New language-learning algorithms risk reinforcing inequalities, social fragmentation, per U-M study