Andrew Trask
Most organisations want the upside of AI but cannot share the data that would make their models useful. Regulators, customers, and competitors all push in opposite directions, and the standard answer is to slow down. The harder question is how to use sensitive data across institutional boundaries without giving it up, and that question is now sitting on the desk of every senior leader running an AI programme.
Andrew Trask is a DeepMind research scientist and founder of OpenMined who helps organisations use AI on data they cannot share, through privacy-preserving infrastructure built for real institutional use.
Full Profile
Why organisations work with Andrew Trask
- He has built the working infrastructure for privacy-preserving AI, not just written about it. PySyft, the open-source framework he created through OpenMined, is used by researchers and engineers to train models on data they never directly see.
- He sits inside one of the most consequential AI research labs in the world, DeepMind, while also leading an 18,000-strong community of privacy-preserving AI engineers. Few speakers operate credibly in both rooms.
- His agenda is structured transparency: making it possible for institutions to collaborate on AI without surrendering the underlying data. That is the exact tension a regulated industry leader is now trying to solve.
- He teaches the material rather than narrates it. Grokking Deep Learning and his Udacity courses have introduced tens of thousands of practitioners to neural networks and to private AI, which is why technical audiences find him credible rather than performative.
Biography highlights
- Senior Research Scientist, Google DeepMind, working on privacy, federated learning, differential privacy, and secure computation.
- Founder of OpenMined, an open-source community of more than 18,000 researchers and engineers building privacy-preserving AI tools.
- Creator of PySyft, the open-source framework for training machine learning models on data the model owner cannot see.
- PhD candidate, University of Oxford, on the Oxford-DeepMind Graduate Scholarship; affiliated with the Future of Humanity Institute and the Centre for Governance of AI.
- Author of Grokking Deep Learning, Manning Publications, and instructor on Udacity’s Secure and Private AI and Deep Learning Nanodegree courses.
- Term Member, Council on Foreign Relations.
Biography
The hardest problem in applied AI is not modelling. It is access. The data that would make a healthcare model genuinely useful sits inside hospital systems that cannot legally share it. The data that would make a financial crime model work is split across banks that compete with each other. The default response has been to do less.
Trask has spent the last decade building the technical answer to that problem. Through OpenMined, the open-source community he founded, he created PySyft, a framework that lets engineers train models on data they never directly see. The underlying methods, federated learning, differential privacy, secure multi-party computation, homomorphic encryption, are what serious AI governance now relies on. He calls the broader agenda structured transparency, and it is becoming the operating vocabulary for AI deployment in regulated settings.
The institutional positioning matters. He is a Senior Research Scientist at Google DeepMind and a PhD candidate at Oxford, funded by the Oxford-DeepMind Graduate Scholarship and affiliated with the Future of Humanity Institute and the Centre for Governance of AI. He is also a Term Member of the Council on Foreign Relations. That combination puts him inside the frontier AI conversation and inside the AI governance conversation at the same time, which is unusual.
The teaching record is the test of whether his ideas travel. Grokking Deep Learning, published by Manning, taught a generation of engineers how neural networks actually work. His Udacity courses on secure and private AI did the same for the privacy-preserving stack. The reason senior technical audiences treat him as a primary source rather than a commentator is that they have already worked through his material.
Key speaking topics
- Privacy-preserving artificial intelligence
- Federated learning and secure computation
- AI governance and structured transparency
- Differential privacy in practice
- AI safety and responsible deployment
- The future of data infrastructure for AI
- Foundations of deep learning for non-specialists
Ideal for
- Chief AI officers, CTOs, and chief data officers in regulated industries (healthcare, financial services, government) deciding how to deploy AI on sensitive data.
- Boards and audit committees pressure-tested on AI governance, data risk, and the gap between AI ambition and AI infrastructure.
- Heads of research, innovation, and product engineering teams building the next generation of AI products.
- Policy and public-sector leaders responsible for AI standards, data sharing, and international AI governance.
Audience outcomes
- A working mental model of the privacy-preserving AI stack, named technique by named technique, instead of the usual abstractions.
- A clearer view of which AI use cases are unlocked by federated learning and secure computation, and which remain blocked by data access.
- A grounded read on where AI governance is heading, from someone inside both DeepMind and the Oxford AI governance ecosystem.
- A direct sense of how to move from AI pilots to deployments that survive regulatory and reputational scrutiny.
Talks
A working tour of the technical toolkit for safe, privacy-preserving, multi-owner AI, built around live demonstrations from OpenMined’s open-source project.
Key takeaways:
- How homomorphic encryption, secure multi-party computation, federated learning, and differential privacy fit together as a single stack.
- Where each technique is production-ready and where it is still research.
- What this changes about which AI products are buildable in regulated industries.
A focused session on training models across multiple data owners without exchanging the underlying data.
Key takeaways:
- Why most current AI deployments leak more than their owners realise.
- How differential privacy and secure multi-party computation reframe the data sharing question.
- The institutional architecture needed to make privacy-preserving AI an operating norm, not an experiment.
A first-principles talk on what privacy actually means in an AI economy, and what state-of-the-art tools now make possible.
Key takeaways:
- Why the current privacy debate confuses consent with control.
- The shift from data minimisation to structured transparency.
- What buyers, regulators, and engineers each need to demand of AI systems.