I spent a few days last week in Ottawa at the Global Conference on Responsible AI and Elections, hosted by the U.N. Development Programme and Elections Canada. More than 200 people from more than 60 countries were in the room. Electoral management bodies from across Africa were well represented, alongside participants from North and South America, the Middle East, Europe and Australia. Even Australia was there, and yes, they heard my Steve Irwin tribute. I’ll spare you all.
As I reflect on what I heard and learned, I’m fascinated by how the conversations we’re having in county offices in the U.S. are the conversations they’re having everywhere.
We’re all grappling with the same questions and constraints
In the United States, election officials are still working out what AI means for their offices. That is true everywhere. What surprised me is how much of the conversation is still stuck one step earlier – on what AI even means. I say this with the humility of someone who is by no means a technologist or software engineer: Even when we’re all speaking in one language (unlike at this conference), I can’t help but have the sense that we all miss so much or struggle to convey what we’re trying to communicate because we don’t have the vocabulary to do it.
Additionally, we are at an inflection point where it is very easy to conflate an existing technology with a new one. The foundations of so many of our technologies predate AI. Optical character recognition is not new. Biometric identification is not new. Pattern detection is not new. Some of those tools have been in use for decades. If we don’t strive for clarity in describing what, when, and how AI is part of the technology we risk increasing confusion. AI has made a lot of tools and technologies easier to develop, buy, and implement, and easier to talk about and imagine use cases. That is not the same as making them AI.
However, where AI has become core to the way the technology operates we should identify this clearly and not conflate the two.
The leapfrog
I think there is an operationally-focused consideration that is especially interesting in the international context.
Countries that have been able to procure and build for a long time have layers of technology already on the ground (and sometimes literally in the ground). Countries that have not are now looking at cloud computing and tools you interact with in plain language. The barrier to entry has dropped. In some cases these officials can skip a generation of infrastructure entirely.
It reminds me of mobile networks. I went to Peru last year, wrongly assuming cell phone coverage would be spotty or nonexistent. I stood at the top of Machu Picchu and got a phone call on my U.S. cell phone. I had assumed connectivity required the kind of buildout and linear digital evolution I grew up watching at home.
AI is following a similar path. The ability to do something no longer depends on having spent 20 years preparing or building the infrastructure to do it. I think that changes more than the speed of adoption. It changes where we should expect innovation to come from, and it changes how we should view and calculate risk.
An election office or electoral management body without decades of legacy systems may have fewer things to unwind, integrate or work around. They can start with the tools and infrastructure available now rather than retracing every step others took to get here. The traditional direction of learning or innovation may become less predictable and less grounded in the hard-earned lessons of the past.
Use AI to build the tool. Do not always make AI the tool
There are a lot of cases where you can use AI to deploy a conventional tool faster, and the end user never touches an AI system at all.
Take a chatbot. You can stand up an AI-powered chatbot that generates answers on the fly. Or you can use AI to read through everything voters have actually asked your office, find the real questions and build a defined, rule-based bot that answers them.
The second path gives you a fixed output that you control, which is much more popular with my security friends. It is faster to build than it used to be. It is easier to edit and expand as the cycle changes. It does not put a generative system between your office and a voter. And it does not carry the recurring cost of attaching an AI service to a public-facing tool.
(Related to this, I always come back to my friend Gretchen Reinemeyer’s example of how she used an AI tool to build out an election turnout forecast in Excel that helped inform how she would staff and resource her voting locations.)
Big picture, small picture
Macro concerns – data privacy, political consensus and public trust – shape the conversation. Those came up in Ottawa constantly and they should.
But the big picture has a way of crowding out the small one. When every conversation starts at the level of democracy itself, election offices never get to the question they can actually answer this week. Should we or can we use this thing for this task? In a digital world with more and more AI integration, what are the off ramps for non-use? Is a policy or position that disallows AI use from the frontline workforce even possible at this point?
This is reflective of the approach taken through our work at the AI + Elections Clinic. Put the tools in front of people. Have them use one. Talk about what happened. Work through a specific use case and decide whether it holds up and what the risks are. The small picture feeds the big picture.
Thank you
A big thanks to the U.N. Development Programme for the invitation and Elections Canada for the hospitality.
Originally posted on the AI and Elections Substack, September 14, 2026. Read the original ↗