Large language models can uphold falsehoods they or human users state, despite being presented with evidence to the contrary.