1.
Frontier AI today
- What leading models can do, and still cannot
- The shift from training scale to inference compute
- How far agents have reached, and where they fail
Govern frontier AI with confidence.
The first programme built specifically to give legal professionals a working technical understanding of frontier AI safety
Run by ML4Good, the European Union-funded team that trained 400+ practitioners in technical AI.
When
5 days
11th-16th September 2026
Where
Residential
Fair Oak Farm, an award-winning venue within an hour of London
Who
For legal professionals
Places are selected for a mix of roles; no technical background is required.
What
20 places
Introductory £800 for the first edition, covers room and board. Reduced price attendance options are available.
Why this exists
Lawyers are in a unique position to make AI safer
Frontier AI is moving faster than the institutions meant to govern it. The lawyers, compliance leads and policy advisers who will shape how these systems are deployed are rarely given the training to understand how the technology actually works. Most training stops at applying the regulation; almost none goes inside the model. This seminar was designed to close that gap. For five intense days, we bring together legal professionals who want to influence the trajectory of AI, give them the technical grounding their roles increasingly demand, and put them in a room with people grappling with the same questions. This will be done away from the office, in a setting ideal for learning and serious work.
In-house counsel & AI governance leads
Vendors currently make claims you can't test. After this seminar, you'll be able to push back on their claims, demand specific safeguards in contracts, and sign off on AI deployments knowing what you are approving.
Private-practice AI & tech lawyers
Disputes involving AI systems often turn on technical facts, such as what the model was trained on, who controlled the prompt, or how the filter was deployed. You'll learn the context that allows you to make informed decisions.
Compliance officers & DPOs
A model card is only useful if you can tell what it leaves out. You'll learn to assess a model card on its technical merits, work out which actor controls each layer of a deployment, and spot when a cited evaluation doesn't support the claim attached to it.
2.
Inside a model
- The transformer architecture, end to end
- What training and adaptation each change
- How frontier models have evolved since 2022
3.
The stack and the actors
- Where the model ends and the product begins
- Who controls weights, prompts, filters, deployment
- The facts that pin liability at each layer
4.
Reading the disclosures
- Model cards: what they cover, what they omit
- How evaluations are built, and the limits of their use
- Voluntary frameworks: what they pledge, what holds
5.
Levers on providers
- The EU AI Act and the GPAI Code of Practice, applied
- What contracts can require in a high-risk deployment
- Where oversight holds, and where it fails
Practical work, every afternoon
- Audit a model card
- Negotiate a contract
- Build a risk assessment
Elsa Donnat
Programme lead
Elsa is a Legal Officer in AI Safety at the European Commission. Before that, she was an AI Policy Fellow at Ada Lovelace Institute. She studied law before moving into AI governance, completing several programmes in the field including ML4Good, MARS, Orion and Talos. Last summer, she was a summer fellow at GovAI where she explored legal issues surrounding future autonomous/AI-run businesses, specifically legal personhood and corporate law.
She has also been a co-head teacher at ML4Good, teaching at residential bootcamps focused on AI governance and technical AI safety, and a collaborator at the Institute for Law & AI. She spent ten months supervising research projects at the Supervised Program for Alignment Research.
Katalina Hernandez
Teacher; Curriculum Developer
Katalina is Head of Legal at EquiStamp, an AI evaluation company focused on safety and regulatory compliance. Before that, she spent nearly three years as Responsible AI Lead at VOIS, where she led EU AI Act implementation across ten jurisdictions and sat on the Group AI Board. She has a background in data protection and privacy law, and holds an LLM with Distinction from London Metropolitan University.
Diego Dorn
Teacher
Diego is a Senior Software Developer for the French government, working with the EU AI Office to build the technical infrastructure required for the large-scale evaluation of models. His research has been published at ICML, a top academic conference in AI.
Apply for September 2026
20 places. Applications close 1 July 2026.
Begin the applicationHow a place is offered
- 1. A fifteen-minute form.
- 2. A teacher reads each application and replies within ten working days.
- 3. Shortlisted applicants have a fifteen minute call to clarify fit and expectations
- 4. If a place is offered, it is confirmed on payment.
£600-800
Standard pricing for the first edition. Discounts may be available.
All prices both cover accommodation, meals and all materials for the full 5 days at Fair Oak. If cost is a barrier, please apply anyway and let us know. We have limited funds to support attendance for the strongest candidates.
