I can help fellows think beyond "how do we use AI?" to "what changes because AI exists?" My work sits at the intersection of organising, campaigning and emerging technology: testing where AI actually improves political work and where it creates new risks, with a focus on keeping humans in the loop rather than replacing political judgement.

Through Campaign Lab I've supported a wide range of experiments: campaign training simulations, hyperlocal chat interfaces, social listening at scale, and agentic experiments. I also train campaigners in AI tools and help organisations write AI policies.

In July we ran the AI & Campaigning Conference, stress-testing persuasion bots, synthetic focus groups, agent swarms, content tools and canvass training in one room. The lessons form a good map of where political AI actually is:

• AI isn't replacing campaigners - it's changing what they spend their time doing. Almost nobody wanted AI to replace campaigning; the question was where AI makes humans better at it.
• Listening may be more valuable than persuading. If AI helps politicians hear more people rather than simply speak more efficiently, that's a very different vision of political technology.
• Agents still need humans and AI can only act on what's legible. Our agent-swarm demo only worked once a campaigner calibrated it, and at one point became comically obsessed with verifying whether an Oxfordshire stream actually existed. If something isn't clearly documented online, AI will struggle with it which matters enormously for hyperlocal politics, where much of what's important lives in people's heads rather than on the internet. Domain expertise and collaboration between campaigners and technologists still matter.
• Synthetic voters are useful, but they're not voters. Incredibly useful for exploration; dangerous if treated as evidence.
• More content isn't better communication. When everyone can publish ten times as much, quality becomes even more important.
• Training people may be one of AI's biggest opportunities simulated voters building volunteer confidence before real doors, not replacing organisers.
• When everyone can build, value shifts from writing software to orchestrating workflows, designing systems, and knowing when to move fast versus when infrastructure demands care.

I'm particularly useful if you're trying to answer questions like: What campaign functions should be automated, and which never should be? How do you run ethical political experiments and test tools safely in real environments? I'm interested in political products that increase participation, responsiveness and capacity rather than simply optimising persuasion and I keep a watching brief on how these capabilities are used for more problematic ends.