Meta Oversight Board finds top AI models less likely to criticize repressive regimes