Epistemic interventions are a promising way to reduce power concentration

Various people are sceptical of epistemic interventions to reduce power concentration.

Here is a half-hour note sketching out what I think the bull case here is:^[1]

Base rates suggest this will matter. Historically, control of the information environment has played a significant role in backsliding^[2] and coups^[3]. So the starting expectation should be that it’s going to be significant
There are reasons to update the base rates upwards:
- AI will probably significantly raise the ceiling on how good epistemics can get, which should make this a more important lever than historically
- I think it’s likely that good and bad epistemics will be self-reinforcing, more so with AI than they have been historically. Which should also make this more important than previously
‘Everyone correctly understands their own interests’ effectively blocks several important ways to backslide^[4]
- Makes it hard to get a plebiscitary majority or a legislative majority (unless power concentration is actually in the interests of the majority)
  - Still possible for the executive to be so disproportionately powerful that it can just present things as a fait accompli in spite of unpopularity—but then there’s a question of why people have failed to anticipate this and coordinate against it
  - Still possible for the executive to collude with powerful private actors against the public interest
‘Everyone correctly understands their own interests’ empowers other branches of government to resist power concentration
- The courts and the legislature are responsive to public opinion
- E.g. if everyone could see that e.g. a better funded Congress was in their interests, then it’s much more likely that Congress would use its legislative powers to vote itself more funding
- E.g. if everyone could see that some executive action was power-seeking, then it’s much more likely that the courts stand up to that
‘Everyone correctly understands their own interests’ probably isn’t actually necessary
- Some kinds of epistemic uplift you only get the benefits if everyone is uplifted. E.g.
  - Lots of the societal benefits of provenance tracing, rhetoric highlighting, reliability tracking come from widespread trust in and use of those tools
  - Some of the paths to impact from coordination tech involve most of civil society coordinating using AI tools
- But other important kinds you only need some important people to be uplifted. E.g.
  - Automated OSINT for journalists
  - If only academics/lawyers/economists were using things like provenance tracing, rhetoric highlighting, reliability tracking, that could still enable them to use their political and advisory capital to much greater effect
  - Similarly, automated superforecasting and scenario planning could help if only adopted by influential elites (though there are also dual use concerns here)
  - Angels-on-the-shoulder type tools like reflection scaffolding and guardian angels could be extremely consequential if adopted by a handful of key decision-makers
Epistemic interventions are unusually tractable
- You don’t need legislation, you can just go and build stuff
- Adoption is hard, but as above, there are important cases where narrow adoption is sufficient
Epistemic interventions are approximately uncorrelated with other intervention classes, so they make the overall portfolio more robust
- Lots of interventions route through a) government or b) lab policy. These are super important, but also all or a)/b) might fail if the environment for that kind of policy remains/becomes unfavourable

Thanks to Ben Stewart and John Bridge for scepticism that prompted me to write this and useful comments; and to Owen Cotton-Barratt, Abra Ganz and Oly Sourbut for comments. I haven’t edited the text in light of comments, but sharing as is because maybe it will prompt more useful discussion.

^
NB I’m leaning into the bull case in the expectation that others will represent the bear case, rather than trying to reach an all-things-considered take in this doc.
^
Boese et al find that freedom of expression and civil society freedoms are usually the first things to go (though other scholars point to different ordering:
- Levitsky and Ziblatt, How Democracies Die (2018): capturing the referees (like courts and law enforcement), then hampering the opposition (through bribery, corruption and lawfare), then changing the rules (the constitution and the electoral system)
- Sato et al: undermining horizontal accountability (separation of powers), then diagonal accountability (media freedom and freedom of speech), then vertical accountability (elections)
^
Singh argues that coups are basically coordination games, and control of the information environment is key.
^
Ways to backslide, from Riedl et al:
- Legislative capture, when there’s a strong party. Examples: India, Turkey, Hungary
- Plebiscitary overrides, when there’s a populist president. Examples: Venezuela
- Executive powergrabs, when opposition parties and institutions are weak. Examples: Tunisia, Brazil
- Elite collusion, when parties and civil society are weak and state capacity is low. Examples: Indonesia, Guatemala