We just got our election ballots.
I had Claude spend a few hours researching ~100 sources and putting together an election guide for me and any friends/family who happen to share similar values. It did a pretty decent job.
It scanned not only news sources and candidate websites, but also dozens of local and national group endorsements, FEC and state campaign finance data, dark money monitors, ballotpedia, etc. Then it found a bunch of podcast interviews and videos of candidate forums and downloaded and transcribed all of them.
The result was a ranked list, based on values I gave it, of each measure and candidate. Each was weighed both against how well they align with the values and how likely they were to cause a spoiler effect in close races (the local Democrat vs Progressive split here is pretty strong, with a united GOP profiting off the divide as always). Personalized (human) judgment calls (like “labor” vs “climate” divides) were asked for where they would break a tie among otherwise good candidates.
It’s the kind of research I used to do by hand over many hours… I’d go to a bar and spend a few hours researching each candidate or measure in depth, taking a sip now and then… “once sober, once drunk” style 
It was exhausting.
Now Claude does it in a few prompts (minus the drinking), and the methodology could be easily replicated across local & national races everywhere…
Today it seems to work well, but this also creates a single-point-of-failure that Anthropic (or any big AI provider) could completely co-opt in order to subvert democracy in the future. That strategy worked well enough even back in the pre-AI, Facebook days; it’s scary to think how it will affect (what’s left of) democracy in this cycle and the next.
I don’t know how you would be able to easily counter that. Maybe having a mixture of open-weight Chinese and European models review the American outputs? Presumably each region’s models would have different forced guardrails and “cultural values” (trained-in censorship or behaviors) and this sort of cross-referencing might be better able to find gaps in the American models’ coverage and reasoning. Hmm… 