Insights
The official theme for this year’s Code for America Summit was “The Future We Build.” Every panel, lightning talk, and keynote that took place on the mainstage encouraged us to look forward, to embrace that future with hope and optimism, and to take an active role in building it. It’s a good message, and the emphasis on hope is certainly needed; civic tech has been put through the wringer in the past year and half. We are stressed, burnt out, and eager to move into a future where we can make bigger moves to improve government for the people it serves.
The unofficial theme was the future of AI. No one was surprised to see AI featured at the conference, but I’m not sure any of us expected to see it to this extent; it was in nearly every keynote, lightning talk and breakout session, often presented with unquestioning enthusiasm, and focused on the “more” that it could accomplish. More with less, and faster than before. But in every hallway conversation or chat over lunch, attendees voiced deep concern about how fast it’s being pushed, and what it could lead to when used with more optimism than caution. Everyone seemed to be asking similar questions: How do we use AI in ways that benefit the people we’re trying to serve, but don’t cause further harm? Blue Tiger’s recent Humans First zine points out: “Efficiency without dignity is just bureaucracy at machine speed.” We’re watching systems, tools, and institutions be broken, and we know that it comes at a human cost. Those in civic tech build, fix, and push for improvement. We do the hard work of getting things done the right way, not just the fast way, because we know and care that the consequences can save or ruin lives. Poorly trained AI without enough oversight could lead to benefits denied or delayed when they’re needed most. We’re hungry for conversations that acknowledge these risks and tackle them directly.
Thankfully, some breakout sessions showed that at the state level, folks are working this out.
My favorite came from the Minnesota Department of Human Services on their work to deliver multilingual services. They opened by stating that they didn’t want to assume that the newest and most talked-about tools would necessarily be the best ones for them: “We’re a prickly little thorn in the side of tech determinism.” It was encouraging to start without the assumption that AI can and should be doing everything.
What followed was an in-depth presentation on all their considerations in choosing which path to take.
- What were the legal risks?
- What communities could they serve with each option?
- Who would benefit, and who could it harm?
Ultimately, they settled on a combination of AI and human expertise, leveraging their Enterprise Translation Office to train AI on each target language, and ensure human oversight on everything the LLMs produce. They eventually cut the time spent on each translation task in half, and are already working on scaling it to include other languages and types of content. Their work is impressive, but the candid description of how they got there was even more appreciated.
They didn’t shy away from describing the LLMs’ risks or weaknesses. Their language experts were open about the challenges of teaching an LLM both low context and high context languages, the nuances of languages that rely heavily on metaphor, and the wildly different grammatical structures some languages use. They even shared the story of an early iteration of the Hmong language LLM they had been confident in, but that failed badly when tested with native speakers, and how they went on to shift their approach.
Most refreshing of all, they acknowledged that AI has as much potential for harm as it does for help, and that when harm is done, it’s often the most vulnerable that suffer. They talked us through their thought processes, successes, failures, how they mitigated risks and made choices. That is the hard, often thankless work that will never make headlines or be featured on your favorite podcast, but makes people’s lives better.
All this to say: We need more open conversations like this. Many of us aren’t rejecting AI, we just want to use it like we would any other tool. With precision, in ways that have been tested, and found to do no harm. We want to take the time to talk openly about pros and cons, risks and benefits. Regardless of the form it came in, I’m grateful that Summit gave us a space to start having those conversations and learning from one another’s experiences.