Working Group Progress

P7006 Work in Progress Guiding Principles

The work in progress draft principles are created by the members of the P7006 working group with the objective to guide the more detailed work on the standard. The principles are work in progress non-normative guidelines and should be viewed as such.


  1. Agent Duty of Care
    1. Human-interpretable transparency
      1. “Black box” processes must be able to be tracked in a way that creates a human-readable chain of custody for all decisions.
    2. Interruptible & Retractable decision-making
      1. A “Kill Switch” shall be incorporated into root-level Agent design so that is is possible for a User to halt and.or reverse the actions of an agent.
    3. Fairness of Agents
      1. All agents should be designed in a way that ensures the highest level of equality among agents.
  2. Agent: Security
    1. Verifiable security
      1. The publication of auditable cryptography across the full stack shall be accessible to Users and non-Users
    2. Reduction of Potential Thread Surfaces
      1. User-specific I/O data used by the Agent to make decisions shall be stored in a manner where the breach of any one system will not lead to a full system compromise.
    3. Notification of system compromise
      1. An Agent or its issuer must notify users upon discovery of a system breach of any system, subsystem or system-dependent network architecture.
  3. Agent: Integrity
    1. Authentication, Appropriation and Authorization
      1. The Agent must provide a User with information regarding any third-parties (including but not limited to ad networks and content intermediaries) that have influenced a decision.
    2. Technically-feasible replicability
      1. The decision-making rationale for each Agent output should be able to be recreated.
    3. Recursive instances
      1.  All copies of agents should respect the same principles defining the original agents.
  4. User: Rights
    1. Human-rights-based architecture
      1. The agent reflects and protects individual rights and freedoms enshrined in the UN’s Universal Declaration of Human Rights.
    2. Assumption of liability by agent
      1. Autonomous decisions are the liability of the agent that can therefore only enter into contracts or agreements with the express (and re-authorized) consent of a user.
    3. Privacy-oriented dynamic UI/UX
      1. A user can easily toggle Agent privacy settings to enable updating system settings to facilitate differential privacy depending on the occasion.
  5. User: Control
    1. Data interoperability & portability
      1. Users shall be able to retain control over portions of the data that has been processed by their Agent for use outside the application of the Agent.
    2. Human extension
      1. The Agent’s agency is constrained to the decision-space allowed for by the human, and, while making autonomous decisions, it is never fully independent of the individual.
    3. Arbitrations mechanisms
      1. Every agent shall have legal and non-legal recourse that gives users the ability to appeal decision-making attributed to the Agent
  6. User: Safety
    1. User-centered agency
      1. The decision space of an Agent shall be constrained to actions that have been expressly consented to by a User.
    2. Agent shall do no harm
      1. An Agent shall not make any decisions that brings physical, emotional or cognitive harm to a User or non-user.
    3. Safeguards from law enforcement
      1. The Agent shall not be compelled by law enforcement (under law or extrajudicial means) to reveal compromising information about a User.
  7. Non-User:
    1. Ethical Due Diligence
      1. Impact assessment of potential Externalities should be carried out on a recurring basis.
    2. Mitigation of data capture
      1. An Agent shall not collect superfluous data on Non-users without express Consent that authorizes PII being captured
    3. Privacy-by-design for non-user and minors
      1. Non-users and Users under the age of legal consent shall have no data stored about their interactions with an Agent by default.