Mitsuoka, submitted to Apart Research Technical Governance Hackathon - February 2026)

Adversarial LLM span detection for social sycophancy observation (with Pranati Modumudi, submitted to Apart Research AI Manipulation Hackathon - Jan 2026)

Comparative LLM methods for social media bot detection (With Andreas Raaskov. 2nd place winner, November 2025 Apart Defensive Acceleration Hackathon)

Protocop - social media moderation toolkit using MCP (Entered into MCP 1st Anniversary Hackathon, November 2025)

Digital humanities

Cervantes Reader (v1) - AI-powered narration/analysis for early modern literature

Artificial Quartet. An early attempt to use generative AI to write a play. Submitted as an entry to the Financial Times 2020 Weekend Festival theater contest.

Resources

AI Safety

ENAIS - The European Network for AI Safety offers worldwide AI safety education. I’m a faciliator for the Winter 2026 AI Safety Collab.

BlueDot provides education about AI safety and governance. I completed the AGI Strategy and Technical AI Safety courses in 2025 and recommend them to anyone interested in AI risk.

ARENA - detailed technical AI safety course. I’ve formed a self-study group on the ARENA Slack channel.

Software Development

SolveIt is a course, a method, and a platform to use LLMs thoughtfully in software development (and other areas, like writing). (This is an affiliate link - I’ll earn a commission and you’ll get a discount if you sign up using this link). It is led by Jeremy Howard whose Fast.AI course on Deep Learning remains as relevant and useful as when I first took it in 2019.

Find me at

LinkedIn

Github