I Let AI Vote for Me in the Nix SC Election (jackson.dev)

Faced with a last-minute Nix Steering Committee election that required ranking all 24 candidates, the author used an LLM to save time rather than read ~188,000 words of candidate statements (~7,800 words each). They encoded personal priorities into a “value set,” batched the markdown candidate files (two at a time to respect token limits), had the model summarize each candidate relative to those values into a single summary file, then asked the model to produce a ranked list. The author spot-checked several items and compared the LLM’s picks to a few candidates they already knew before submitting the ballot, treating the model’s output as an initial, verifiable draft rather than an unexamined decision. For the AI/ML community this is a concrete example of LLMs as productivity amplifiers for high-information civic and OSS tasks: practical workflow choices (data in markdown, batching to handle tokens, value-conditioned prompts, and lightweight human verification) let users scale decision-making without reading everything. It also highlights important trade-offs—value alignment, bias amplification, auditability, and the need for transparent pipelines—so LLMs are best used as assistants that produce auditable summaries and rankings, not as sole arbiters of governance choices.