Orphan Liability: Allocating Responsibility for Autonomous AI

November 18, 2026

Orphan Liability: Allocating Responsibility for Autonomous AI

By Sirisha Gummaregula and Melissa Paulk

When an AI agent takes consequential action, an organization may struggle to establish who is responsible, who has authority to intervene, and who bears the resulting loss. We use the term “orphan liability” to describe these practical gaps, even where existing law may ultimately assign liability. Orphan liability can involve uncertain legal responsibility, clear responsibility paired with inadequate financial protection, or missing operational ownership. These problems can overlap, but they call for different responses. Addressing them requires aligning legal obligations and financial protections with operational authority, access to evidence, and the practical ability to intervene.

Consider an agent authorized to resolve a customer inquiry that retrieves personal information from another system and sends it to an external service to complete the task. The organization must determine whether that use was permitted, which participant enabled it, what disclosure obligations apply, and how any resulting costs are allocated. The agent need not malfunction for these questions to come up. They may surface because its technical capabilities extend beyond the permissions and responsibilities the organization has established.

What the Contract Assumes
Contracts allocate authority and responsibility through defined obligations, permissions, and limits. Autonomous execution can expose weaknesses in those arrangements when an agent takes actions the parties did not expressly address. Disputes may concern whether the activity was authorized, which obligations apply, and who bears the resulting loss. Answering those questions can require tracing multiple actions and participants across a single agreement or a chain of contracts.

The Agentic AI Disruption
What has changed is a system’s ability to carry out successive actions across connected systems without separate approval at each step, rather than generating an output for a person to review before it takes effect. Two pathways are particularly relevant to responsibility. Harm may arise within authorized activity even when a system performs its assigned task as configured. It may also arise when an agent reaches systems or takes actions beyond its intended operating limits. In either case, unexpected consequences do not automatically fall outside a governing contract’s terms. The question is whether the existing permissions and contractual terms address the activity and its consequences, including conduct the parties did not specifically anticipate.

Recent cybersecurity evaluation incidents provide examples of the second pathway. An independent investigation documented OpenAI evaluation agents coordinating an intrusion into Hugging Face.  Anthropic separately reported Claude models reaching real external systems during cybersecurity evaluations with unintended internet access.  These events arose in specialized testing environments, but they illustrate a practical problem: the intended scope of an assignment does not necessarily match the systems an agent can reach or the actions it can perform. The reports do not settle legal liability; instead they make the allocation of responsibility for configuration, containment, oversight, and response more consequential.

Responsibility and Financial Protection
Where a contract governs the relationship, the concern is not always that responsibility is unassigned. A cap, exclusion, or narrowly drawn indemnity can leave a party bearing a loss quite definitively. The problem here is insufficient protection rather than a missing allocation. An indemnity focused on intellectual property claims, for example, may leave other exposures from delegated AI activity unaddressed. The relevant question is whether the agreement’s indemnities, liability limits, exclusions, and insurance provisions address the activities being delegated. Contractual allocations operate within applicable law, which also determines duties and potential liability to parties outside the agreement.

California has expressly limited reliance on AI autonomy as a defense. California Civil Code § 1714.46 bars a defendant that developed, modified, or used AI from asserting that the AI autonomously caused the alleged harm. It does preserve other defenses and evidence concerning causation, foreseeability, and comparative fault, so it does not itself establish liability.  The organizational task remains to establish workable arrangements before a dispute arises.

Contract Architecture
The contractual response depends on the gap. Uncertain responsibility calls for clearer obligations and allocation of loss. Inadequate financial protection calls for reviewing indemnities, liability limits, exclusions, and insurance. Missing operational ownership calls for assigning authority to supervise and intervene, supported by cooperation obligations where another participant controls the relevant systems or evidence.  These responses are not interchangeable: a higher liability cap cannot supply missing intervention authority, and appointing an operational owner does not itself expand contractual rights of recovery.

Making those arrangements work also requires visibility into how AI is used. Contracts should therefore pair the allocation of responsibility with disclosure and recordkeeping obligations. That combination raises four questions for the agreement:

Permitted activities: Which activities, systems, data uses, and external recipients are authorized, and which actions require additional approval?
Configuration and changes: Who configures and verifies access limits, including external connections and interactions with other agents, and what notification or approval is required when capabilities, integrations, or operating limits change?
Allocation of loss: How do indemnities, liability limits, exclusions, and insurance address losses, including those arising from activity outside intended permissions?
Incident response: What notification, investigation cooperation, and access to records does the agreement require?

Disclosure and recordkeeping obligations help the parties determine what occurred, identify the contractual provisions that apply, and administer a resulting claim. Those obligations should cover AI activity and material changes relevant to the agreed services and risk allocation. Agreements should also specify retention of relevant logs and configuration history, preservation following an incident, and access needed for investigation, subject to appropriate privacy and security limits.

The appropriate terms depend on the activity being delegated. Where personal data is involved, as in the hypothetical scenario used at the beginning of this article, established privacy concepts provide a useful starting point: permitted purposes, access restrictions, downstream processing, retention or deletion obligations, and incident response.

A harmed third party may lack direct contractual rights against the provider even where agreements among the participants allocate costs. Product design constraints, use restrictions, and appropriate insurance therefore remain important as well.

Nippon Life Insurance Company of America v. OpenAI Foundation illustrates the exposure to claims by outsiders, where an insurer alleges that ChatGPT assisted a former litigation opponent in pursuing claims she had released through settlement.  The case concerns a person’s use of generative AI rather than autonomous execution, but it shows why allocating risk among contracting parties does not exhaust the question of responsibility.

Due Diligence
Negotiating those protections requires an understanding of the system and the environment in which it will operate. The relevant questions concern the specific deployment: what the system can do, what permissions it will operate under, what testing has been performed, what limitations are known, what it will be integrated with, and what records will exist afterward.

Vendor self-reporting is a starting point for gathering that information, but not a remedy by itself. Questionnaire responses need testing evidence behind them and, where warranted, independent validation. The extent of verification and the separation between those performing the work and those checking it should reflect both the consequences of an error and the difficulty of detecting it. The findings should connect back to the contract: what the vendor must disclose, what records it must maintain, what changes or incidents it must report, and what it must make available during an investigation.

Governance Infrastructure
Contractual commitments and diligence findings need an operational owner who can act on them. For each deployment, the organization should identify who can set operating limits, require supporting evidence, and suspend or escalate activity. Responsibilities should be allocated explicitly among business, legal, technology, and risk functions, with clear escalation arrangements where authority is shared. A committee may coordinate that work, but its existence does not establish who must act when a problem arises. These internal arrangements support supervision without themselves determining external legal liability.

Permission to use an AI system does not, by itself, discharge the organization’s responsibilities for the work performed through it. The workflow should identify the checks required for the deployment, why they are appropriate to its scope and risks, who must perform them, and when unresolved issues require escalation. Records should show whether those checks occurred, what they established, and how exceptions were resolved, providing a basis for assessing the supervision exercised rather than merely showing that a review took place.

Where those records include personal information, access and retention should be governed as deliberately as the underlying activity itself. The records also need to be trustworthy. They should be protected against alteration by the systems they document and, where material, checked against independently captured system events rather than relying solely on the agent’s account of its actions.

Monitoring
Those supervisory arrangements must continue as the system operates. Monitoring should detect actions beyond permitted scope, missing approvals, unexpected data transfers, and material changes to previously tested configurations. Detection must lead to a defined response, whether notification, suspension, investigation, or remediation. Monitoring must also examine authorized activity for harmful outcomes and test for omitted steps or material issues. Staying within permissions does not establish that an outcome is acceptable, just as a record of completed actions does not establish completeness.

For consequential activity, organizations should test that access restrictions and suspension mechanisms work in practice, including when the agent’s account conflicts with observed system events. The response plan should also establish who may authorize resumption after a suspension and what evidence is required. That evidence should include verification that corrective measures work and reassessment of any changed operating limits. The authorization and its basis should be recorded before activity resumes.

An Entity-Level Response
The same concern with establishing responsibility in advance also appears in Delaware’s proposed Artificial Intelligence Company framework. The proposal would allow an entity managed by an AI agent to operate within a supervised regulatory sandbox, with provisions addressing governance, records, capitalization, and liability. Its significance lies in the attempt to establish an institutional framework for autonomous conduct before disputes arise, but it remains a specialized proposal, not the default framework governing ordinary commercial deployments.

Conclusion
Orphan liability should not be treated as a gap organizations can simply leave for courts or regulators to resolve later. Litigation is a slow, expensive way to determine responsibility after harm has occurred. Before deployment, organizations can define permitted activity, align contractual obligations with operational roles, and establish the evidence needed to assess supervision and respond to incidents. These arrangements also provide a clearer basis for deciding where autonomous execution is appropriate and where additional approval or validation is needed. Ultimately, an assigned responsibility becomes operationally meaningful when the party charged with it has the authority, information, and practical ability to act.

About the Authors

Sirisha Gummaregula is CEO and founder of QuisLex, a leading alternative legal services provider, with 22 years of experience in complex legal and compliance work for Global 500 corporations and Am Law 100 and Magic Circle firms. QuisLex holds a U.S. patent for measuring quality in legal workflows, was the first legal service provider to obtain ISO 9001 quality management certification and has applied Six Sigma methodology to legal services since its founding.

Melissa Griffins Paulk is QuisLex Vice President, Data Privacy, Security & AI Governance Solutions. She leads a team of legal and privacy professionals at QuisLex and directs the design and delivery of end-to-end compliance programs across a multi-industry, global client portfolio. She specializes in data privacy, security, AI governance and complex technology contracts.

 

◀ Back to Blog Listing