Assess and document reliance on the Publicly Available Exception for training data
Severity Criticalparas 3.1–3.9
For every scraped corpus used in pre-training or fine-tuning, assess whether the personal data in it is publicly available. A digital barrier does not by itself defeat public availability, but the assessment must weigh the barrier’s purpose and effect, the number and complexity of steps needed to reach the data, and whether the data is obtainable without restriction elsewhere. Where it is reasonably arguable that the data behind a barrier is not publicly available and the organisation nonetheless relies on the exception, it must record the assessment and reasoning in a DPIA or other written record and produce it if PDPC asks.
- Applies when
- Model Providers and dataset suppliers that build training corpora by web scraping.
- Evidence to retain
- · Inventory of scraped sources annotated with barrier type
- · Per-source public-availability assessment records
- · Data protection impact assessment (DPIA)
- · Record of robots.txt and terms-of-service compliance checks for the crawl
Estimated first-time effort: 10–25 person-days (estimate, not a commitment)
Provide an AI-Specific Notification before training on User Data
Severity Criticalparas 4.1–4.8 (duty at 4.4; encouraged contents at 4.6)
Where User Data is used for large-scale training or fine-tuning of generative AI models, a general “new product development” statement is not sufficient consent: para 4.4 states that organisations must provide AI-Specific Notifications for this purpose. The form and wording are for the organisation, but individuals must be able to understand the types of personal data affected, how they will be used to train or fine-tune the model, and the functions of those models. Para 4.6 goes on to say organisations are encouraged to provide, to the extent practicable, four items — the model functions requiring personal data, a clear description of the types of personal data used, how the data will be used to train or fine-tune the model, and how individuals can decline or withdraw consent. Those four are best practice rather than a mandatory checklist, but should be treated as PDPC’s review baseline. Layered notification is acceptable, as is a single notification covering several models; consent to processing beyond what is reasonable to provide the product must not be a condition of supply.
- Applies when
- Chinese AI companies that offer apps, platforms or consumer services directly to individuals in Singapore and wish to reuse their conversations, uploads or behavioural data for model training.
- Evidence to retain
- · AI-Specific Notification text with version history
- · Privacy policy update announcement and in-product notice screenshots
- · Opt-out and withdrawal mechanism plus execution records
- · Reconciliation records between training datasets and users who withdrew consent
Estimated first-time effort: 8–20 person-days (estimate, not a commitment)
Map the organisation’s role in the generative AI stack and implement the matching obligations
Severity Highparas 6.1–6.3, 7.1–7.5, 8.1–8.3, 9.1
Determine, per business line, whether the company is a Model Provider, System Provider or System Deployer, and whether in that capacity it is an organisation or a data intermediary, then configure the obligation set accordingly. A Model Provider that processes personal data for its own purposes — including harvesting downstream prompts and inputs for model development — is an organisation subject to all PDPA obligations; when running inference or hosting data for downstream customers it is a data intermediary owing at least the protection and retention-limitation obligations. The same entity may switch roles between scenarios, and each must be documented in policy.
- Applies when
- Chinese AI companies that simultaneously run foundation models, model APIs, vertical SaaS and internal deployments. Open-weight Model Providers with no access to downstream personal data are not subject to PDPA obligations in respect of that data.
- Evidence to retain
- · Business-line to role to obligation mapping table
- · Contract clauses distinguishing organisation and data intermediary capacity
- · Internal policies and procedures per role
- · Statement of downstream data-access boundaries for open-weight releases
Estimated first-time effort: 5–12 person-days (estimate, not a commitment)
Assign ownership and safeguards for prompts, outputs and agent activity data
Severity Highparas 7.5, 8.2–8.3, 9.1–9.5
System Deployers must identify and register the new data sources their systems create — end-user prompts and inputs, generated outputs, agent or tool activity data, and connected internal enterprise data — assign responsibility for each, implement matching safeguards, and educate internal and external users on what categories of personal data may be entered and how it will be processed. Where systems have agentic functionality, deployers must regularly review the sufficiency of safeguards and be transparent about the privacy-utility trade-off. Model and System Providers should disclose downstream the safeguards implemented, testing and performance metrics such as the likelihood of data leakage, and incident response and breach-notification procedures.
- Applies when
- Deployers launching AI assistants, support bots, document workflows or agentic systems in Singapore, and the Model and System Providers supplying them.
- Evidence to retain
- · Register of new data sources with assigned ownership
- · Guidance on data types permitted in prompts plus staff training records
- · Access controls and redaction configuration for prompt logs
- · Downstream-facing safeguards information sheet including data-leakage test metrics
- · Privacy-utility trade-off assessment for agentic use cases
Estimated first-time effort: 10–25 person-days (estimate, not a commitment)
Handle access and correction requests concerning personal data in models and systems
Severity Mediumparas 10.1–10.5
Access and correction obligations apply to personal data collected, used or disclosed for model or system development and deployment. Organisations must accede unless a PDPA exception applies — for example where the burden or expense of access would be unreasonable to the organisation or disproportionate to the individual’s interests, or where there are reasonable grounds that a correction should not be made. Data transferred to a data intermediary remains under the organisation’s control and must be covered in the response. The guidelines acknowledge the technical difficulty created by embeddings and context windows, but this is not an automatic exemption.
- Applies when
- Any entity acting as an organisation in Singapore that uses personal data to build or operate generative AI models and systems.
- Evidence to retain
- · Searchable index or data-lineage records for training corpora and User Data
- · Register of access and correction requests
- · Written justification and technical-limitation explanation where an exception is relied on
- · Records of retrieval and correction instructions issued to data intermediaries
Estimated first-time effort: 8–20 person-days (estimate, not a commitment)