Singapore's Personal Data Protection Commission issued guidance on the use of personal data in generative AI on 20 July 2026 (referred to here as the generative AI advisory guidelines), setting out how the Personal Data Protection Act 2012 applies when organisations build, supply and deploy generative AI. The guidelines follow the GenAI lifecycle from development through deployment to handling access and correction requests, and they assign duties to three roles: Model Providers, System Providers and System Deployers. This guide explains what each part says, what it means for security and identity teams, and how to turn it into controls.
Earlier NHI Mgmt Group pages cover Singapore's AI governance frameworks and the PDPA in general; see the Singapore PDPA Guide, the Singapore AI Governance and the Singapore Model AI Governance Framework for Agentic AI Guide. This page focuses only on the July 2026 generative AI guidelines.
Status as of 17 Nov 2026. This guide is a practitioner summary and is not legal advice. Paragraph references in brackets point to the PDPC guidelines document; confirm the current text with the PDPC before relying on it.
Key takeaways
- The guidelines were published on 20 July 2026 following a public consultation; check PDPC's site for the consultation dates and response.
- They are read with two earlier PDPC documents: the Advisory Guidelines on Use of Personal Data in AI Recommendation and Decision Systems, and the Advisory Guidelines on Key Concepts in the PDPA.
- Web-scraped public data can rely on the Publicly Available Exception, but data behind digital barriers such as paywalls, registration, API keys, geo-blocks, age-gates, rate limits or bot blockers needs a documented assessment first.
- Using customer or user data to train or fine-tune a model, without an exception or deemed consent, needs an "AI-Specific Notification". A general notice citing "new product development" is not enough for large-scale training.
- System Deployers carry primary responsibility for making sure the systems they choose meet PDPA obligations, and must protect new data sources such as prompts, outputs and agent or tool logs.
- Agentic systems need regular safeguard reviews, and the PDPC points to IMDA's Model AI Governance Framework for Agentic AI.
- Access and correction rights apply to personal data used in development or deployment, including data held by a data intermediary on your behalf.
The guidelines at a glance
| Item | Detail |
|---|---|
| Title | Advisory Guidelines on Use of Personal Data in Generative AI |
| Issuer | Personal Data Protection Commission (PDPC), part of IMDA |
| Published | 20 July 2026 |
| Consultation | 2 June 2026 to 1 July 2026, response published 20 July 2026 |
| Legal basis | Personal Data Protection Act 2012 (PDPA) |
| Structure | Development (collecting and using personal data to build models), deployment (processing personal data in deployed models and systems), post-deployment (access and correction requests) |
| Roles | Model Providers, System Providers, System Deployers; dataset curators handling personal data are also subject to the PDPA |
| Companion texts | AI Recommendation and Decision Systems guidelines; Key Concepts guidelines; IMDA Model AI Governance Framework for Agentic AI |
The guidelines are advisory. They explain how the PDPC reads existing PDPA obligations in a generative AI setting rather than creating new statutory duties, but they are the clearest statement of what the regulator expects and should be treated as the benchmark in any investigation.
Scope: who the guidelines cover
The guidelines define three stakeholder roles (6.2):
- Model Providers develop and distribute generative AI models.
- System Providers develop and distribute generative AI systems, whether bespoke or commercial off-the-shelf.
- System Deployers use or enable generative AI systems under their authority.
One organisation can hold different roles in different contexts, acting as an organisation in one and as a data intermediary in another (6.3). A software company that fine-tunes a model on its own support tickets is a Model Provider for that work; the same company hosting a chatbot for a client is likely a System Provider acting as a data intermediary for the client's data. The guidelines also note that open-weight model providers without access to downstream personal data are not subject to PDPA obligations for that data (footnote 19).
For most enterprises the important role is System Deployer. Any business that rolls out a copilot, a customer chatbot, a retrieval augmented generation assistant or an AI agent that touches personal data is a deployer and carries primary responsibility for the outcome.
Key obligations
Development: scraping public data
Organisations may consider the Publicly Available Exception instead of consent when web-scraped datasets contain personal data (3.1). Publicly available data is data generally obtainable with few or no restrictions, including data observable at open public events (3.2). Data on open sites with no restrictions may be collected, subject to a reasonable person test (3.3).
The harder case is data behind a digital barrier. The guidelines list paywalls, registration, API keys, geo-blocks, age-gates, rate limits and bot blockers (3.4), and ask organisations to assess four factors before relying on the exception (3.5):
- the purpose of the barrier, for example data monetisation;
- its effect, meaning whether the public at large or only a specific group can access the data;
- the number and complexity of steps needed to access the data;
- whether the data is accessible without restriction from other sources.
Examples that remain publicly available despite a barrier include fee-based public registers, news sites with limited free articles and large open forums that require only registration (3.6). The assessment must be documented in a data protection impact assessment or other written record that can be produced if the PDPC asks (3.7). The guidelines also remind organisations that contractual terms of use, licensing, intellectual property and criminal law sit outside PDPA compliance (3.9); in a forum example, terms that prohibit scraping mean the official API must be used (4.1).
Development: reusing user data
"User Data" is personal data provided to an organisation, or created through use of its products or services (4.1). The guidelines start with the business improvement and research exceptions, or deemed consent (4.2), and refer to the AI Recommendation and Decision Systems guidelines for their detailed conditions. Reliance can fail where the new use is a major departure from past activities; in one example the use was "so significant a departure from past commercial activities" that fresh consent was needed (4.2).
Where consent is needed, a general notice is not sufficient for large-scale training or fine-tuning (4.3, 4.4). AI-Specific Notifications should cover (4.6):
- the model functions that require personal data;
- the types of personal data used;
- how the data will train or fine-tune the model;
- how individuals can decline or withdraw consent.
Notifications can be layered, and one notice can cover several models if each context is captured (4.7). Consent cannot be made a condition of a product or service beyond what is reasonable to provide it (4.8). The guidelines flag particular concern where children's, health or credit data could be exposed or reconstructed from a model (4.5).
Development: anonymisation and minimisation
The PDPC encourages organisations to anonymise datasets as much as possible and practise data minimisation (5.1). Anonymised data is not personal data and falls outside the PDPA (footnote 14). Where personal data is necessary, physical, technical, process and legal controls apply (5.1).
Deployment: Model Providers
- Must comply with all PDPA obligations as organisations, including for end-user prompts used in development (7.1).
- Must stop retaining data once its purpose is served, under section 25 (7.2). Keeping training data for future models needs a retention policy with a rationale and regular review (7.3).
- When acting as data intermediaries during inference or hosting, must make reasonable security arrangements under section 24 (7.4).
- Should document access controls, data residency, retention, incident response and breach procedures (7.5).
Deployment: System Providers
- Are organisations when using their own datasets and data intermediaries when processing data for deployers (8.1).
- Should periodically review security, because prompt injection risks are made worse by unclear boundaries between control instructions and user data (8.2).
- Should share safeguards with deployers, including input and output filters, privacy-enhancing technologies, data-leakage testing metrics and breach notification timelines (8.3).
Deployment: System Deployers
- Bear primary responsibility for ensuring the systems they choose meet PDPA obligations, and must gather information on upstream safeguards (9.1).
- Under purpose limitation (section 18), must define the purpose and the amount of data needed, and avoid illegal or harmful processing (9.2).
- Must protect new data sources such as prompts, outputs and agent or tool logs, designate responsibility and educate users on what they may input (9.3).
- Should document written policies and make them available (9.4).
- For agentic systems with multi-step planning, action-taking and access to external tools, should review safeguards regularly and be transparent about privacy and utility trade-offs (9.5).
The worked examples are practical: restricting the data a medical document triage system can access, scanning HR chatbot prompts for identifiers and limiting who can read the logs (9.4), and applying role-based controls, data classification and human escalation to a sales AI agent (9.5).
Post-deployment: access and correction
Sections 21, 22 and 22A of the PDPA apply to personal data used for development or deployment (10.1, 10.3). Data held by a data intermediary counts as under the organisation's control, though the intermediary need not forward requests (10.2). The guidelines accept that identifying data in large datasets and removing it from models is hard, especially where training data is held as embeddings or in temporary context windows (10.4). Recommended practices (10.5) are to verify accuracy at collection, clean data and keep provenance records; to review requests case by case, correcting data in places such as a RAG database and removing inaccurate data before future training; and to track maturing removal techniques, using output filters in the meantime.
Enforcement and penalties
The guidelines do not create penalties of their own. Enforcement runs through the PDPA, so a failure the guidelines describe, such as weak protection of prompt logs or training on user data without valid consent, would be assessed as a breach of the relevant PDPA obligation: consent, notification, purpose limitation, protection, retention limitation or access and correction. The PDPA's financial penalty regime, including the turnover-based maximum introduced by the 2020 amendments, is summarised in the Singapore PDPA Guide.
Two enforcement signals in the guidelines are worth acting on now. First, assessments under the Publicly Available Exception must be documented and produced on request (3.7), so an undocumented scraping decision is itself a weakness. Second, deployers are expected to have gathered information on upstream safeguards (9.1), so "the vendor handles it" is not a defence.
What it means for security and identity teams
Much of the guidelines' deployment section is a security brief in privacy language.
- Prompts and logs are personal data stores. Paragraph 9.3 names prompts, outputs and agent or tool logs as new data sources the deployer must protect. That puts AI gateway logs, observability traces and agent memory inside the scope of your protection obligation, with access control, retention and monitoring to match.
- Prompt injection is a PDPA risk. Paragraph 8.2 links prompt injection to unclear boundaries between control and user data. A successful injection that leaks customer data from a retrieval index is a protection failure, not only a model safety issue. MITRE ATLAS tracks the technique as AML.T0051 LLM Prompt Injection.
- Agents need identity controls. The sales agent example in 9.5 calls for role-based access, data classification and human escalation. In practice that means each agent gets its own non-human identity with scoped permissions rather than a shared service account, its tool calls are logged, and high-risk actions route to a person. The Singapore Model AI Governance Framework for Agentic AI Guide covers the IMDA framework the PDPC references.
- Vendor due diligence becomes evidence. Paragraphs 7.5 and 8.3 list what providers should document: access controls, data residency, retention, incident response, breach procedures, filters and data-leakage testing metrics. Use that list as your questionnaire.
- Retrieval indexes are correctable stores. Paragraph 10.5 treats a RAG database as a place where personal data can be corrected. Build a way to find and fix records in embeddings and source stores before you need it.
Compliance steps
- Map your roles. For each generative AI use, record whether you act as Model Provider, System Provider, System Deployer or data intermediary, and for whom.
- Inventory data sources. List training sets, fine-tuning sets, retrieval indexes, prompt and output logs, and agent or tool logs, with owners and retention periods.
- Document scraping decisions. For any web-scraped data, record the digital barrier assessment against the four factors in a DPIA or written record.
- Review training consent. Where user data trains or fine-tunes a model, confirm the exception relied on or issue AI-Specific Notifications covering the four required elements, with a working opt-out.
- Minimise and anonymise. Strip identifiers from training and evaluation data where the purpose allows, and redact inputs where the use case does not need them.
- Collect upstream safeguard evidence. Request providers' documentation on access controls, residency, retention, filters, leakage testing and breach timelines, and keep it on file.
- Secure the deployment. Apply access control and retention to prompt and output logs, deploy input and output filters, test for prompt injection and data leakage, and scope agent permissions.
- Prepare for requests. Define how access and correction requests reach AI data stores, including data held by intermediaries, and how corrections feed future training runs.
- Review agentic systems on a schedule. Set a review cycle for agent safeguards and record the privacy and utility trade-offs you accepted.
Practitioner checklist
- Role register for every generative AI system (Model Provider, System Provider, System Deployer, intermediary).
- DPIA or written record for each scraped dataset, covering barrier purpose, effect, access steps and alternative sources.
- AI-Specific Notification text reviewed for function, data types, training use and withdrawal route.
- Retention policy with rationale and review date for any training data kept for future models.
- Prompt, output and agent log stores classified as personal data, with least-privilege access (NIST SP 800-53 AC-6 Least Privilege) and defined retention.
- Prompt injection and data leakage tests run before release and after significant changes.
- Each AI agent has its own scoped identity, logged tool use and human escalation for high-risk actions.
- Provider safeguard documentation on file and reviewed at contract renewal.
- Access and correction procedure covers retrieval indexes, intermediaries and future training runs.
- Written AI data policy published to staff, including what may and may not be entered into prompts.
Standards and references
- PDPC: Advisory Guidelines on Use of Personal Data in Generative AI (PDF, 20 July 2026)
- PDPC: Guidance for organisations on responsible use of personal data in generative AI (press release)
- Allen & Gledhill: PDPC issues Advisory Guidelines on Use of Personal Data in Generative AI
- MITRE ATLAS: LLM Prompt Injection
- NIST: AI 600-1 Generative Artificial Intelligence Profile
- ISO: ISO/IEC 42001:2023 Artificial intelligence management system
- OWASP: LLM02:2025 Sensitive Information Disclosure
Related NHI Mgmt Group resources: Singapore PDPA Guide · Singapore AI Governance · Singapore Model AI Governance Framework for Agentic AI Guide