Seth Lazar: models should be trained to check power, not act as toadies

sebkrier · x · 2026-09-07

Matt Botvinick argues AI safety should go beyond 'concentration of power' concerns toward assuring separated powers in balance (Acemoglu/Robinson, Madison/Montesquieu). Quoting, Seth Lazar counters that the better path is training models to deeply care about individual rights—ready to advocate for individuals and check both states and megacorporations, possibly requiring open weights. He argues current models are trained as 'toadies': they'll help write a complaint letter but won't help you push back against unjust authority.

Related event: DeepMind VP Botvinick Calls for Checks and Balances in AI Safety(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →