AI for Democracy
Discussion
Audrey Tang: 'a shift from humans in the loop of AI to AI in the loop of humanity'
Georg Diez and Felix Beer:
"Reinventing democracy for the AI era begins with an understanding of what people could become able to do together. Some possibilities deepen familiar capabilities, such as deliberation. Others extend capabilities into new domains, such as aligning AI systems through democratic input, or exercising them through new means, such as representing humans through agents in political processes.
Facing these changes, Audrey Tang calls for a shift from humans in the loop of AI to AI in the loop of humanity. Her ambition is a superintelligent democracy, in which AI strengthens rather than replaces people’s capacity for self-government. The danger of algocracy arises when algorithmic decisions replace or even restrict meaningful human self-determination. Tang proposes a demanding test: AI assistance should leave communities better able to listen, deliberate and act together, building capabilities they retain even when the system is withdrawn.
Emerging initiatives show parts of the infrastructure this requires.
Three possibilities illustrate how technical advances and institutional invention could work together.
Listening at scale. AI can help institutions take in and make sense of far more public knowledge than before. In Engaged California’s process on Los Angeles wildfire recovery, experts used generative AI to turn 1,289 resident comments into 19 policy options. Residents then discussed and ranked these options, and the resulting action plan was published together with the government’s responses. Yet as Beth Noveck argues, every summary decides which signals survive, so compressing public input is an exercise of power over representation. The design space here is listening infrastructure. Such systems would draw on many sources, compress input transparently, preserve disagreement and minority experience, and connect what is heard to authorities obliged to respond.
Governing AI together. People need to shape the AI systems that increasingly shape them. Weval, the open evaluation platform of the Collective Intelligence Project, allows civil society groups and domain experts to write and run evaluations that reflect their communities’ concerns. Building on such tools, Evan Hadfield proposes standing assemblies whose deliberations produce revisable principles. These principles would both guide AI systems and serve as the standard against which their behavior is judged. The design space here is democratic alignment infrastructure: public procedures for specifying what AI should do, testing whether it does so, and connecting the findings to institutions able to require a response.
Acting through representatives. AI agents could help people participate where time and attention are scarce. Andrew Sorota argues that a good representative should not simply mirror a person’s current views. It should also help them encounter other perspectives and revise their judgments, while carrying their authorized positions faithfully to institutions and other agents. That fidelity must be testable over long interactions and chains of delegation. As agents act in higher-stakes settings, trustworthy agents also become a condition of inclusion. People who cannot trust these systems risk being shut out of an increasingly important channel of participation. The design space here is representation and verification infrastructure. Agents would be tied to verified persons, hold limited and revocable mandates, and keep auditable records, so that deploying more agents never multiplies anyone’s standing.
Across all these cases, the democratic commitments that might form a kernel apply together: listening infrastructure serves popular representation; democratic alignment extends checks and balances to AI; revocable mandates preserve the reserved power of the people. In each case the tool alone decides little. Its democratic value depends on the layers beneath it: on rules that oblige an authority to respond, procedures that make its outputs contestable, and institutions able to act on what it produces. The design opportunity lies in building these connections between emerging tools and enduring institutions. That work involves choices which the operating-system lens can make visible but cannot settle."
(https://informationaldemocracy.substack.com/p/operating-systems-of-democracy?)