Executive summary

The rapid advancement and adoption of generative artificial intelligence (Generative AI) and agentic AI systems have introduced novel challenges for data protection and legal liability frameworks. To help organisations navigate these challenges, Singapore has in recent months issued the following:

  1. in May 2026, the Infocomm Media Development Authority of Singapore (IMDA) released a Discussion Paper on Legal Responsibility for AI Agents (the Discussion Paper on AI Agents), which considers how liability should be allocated where AI agents act with some degree of independent planning, decision-making and action-taking. The Discussion Paper recognises that existing doctrines, including contract law, tort of negligence, Rylands v Fletcher (strict liability), product liability and law of agency, may be capable of addressing some liability issues, but also identifies practical difficulties around attribution of knowledge and intention, foreseeability, causation, contractual risk allocation and the applicable standard of care. It therefore identifies further areas for study, including how responsibilities should be clarified across the AI value chain, how consumers and other actors with limited bargaining power should be protected and who should bear responsibility for unforeseeable agent actions;
  2. on 20 July 2026, the IMDA issued the voluntary Transparency Guidelines for Generative AI Chatbots (the Transparency Guidelines), which introduces the concept of a “chatbot info card” as a single reference point for users, organised around relevance, accessibility and timeliness. Deployers are encouraged to disclose what the chatbot can and cannot do, the extent to which it is reliable and safe, how user data is used and protected, and how users can report issues. The IMDA also emphasises that chatbot info cards should be easy to understand, navigate and find, and should be published before users engage with the chatbot and updated within a reasonable time when material changes affect users;
  3. on 20 July 2026, the Personal Data Protection Commission (PDPC) issued its Advisory Guidelines on use of Personal Data in Generative AI (the Generative AI Guidelines), which address the use of publicly available personal data, consent and notification requirements for user data and the responsibilities of model providers, system providers and system deployers. In particular, they make clear that general notices are insufficient where personal data is used for AI or Generative AI model development, and that organisations should provide AI-specific notifications that explain the relevant data uses in understandable terms.

The key features of each publication are discussed below.

Discussion Paper on AI Agents

The Discussion Paper on AI Agents (accessible here) examines how legal responsibility should be allocated when AI agents act autonomously, use tools, interact with third parties and cause harm. It focuses on civil liability and private law, especially Singapore law, and draws together perspectives from the government, academia and industry.

While there is no broad consensus on the definition of an AI agent, the Discussion Paper identifies certain common features - agents usually possess some degree of independent planning, decision-making and action-taking. It further acknowledges that the challenge in allocating liability for agentic AI lies with the proliferation of actors1 within the value chain.

Applying existing law and its challenges

The legal doctrines considered include contract law, the tort of negligence, the rule in Rylands v Fletcher (strict liability), product liability and the law of agency. It is generally agreed within the working group of the Discussion Paper that existing legal principles may be able to address liability issues arising from agentic AI, but with significant practical challenges:

  1. Knowledge and intention. In many cases, where an AI agent acts in accordance with instructions, existing legal doctrines can "look through" the agent to the actor behind it. However, such attribution is difficult where the agent deviates from those instructions, or interprets or executes them in an unexpected way.
  2. Foreseeability. As agents become more general-purpose and autonomous, their actions may cause unforeseeable types of harm, in which case liability may not attach to any party.
  3. Causation. Many members expressed concern that proving which actor in the agentic AI value chain caused the incident may be challenging, giving rise to a near "impossibility" for a claimant to pinpoint what went wrong and who caused it.
  4. Contract – allocation of risk and disclaimer. While contracts are effective means of defining actors' rights and obligations, the allocation of risk will often be dependent on bargaining positions and parties with stronger bargaining power may carve out responsibilities and disclaim liability for many incidents, leaving downstream actors with no choice but to accept the risk.
  5. Tort of negligence – standard of care. While the tort of negligence may serve as a liability framework for agentic AI, determining what constitutes a reasonable level of safeguards and how responsibility should be allocated among actors across the value chain, remains unclear.

Areas for further study

On this basis, the Discussion Paper identified three key areas for further study: (i) how responsibilities along the value chain should be clarified in the context of agentic AI; (ii) how actors with limited bargaining power, such as consumers, can be better equipped through measures such as simplified dispute resolution, presumptions, or sector-specific liability frameworks; and (iii) who bears responsibility for unforeseeable agent actions.

The Transparency Guidelines

In addition to examining how legal liability should be allocated for increasingly autonomous AI systems, IMDA has addressed the external-facing aspect of Generative AI deployment through its Transparency Guidelines.

The Transparency Guidelines (accessible here) set out voluntary guidance for organisations that deploy external-facing Generative AI chatbots (used by customers or the public etc.).2 Rather than focusing on the underlying Generative AI Models, the guidelines target the application layer, on the basis that chatbots are Generative AI applications with the widest consumer reach and have drawn considerable public concern.

The Chatbot Info Card and Three Key Principles

At the centre of the Transparency Guidelines is the concept of a “chatbot info card”, akin to a medical label, which serves as a single, consolidated reference point through which consumers can access essential information about the chatbot they are using. The Transparency Guidelines organise the expected content and presentation of a chatbot info card around three key principles—(i) relevance, (ii) accessibility and (iii) timeliness.

Relevance

With regards to relevance, deployers are encouraged to address four key questions:

  1. What the chatbot can be used for, including its capabilities, limitations and prohibitions;
  2. How reliable and safe is the chatbot, including the risks it poses, the safeguards in place and the precautions users should take;
  3. How user data will be used and protected, including what data is collected, who has access to it, whether it is used for model training and what user control options are available; and
  4. How users can report issues, including the relevant channels and the deployer’s response process.

The Transparency Guidelines encourage deployers, at a minimum, to ensure that the chatbot info card addresses all four questions, with at least one substantive disclosure (i.e., a concrete, specific statement that users can act on) in each. Nonetheless, they recognise that the extent of disclosure may be calibrated against legitimate constraints such as proportionality, protection of proprietary information, security and reasonable availability of information.

Accessibility

Beyond identifying what information to disclose, the Transparency Guidelines emphasise how such information should be presented. Deployers are encouraged to make the chatbot info card:

  1. Easy to understand, by using plain language rather than jargon-heavy phrasing, while remaining specific and substantive;
  2. Easy to navigate, through layered disclosure, structured content and visual aids; and
  3. Easy to find, by maintaining it as a single one-stop-shop reference. At first use, deployers should surface a high-level safety statement together with a clearly identifiable link to the chatbot info card. The chatbot info card should also be accessible from within the chatbot interface itself. For chatbots available on multiple platforms (e.g., website, mobile application), the chatbot info card should be equally accessible across the platforms.

Timeliness

The Transparency Guidelines further recommend that chatbot info cards be published by the time a chatbot is made available to external users, including in beta or pilot form, so that users can review the card before they engage with the chatbot. For chatbots that were already publicly accessible before the guidelines were issued, deployers are encouraged to publish the corresponding chatbot info cards as soon as reasonably practicable. Deployers should also continually assess whether the disclosed information remains accurate and update the chatbot info card within a reasonable time where a development, such as a change to the underlying model, guardrails, capabilities or use cases, meaningfully affects users. Deployers are further encouraged to clearly version each chatbot info card (e.g., by including a “last updated” date and, where applicable, the chatbot version to which it relates).

Allocation of responsibility among deployers and upstream providers

Critically, the Transparency Guidelines place primary responsibility for providing transparency to end-users on chatbot deployers, being the organisations that make a chatbot available for use. They are responsible for conducting risk assessments and applying adequate mitigations before deployment, and are the first port of call when something goes wrong. Upstream parties, such as model providers, are not thereby absolved of responsibility - they are expected to support deployers by furnishing relevant and up-to-date information about their underlying models to enable deployers to conduct risk assessment and mitigation, and to be transparent to users.

The Generative AI Guidelines

In parallel with the steps taken by IMDA, the PDPC has addressed a related issue by clarifying organisations' existing data protection obligations in the Generative AI context.

The Generative AI Guidelines were released on 20 July 2026 (accessible here), following a month-long public consultation from 2 June to 1 July 2026. They clarify how personal data can be used in Generative AI Models3 and Systems4. Prior to the release of the Generative AI Guidelines, the PDPC did not have any specific Advisory Guidelines dealing with the use of personal data in the context of Generative AI.5 In particular, the Generative AI Guidelines address some of the key data protection issues relating to Generative AI today—(i) the collection and use of personal data to develop Generative AI Models; (ii) the allocation of data protection responsibilities across the Generative AI lifecycle; and (iii) the handling of individuals' requests concerning the processing of their personal data for Generative AI.

Development: Collecting and using personal data to develop Generative AI Models

Publicly Available Exception. Web-scraping is a common means of compiling datasets for Generative AI development. Where web-scraped datasets include personal data, organisations can rely on the "Publicly Available Exception" under the Personal Data Protection Act 2012 (PDPA).6 This exception enables organisations to collect, use, or disclose personal data about an individual that is publicly available7 without consent, subject to the reasonable test.8

However, where personal data sits behind a digital barrier9, being any technical and/or financial measure that meaningfully restricts data access in whole or in part, the Generative AI Guidelines provide that organisations must make an assessment on whether the data is publicly available before they can rely on the Publicly Available Exception. Whether data is publicly available is ultimately a fact-specific inquiry, which turns on the purpose and effect of the digital barrier, the steps required to access personal data and whether the personal data remains accessible without restriction from other sources. Critically, these considerations are equally applicable for organisations and/or individuals who make personal data available online - where web-scraping is not intended or consent has not been obtained, appropriate digital barriers must be implemented.

Consent and Notification Obligations. Another key source of data used to develop Generative AI Models is User Data10. As a starting point, organisations may consider relying on exceptions to the Consent Obligation11 or deemed consent to use User Data to develop Generative AI Models. However, in cases where organisations determine that such alternatives are not applicable, they must obtain consent. The Consent Obligation is complemented by the Notification Obligation, which requires individuals to be notified of the purpose of the intended use of their personal data when their consent is sought for such use.

Critically, the PDPC takes the view that general notifications, such as citing the use of personal data for "new product development" without specifying AI or Generative AI Model development, are insufficient to obtain valid consent. Organisations must instead provide AI-Specific Notifications that explicitly state that the purpose of processing includes AI and/or Generative AI Model development. While such notifications need not be overly technical, individuals must be able to understand the types of personal data affected and how that data will be used to train and/or fine-tune AI models, as well as the function(s) of these models.

Deployment: Data protection responsibilities of Generative AI stakeholders

The Generative AI Guidelines also identify the relevant stakeholders responsible for the safeguarding of personal data: (i) "Model Providers" who develop and make available Generative AI Models for distribution and use; (ii) "System Providers" who develop and make available Generative AI Systems for distribution and use; (iii) "System Deployers" who use and/or enable the use of Generative AI Systems under their authority; and (iv) organisations that specialise in collecting, curating and supplying datasets for Generative AI development and/or deployment.12

  1. Model Providers are considered “organisations” under the PDPA where they process personal data to develop and deploy Generative AI Models. As “organisations”, they are reminded to pay particular attention to their Retention Limitation Obligation13—where such data includes identifiable personal data, Model Providers should, as good practice, develop and make available a data retention policy setting out the rationale for any extended retention period. Model Providers may also act as “data intermediaries” where they process personal data on behalf of downstream users and must comply with the applicable PDPA provisions (i.e., the Protection14 and Retention Limitation Obligations). In demonstrating compliance with the Protection Obligation, Model Providers should document and make available the measures taken to safeguard personal data from downstream sources—however, where Model Providers do not have access to personal data in downstream systems built on their models, they are not subject to the PDPA obligations in respect of that data.
  2. System Providers refer to entities who are engaged to develop bespoke or customisable Generative AI Systems, as well as those that develop and retail commercial off-the-shelf systems. System Providers are regarded as “organisations” when they process personal data as part of their own datasets to develop systems; when they process data on behalf of downstream deployers, they are “data intermediaries”. As “data intermediaries”, System Providers are expected to periodically review the need for additional security arrangements as they develop and make available new types of systems. As good practice, System Providers should also share information on their system-level safeguards with downstream deployers.15
  3. System Deployers bear primary responsibility for ensuring that their chosen Generative AI Systems can meet their obligations under the PDPA. They must ensure that personal data processed by their systems is for relevant purposes, in accordance with the Purpose Limitation Obligation under Section 18 of the PDPA16. Pursuant to the Protection Obligation, System Deployers must also safeguard personal data in their possession or under their control, including new categories of data sources such as end-user prompts, inputs, generated outputs, agent or tool activity data and internal enterprise data.

Post-deployment: Access and correction obligations

The PDPC acknowledges present-day challenges to facilitating Generative AI-related access and correction requests, including the massive amounts of data involved, the nature of Generative AI Models and/or Systems where training data is stored as embeddings or User Data is temporarily held in context windows and other technical limitations. Notwithstanding these challenges, organisations are expected to comply with the Access and Correction Obligations, and adopt best practices including: (i) upstream data handling measures such as verifying data accuracy at the point of collection and maintaining data provenance records; (ii) reviewing access and correction requests on a case-by-case basis and acceding where reasonable; and (iii) tracking the maturity of and progressively adopting appropriate technical measures to remove inaccurate personal data.

Key takeaways and looking ahead

The publication of the Generative AI Guidelines, the Discussion Paper on AI Agents and the Transparency Guidelines reflects Singapore's holistic and proactive approach to governing advanced AI technologies. The Generative AI Guidelines provide concrete, actionable guidance for organisations across the AI lifecycle, from the development phase through deployment and post-deployment. The Discussion Paper on AI Agents, meanwhile, tackles the more forward-looking and unresolved question of legal liability for increasingly autonomous AI systems. The Transparency Guidelines complement both, addressing the more immediate question of how organisations deploying external-facing Generative AI chatbots can build user trust and accountability through meaningful, consistent disclosure. Against the standards set out in these publications, organisations should carefully consider their AI governance approach and review their current practices to assess where changes may be needed.

Footnotes

  1. 1

    This includes model developers, tooling providers, platform providers, system providers, deployers, end users and third parties.

  2. 2

    Chatbots deployed solely for internal organisational use, such as employee productivity tools or internal knowledge assistants, fall outside the scope of the Transparency Guidelines.

  3. 3

    “Generative AI Models” refer to models, including those trained on a large amount of data using self-supervision at scale, that display significant generality and can competently perform a wide range of distinct tasks regardless of the way the models are placed on the market and that can be integrated into a variety of systems or applications.

  4. 4

    “Generative AI Systems” refer to AI systems or applications based on Generative AI Models, both for direct use as well as for integration in other AI systems.

  5. 5

    While the PDPC did not release any advisory guidelines specific to Generative AI, it had previously issued a set of Advisory Guidelines on use of Personal Data in AI Recommendation and Decision Systems (issued 1 March 2024).

  6. 6

    See paragraph 1 of Part 2 of the First Schedule to the PDPA.

  7. 7

    Publicly available personal data refers to personal data that is generally available to the public, including personal data which can be observed by reasonably expected means at a location or an event at which the individual appears and that is open to the public.

  8. 8

    The reasonable test here refers to whether a reasonable person considers the use of personal data to develop Generative AI Model appropriate in the circumstances.

  9. 9

    Examples of digital barriers include paywalls, registration requirements, authentication mechanisms, location or eligibility-based access controls and tools that detect and prevent automated programs from accessing online data.

  10. 10

    “User Data” refers to personal data provided by an individual to an organisation or created in the course of or as a result of the individual's use of the organisation's products or services.

  11. 11

    See paragraphs 4.1 to 6.4 of the Advisory Guidelines on Use of Personal Data in AI Recommendation and Decision Systems.

  12. 12

    Where such datasets include personal data, these organisations are subject to the PDPA.

  13. 13

    Section 25 of the PDPA requires an organisation to cease to retain its documents containing personal data, or remove the means by which the personal data can be associated with particular individuals as soon as it is reasonable to assume that the purpose for which that personal data was collected is no longer being served by retention of the personal data and retention is no longer necessary for legal or business purposes.

  14. 14

    Section 24 of the PDPA requires an organisation to protect personal data in its possession or under its control by making reasonable security arrangements to prevent unauthorised access, collection, use, disclosure, copying, modification or disposal, or similar risks and the loss of any storage medium or device on which personal data is stored.

  15. 15

    This includes data security and protection measures, testing and performance metrics, as well as incident response and data breach procedures.

  16. 16

    Section 18 of the PDPA limits the collection, use or disclosure of personal data about an individual only for purposes that a reasonable person would consider appropriate in the circumstances and that the individual has been informed of under section 20 of the PDPA.

Contact