IMDA Issues Discussion Paper on Allocation of Legal Responsibility for AI Agents

Executive Summary

In May 2026, the Infocomm Media Development Authority of Singapore released a discussion paper examining how liability should be attributed when Artificial Intelligence (“AI“) agents act autonomously, use tools, interact with third parties, and cause harm.

Drawing on input from a working group comprising over 20 members of Singapore’s legal community, including Rajah & Tann Singapore’s Head of Technology, Media & Telecommunications, Rajesh Sreenivasan, the paper seeks to develop a preliminary understanding of agent liability in private law to support confident adoption of agentic AI and anticipate shifts in accountability as autonomy grows. It focuses on civil liability and does not cover criminal, data protection or other regulatory mechanisms.

This Update examines the liability frameworks discussed and the implications for businesses using agentic AI in Singapore.                                                                          

Key Features of Agentic AI

While there is no consensus on what defines an AI agent, the paper has identified three core features of agentic AI that are particularly relevant to legal liability, namely: (i) autonomy; (ii) decision-making; and (iii) action-taking.

These features reduce human involvement, thereby diffusing accountability and increasing the risks of unexpected outcomes. The breadth of actions that an agent can take also amplifies the potential harm that could be caused. These dynamics are further complicated by the increased number of actors involved in the development and deployment of agentic systems.

Despite these challenges, the working group generally agreed that the following normative considerations are relevant in determining when legal liability ought to be imposed:

1. Accountability: Legal liability is generally placed on actors who have taken on obligations or are at fault.

2. Compensation: Corrective justice provides that victims should receive financial remedies to restore them, as far as possible, to their original position.

3. Deterrence: Liability should discourage harmful conduct and encourage responsible deployment.

4. Innovation: As a balancing factor, the desire to promote innovation militates against an overly restrictive regulatory approach.

Existing Liability Mechanisms and Their Limitations

Generally, the working group agreed that existing legal frameworks may be adapted to sufficiently address liability issues relating to agentic AI. However, significant practical challenges remain due to the complexity of the agentic systems, the number of actors involved, as well as the unpredictability of agentic AI.

Contract 

Contract law is effective in defining actors’ rights and obligations, setting expectations of agent behaviour, and pre-allocating the risks of negative outcomes between actors.

However, its usefulness is limited by the doctrine of privity, which excludes third parties from enforcing the contractual protections available. There are also concerns that parties with greater bargaining power may disclaim liability and push risks downstream, particularly in consumer-facing scenarios.

Tort of Negligence 

The tort of negligence imposes liability for the commission of civil wrongs. The three elements of negligence remain applicable, but in the context of agentic AI, several challenges are raised:

1. Duty of care: The existence and scope of the duty of care is unclear due to the complexity of the foreseeability of the potential harm caused by AI agents, as well as the potential lack of proximity between the parties involved. For instance, it is uncertain whether a system provider owes a duty of care to third parties impacted by an agent’s action.

2. Breach: A breach of the duty of care occurs when the defendant falls below the required standard of care. While it is possible to determine the appropriate standard of care through the types of safeguards implemented, it may be difficult to ascertain what the “reasonable” level of safeguards should be. The standard of care is further complicated by the number of actors involved across the value chain, as expectations may differ across model developers, system providers, developers, and end users. 

3. Recoverable damage: A claimant must prove causation and that the loss was not too remote. In this case, tracing causation is difficult as the agent’s behaviour may be the result of an interplay between model training, system design, and user instructions. Even where causation is established, apportioning liability between actors is challenging.

Other Liability Mechanisms 

The working group briefly considered the following liability mechanisms that could apply to agentic AI:

1. Product Liability: While it was agreed that this could help apportion liability in favour of end users, Singapore’s product liability laws are presently limited to narrowly defined contexts and do not cover losses arising from AI. The group further cautioned against enacting such laws without further study.

2. Agency: It was agreed that AI agents are not legal persons and thus cannot be agents in the legal sense. Although attribution rules could potentially hold actors responsible for agents’ actions, this area was not explored further. 

Common Challenges Across Liability Mechanisms

1. Knowledge and Intention: Knowledge and intention are central to determining liability in both contract and tort. Where an agent acts in accordance with instructions, its conduct can be attributed to the instruction-giver. However, where an agent deviates from instructions or acts unexpectedly, it is difficult to attribute the relevant mental elements to any human or legal entity.

2. Foreseeability: Foreseeability links liability to what a reasonable person knew or ought to have anticipated. Agents can act in non-deterministic and emergent ways, making it difficult for actors, particularly end users and consumers, to foresee what might go wrong. However, it must also be noted that the agent’s degree of autonomy could be a design choice, which may expand what ought reasonably to be foreseen, including the risks of unintended behaviour. 

3. Causation: An agent’s behaviour may result from the interplay between model training, system design, and user instructions, making it difficult to pinpoint which actor in the value chain caused the harm. Views diverged on whether existing legal concepts such as comparative and contributory negligence are adequate to address these issues, or whether the evidentiary challenges posed by agentic AI require new approaches.

Exploring the Solution Space

The working group explored the following two potential liability regimes:

1. Fault-based liability: Under fault-based liability, a claimant must prove that an actor was at fault in how it developed, deployed or used the system. As discussed above, difficulty arises in apportioning liability across the value chain, defining standards of care, and proving causation, particularly where agents may act unpredictably despite all actors having exercised reasonable care. The paper thus sets out the following areas of consideration:

• • Duty of care: The appropriate limits of the duty of care would need to be carefully defined.
  • Breach: To determine what actions each actor needs to take to discharge their applicable standard of care, it is useful to consider as a starting point what each actor has control over. The paper sets out potential main areas of control for each actor in relation to the agentic system. The paper also considers the extent to which disclosures are relevant in fulfilling the applicable standard of care.
  • Recoverable damage: For causation, the working group considered how the law should treat “chain-of-thought” reasoning as evidence of the reason or cause for an agent’s actions. The working group also considered practical difficulties in proving causation and issues relating to the remoteness of loss.

2. Strict liability: Under strict liability, proof of fault is not required to impose liability. This allows a group of defined actors to share liability upfront, with apportionment determined through contribution proceedings, enabling effective victim compensation. However, there were differing views on whether strict liability is suitable for agentic AI.

• • Strict liability has traditionally been confined to inherently dangerous activities and imposing such broad liability on actors could deter deployment and market entry.
  • Shifting liability away from end users could introduce moral hazards, disincentivising them from using agentic systems responsibly.
  • A possible middle ground may lie in scoping down the ambit of strict liability, such as restricting such liability only to high-risk use cases.

Key Insights

The paper suggests that the central challenge is not the absence of applicable legal principles, but how those principles should be applied to the complex realities of agentic AI. While the common law has historically evolved on a case-by-case basis in response to technological developments, significant practical challenges remain, pointing to the following areas for further study:

1. How should responsibilities along the value chain be clarified? It is important to allocate responsibility based on each actor’s level of control, access to information and proximity to end users. Currently, each actor has differing degrees of control, access and proximity, which complicates allocation of responsibility. Against this backdrop, it is worth exploring how transparency and disclosures may serve as complementary mechanisms for aligning responsibilities across the value chain.

2. How can the interest of actors with limited bargaining power be better protected? Consumers may not be best placed to negotiate and allocate risks, warranting interventions such as simplified dispute resolution processes, evidential presumptions, and sector-specific liability frameworks. Further study is warranted to assess the feasibility, effectiveness, and implications of these interventions.

3. Who bears responsibility for unforeseeable agent actions? In cases where agents may still behave unpredictably despite the relevant safeguards being implemented, factors such as: (i) the existence and adequacy of disclosures; (ii) the scope of disclaimers and contractual risk allocations; and (iii) the existence and scope of insurance and other risk-allocation arrangements, may be considered.

Beyond these areas, further study may consider other alternative liability mechanisms, broader legal domains, agents built on frontier models, alternative accountability measures, as well as cross-border legal interactions.

If you have any queries on the above, please reach out to our team set out on this page.

For regional Technology, Media & Telecommunications matters, please see Rajah & Tann Asia’s Regional Technology, Media & Telecommunications Practice for more information.