Language models show human-like content effects on reasoning tasks