The more consequential an AI action is, the clearer the account of responsibility should become.