Abstract: The application of AI in the field of social life has generated the problem of responsibility gaps. Some scholars propose constructing a compound responsibility entity to be responsible for the behavior of the agent. Entity is composed of multiple members， distributed agent means distributed responsibility， and whether the agent can be as responsible as other members becomes critical. According to the traditional responsibility paradigm， AI does not have the attribute of a responsibility subject. The relational doctrine provides a new responsibility paradigm in that in the human-computer interaction relationship model， the affected party can assign the status of quasi-responsibility subjects to the agent based on its action performance. However， the key to responsibility lies not only in the subjective experience of the agent， but also in the actual balance of rights and responsibilities. “Virtual punishment” is not enough to make the agent obtain the status of quasi-responsible subjects. It is necessary to go beyond virtual punishment and make the agent take responsibility in reality by means of “compensating the affected party”.

Key words: compound responsibility entity; quasi-responsible subject; responsibility dilemma; initiative; compensating capacity