IMDA issues discussion paper on legal responsibility for AI agents: Allen & Gledhill

30 July 2026

On 20 May 2026, the Info-communications Media Development Authority of Singapore (“IMDA”) issued a discussion paper, titled “Legal Responsibility for AI Agents” (“Paper”), which examines how legal responsibility should be allocated when artificial intelligence (“AI”) agents act autonomously, use tools, interact with third parties, and cause harm. The Paper focuses on civil liability and private law while recognising that agentic AI may also raise other legal issues. It does not cover issues that may arise from criminal law, data protection law, and other regulatory mechanisms nor does it provide specific policy recommendations.

The Paper explains that AI agents are becoming increasingly autonomous - taking actions, learning from experience, and interacting with third parties in potentially unforeseeable ways. While evaluation and governance frameworks are being developed to address these new risks, there are calls for a more systematic investigation into how legal liability applies to AI agents.

Against this backdrop, IMDA convened a working group of members of Singapore’s legal community (“Working Group”) to develop a shared understanding on how current civil liability and private law frameworks relate to responsibility for agentic AI, including any challenges that may need to be addressed.

Allen & Gledhill Partner Melissa Mak is a member of the Working Group.

Main points

The Paper summarises the following main points from the Working Group discussions:

Autonomy, decision-making, and action-taking are key features of agentic AI relevant to liability: Autonomous agents that can make decisions with reduced human involvement can diffuse accountability for outcomes and increase the risk of misaligned or unexpected behaviour. The larger action space can also increase the impact of negative outcomes. Further, liability must be allocated between the increased number of actors involved in the development and deployment of an agentic system.
Application of existing legal frameworks: A majority of the group considered that many cases may be capable of being addressed through the common law, such as contract and the tort of negligence. However, the law may need to be adapted for agentic AI. There will likely be significant practical challenges faced by claimants due to the technical complexity of agentic systems and the number of actors involved, particularly for consumers and those with limited bargaining power. The unpredictability of agentic AI was also seen as a significant challenge, especially in scenarios where all parties took relevant safeguards, but the agent still caused harm in an unexpected way. This raised the question of who would bear responsibility for such unforeseeable losses.
Attribution of fault: In exploring the solution space, fault-based liability and strict liability were considered in the context of a hypothetical of a computer-use agent that deviated from instructions and caused losses to third parties. It was noted that under the tort of negligence, there were potential issues in identifying where the fault lay, whether actions had fallen below the required standard of care, and remoteness of the loss due to the unforeseeability of the agent’s actions. While strict liability could shift the cost of complex apportionment disputes away from end-users and third parties, it may lead to unscoped liability and moral hazards.

Areas for further study

The following three key areas were identified for further study:

How should responsibilities along the value chain be clarified in the context of agentic AI? Model developers are usually best placed to shape the agent’s underlying capabilities or safety properties but have limited insight into how the agent will eventually be used. Down the chain, there may be more specificity in the use case but less opportunity to intervene in the agent’s base behaviour. This may mean differentiated responsibilities along the chain, while considering the relevance of disclosures or transparency as a complementary mechanism to fulfil such responsibilities.
How can actors with limited bargaining power be better equipped?
If left to the market, parties with less bargaining power, such as consumers, may end up accepting most of the risk. This is not unique to agentic AI. The resulting risk borne is greater due to the increased capabilities and autonomy of agents, and the difficulty of proving fault. Further study is required to assess what measures may be appropriate, such as simplified dispute resolution processes, introducing legal or evidential presumptions, or sector-specific liability frameworks.
Who bears responsibility for unforeseeable agent actions? In some cases, even where all actors along the value chain have taken the relevant safeguards, agents may still behave unpredictably or act in ways that were not anticipated, resulting in harm. In assessing whether, and to what extent, loss should be attributable to one or more actors, it may be relevant to consider various factors relating to transparency, the extent to which the allocation of risk reflects the distribution of benefits across the value chain, and the reasonableness of reliance placed on the agent.

Reference materials

The discussion paper is available on the IMDA website www.imda.gov.sg.