Executive Summary

When AI systems cause harm, determining who bears responsibility is philosophically and practically challenging. The "many hands" problem in AI development—the distributed nature of design, training, deployment, and use—defies traditional individualistic frameworks of moral responsibility.

Drawing on the philosophy of collective responsibility, this paper identifies three key insights for AI safety:

1. The problem of many hands: AI harms often arise from distributed actions where no single individual's contribution is decisive, challenging individualistic blame assignment 2. Forward-looking collective responsibility: All actors with capacity to prevent AI harms bear some responsibility for doing so, not just those who caused the harm 3. Institutional design: We should design AI governance structures that enable collective responsibility rather than hoping individual accountability will suffice

---

Introduction: The Accountability Gap

AI systems are developed by teams, deployed by organizations, and used by individuals. When harm occurs, who is responsible?

Consider an AI system that produces biased outcomes:

  • The data was collected by one team
  • The model was trained by another
  • The deployment decisions were made by executives
  • The user interface was designed by yet another group
  • The system was used in ways not fully anticipated

Who bears moral responsibility for the resulting harm?

Traditional moral philosophy, with its focus on individual intentions and actions, struggles with this question. The philosophy of collective responsibility provides tools for thinking about accountability in distributed systems.

---

The Problem of Many Hands

Individual vs. Collective Responsibility

Traditional moral responsibility requires: 1. Causal contribution: The agent caused the harm 2. Intentionality: The agent intended or was negligent about the harm 3. Knowledge: The agent knew or should have known about the harm

In AI development, these conditions often fail to be met by any single individual:

Causal contribution is diffuse:

  • Each contributor's work is necessary but not sufficient for the harm
  • No single person's decision was decisive
  • Harm emerges from the interaction of many decisions

Intentions are fragmented:

  • Each actor may have good intentions individually
  • The collective outcome differs from individual aims
  • No one "intended" the harmful result

Knowledge is distributed:

  • Each actor has partial knowledge of the system
  • No single person fully understands the system's behavior
  • Risks may be invisible to individuals but visible to the collective

The Accountability Gap

This creates an accountability gap: there is harm, but no one seems responsible. This is not a bug but a feature of how responsibility is conceptualized.

Implication for AI safety: Individual accountability mechanisms (lawsuits, professional ethics, personal conscience) may be insufficient for AI harms. We need collective frameworks.

---

Collective Responsibility: A Primer

Backward-Looking vs. Forward-Looking

Collective responsibility has two forms:

Backward-looking: Who is to blame for harm that occurred?

  • Focuses on causation, fault, and desert
  • Challenges: Can groups have intentions? Can groups be blameworthy?

Forward-looking: Who bears responsibility for remedying or preventing harm?

  • Focuses on capacity, obligation, and remedy
  • Less metaphysically demanding
  • More practically useful

Key Frameworks

Bratman's shared intentions: Collective action involves:

  • Each intends that "we" do something
  • These intentions are common knowledge
  • Subplans "mesh" to enable coordination

Gilbert's plural subjects: Groups form "plural subjects" through:

  • Joint commitment to action as a body
  • Collective intention that transcends individuals

May's relational responsibility: Collective responsibility exists when:

  • Individuals are related to enable collective action
  • Some are authorized to represent the collective

---

Application to AI Safety

Backward-Looking: Blame for AI Harms

When an AI system causes harm, who is blameworthy?

Individual approach (limited):

  • Blame the developer who wrote the buggy code
  • Blame the executive who approved deployment
  • Blame the user who misused the system

Problems:

  • Each individual's contribution may have been reasonable
  • Harm emerges from interaction, not individual decisions
  • Scapegoating individuals may miss systemic issues

Collective approach (more accurate):

  • The development team as a collective bears responsibility
  • The organization as a collective bears responsibility
  • The ecosystem of actors (developers, regulators, users) shares responsibility

Implication: Legal and ethical frameworks should recognize collective blame, not just individual blame.

Forward-Looking: Responsibility to Prevent Harm

Who bears responsibility for preventing AI harms?

Capacity-based responsibility:

  • Those with capacity to prevent harm have responsibility to do so
  • This includes: developers, companies, regulators, users, civil society
  • Responsibility is shared, not divided

Collective obligation:

  • The AI safety community has a collective obligation to prevent harm
  • No single actor can fulfill this obligation alone
  • Coordination is required

Implication: AI safety is everyone's responsibility, but this doesn't mean no one's responsibility. It means shared responsibility requiring coordination.

---

Institutional Design for Collective Responsibility

The Challenge

If AI responsibility is collective, how do we make it effective?

Three problems: 1. Diffusion: When everyone is responsible, no one feels responsible 2. Coordination: Collective action requires coordination mechanisms 3. Enforcement: How to hold collectives accountable?

Design Principles

1. Designate collective responsibility explicitly:

Don't assume individual responsibility will aggregate to collective responsibility. Create explicit collective obligations:

  • Organization-level safety obligations (not just individual)
  • Industry-wide safety commitments
  • Cross-sector coordination mechanisms

2. Create clear accountability structures:

Even within collective responsibility, individuals have roles:

  • Safety officers with explicit responsibilities
  • Clear escalation paths for concerns
  • Designated decision-makers for key choices

3. Enable collective knowledge:

Address the knowledge distribution problem:

  • Information sharing mechanisms
  • Common knowledge creation (transparency)
  • Shared monitoring and evaluation

4. Build collective capacity:

Responsibility requires capacity to act:

  • Collective resources for safety work
  • Mechanisms for collective decision-making
  • Infrastructure for coordination

5. Enforce at the collective level:

Liability and accountability mechanisms:

  • Organizational liability (not just individual)
  • Industry-wide accountability mechanisms
  • Collective sanctions for coordination failures

---

Case Study: AI Development Lab

Consider an AI development lab as a collective responsible entity:

Backward-Looking

If the lab's AI causes harm:

  • The lab as a collective bears responsibility
  • Blame should not be deflected to "a few bad apples"
  • Systemic factors (culture, incentives, processes) matter

Forward-Looking

The lab has collective responsibility to:

  • Develop safety practices
  • Monitor for harms
  • Coordinate with other labs
  • Contribute to collective safety infrastructure

Design Implications

For the lab to fulfill collective responsibility:

  • Explicit safety commitments at organizational level
  • Clear internal accountability structures
  • Transparency with external stakeholders
  • Resources allocated for safety work
  • Mechanisms for collective decision-making

---

Objections and Responses

Objection 1: Collective Responsibility Lets Individuals Off the Hook

Response: Collective responsibility does not replace individual responsibility—it supplements it. Individuals within collectives still have obligations. But focusing only on individuals can miss systemic factors and create accountability gaps.

Objection 2: Groups Can't Have Intentions or Moral Agency

Response: Two responses: 1. Forward-looking responsibility doesn't require collective intentions—it requires capacity and obligation 2. Shared intention frameworks (Bratman, Gilbert) show how collective intentions can exist without positing mysterious "group minds"

Objection 3: This Is Too Abstract—How Do We Implement It?

Response: Implementation includes:

  • Organizational liability regimes
  • Professional collective obligations (like medical boards)
  • Industry-wide safety agreements
  • Regulatory frameworks that recognize collective responsibility

---

Implications for AI Safety Governance

For Labs

  • Recognize organizational responsibility beyond individual ethics
  • Build collective decision-making and accountability structures
  • Contribute to industry-wide coordination

For Regulators

  • Design liability frameworks that include collective responsibility
  • Create mechanisms for industry-wide accountability
  • Enable collective action while maintaining individual accountability

For Researchers

  • Study how collective responsibility works in practice
  • Develop metrics for collective safety behavior
  • Design mechanisms that make collective responsibility effective

---

Conclusion

AI safety presents a problem of collective responsibility. Harms arise from distributed actions where no single individual is clearly to blame. Traditional individualistic frameworks create accountability gaps.

The philosophy of collective responsibility provides tools:

  • Forward-looking responsibility based on capacity, not just causation
  • Shared obligations that require coordination
  • Institutional designs that enable collective accountability

The challenge is not to choose between individual and collective responsibility, but to design governance frameworks that include both. Individual ethics matters. But so does collective accountability. AI safety requires both.

---

References

  • Bratman, M. (1999, 2014). Faces of Intention; Shared Agency.
  • Gilbert, M. (1989, 2000, 2006). On Social Facts; Sociality and Responsibility.
  • Lewis, H. D. (1948). Collective Responsibility.
  • May, L. (1987). The Morality of Groups.
  • French, P. (1998). Collective and Corporate Responsibility.

---

This paper draws on the Stanford Encyclopedia of Philosophy entry on Collective Responsibility.