OpenAI just published the results of a grant program that handed $100,000 each to ten teams tasked with figuring out how the public should actually govern AI behavior. Nearly 1,000 teams from 113 countries applied, so the appetite is clearly there. But the findings from the selected groups expose just how thorny this whole democratic AI thing really is. One of the most striking insights is that public opinion on AI isn’t a stable target; many teams discovered that people’s views shifted frequently, even day-to-day. That’s a practical nightmare for anyone trying to hard-code values into a model. Do you capture a fleeting Tuesday afternoon sentiment or try to distill some deeper, harder-to-change principle? The report suggests the process needs to be both thorough and recurring, which is a tall order.
The projects themselves were a creative bunch, ranging from a chatbot that generated ‘value cards’ for participants to review, to platforms for crowdsourced audits and even mathematical formulations for representation guarantees. A common thread was using AI itself to grease the wheels of the deliberation process through custom chat interfaces and data synthesis. But the real challenges weren’t technical. They were human. Recruiting a genuinely diverse set of participants was brutal. Teams consistently hit a wall trying to cross the digital divide, and those they did recruit online skewed more optimistic toward AI, likely warping the feedback. Even when participants from global majority countries were included, the tools often fell short because of language and context limitations.
The Inclusive.AI team tackled the nuance problem by letting people distribute voting tokens across many statements instead of forcing a single binary choice, capturing not just what people think but how intensely they feel it. It’s a clever approach that hints at the sophistication needed to avoid flattening complex public sentiment into something useless. OpenAI is now taking these prototypes and planning to weave them into an end-to-end process for actually shaping its models, moving from isolated experiments to real governance infrastructure. The call is out for researchers and engineers to join the effort.
I’ll be blunt: the most honest part of this update is the admission that we’re not even close to solving the outreach problem. If you can’t hear from people who aren’t already tech-optimists with internet access, you’re not governing AI for humanity. You’re governing it for a very specific, privileged slice of it. That’s not a technical bug to be patched with better AI tools; it’s a fundamental legitimacy crisis for the whole endeavor. The code is now public, and the ideas are intriguing, but the hard part—actually listening to everyone—remains largely unsolved.
