Post

Will AI give us a shared basis for political knowledge, or will Democrats and Republicans end up asking increasingly different chatbots? New preprint: We find a self-reinforcing loop that can pull AI models apart politically. We call it the Centrifugal Alignment Spiral 🧵 arxiv.org/pdf/2609.28486

The social media era made us worry about echo chambers, partisan sorting, and the loss of shared reality. LLMs looked like the opposite: a few general-purpose systems synthesizing across perspectives. But these models learn from their users. That opens a new route to fragmentation. ꩜

The spiral needs three things: 1) politically different users pick different models 2) models learning from their users' feedback drift apart 3) those differences then shape who picks which model next

Step 1: people already sort. We ran an experiment with 1,884 Americans. Republicans were more likely to choose Grok and Democrats Claude, even when paid for accurate answers. 71% returned to a model they'd preferred before.

Step 2: We trained identical copies of three open models on feedback from 90% Dem vs 90% Rep audiences (a deliberate stress test). On unseen political questions, paired models chose different answers 12–41% of the time. Every run moved in the expected partisan direction.

Even when we removed an entire issue area from training, some models still diverged on it along partisan lines. Partisan feedback seems to teach models something broader than the specific answers it rewards.

None of this is inevitable. Pooling feedback across groups cut divergence from 32% to 4%, and stronger regularization dampens it. But if different companies attract different publics, no single firm can fix this alone.

More broadly, we must recognize that model alignment is emerging as an important political arena. We cannot leave it to market share, smoke-filled rooms, or opaque feedback pipelines. It must be made subject to democratic deliberation.