Singapore: PDPC Seeks Public Consultation on Proposed PDPC Guidelines on Personal Data in Generative AI

• Where personal data is publicly available, organisations may rely on the "Publicly Available Exception" to collect, use, or disclose such data without consent.
  
  However, where online data is placed behind a digital barrier, organisations must consider whether the Exception would still apply before collecting and incorporating such data into the development of the AI system. Factors to consider include the purpose and effect of the barrier and whether the data can be accessed without restrictions from other sources. 
  
  Where an organisation intends to scrape personal data stored behind digital barriers, it is best practice for the collecting organisation to notify the other organisation of its intention to do so.
  
  
• Where user data (i.e. 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) is used to develop generative AI models, unless deemed consent or exceptions apply, organisations must obtain express consent from the individuals.
  
  Organisations should, where practicable, provide clearly visible AI-specific notifications including:
  • what functions of the generative AI model use personal data;
  • the types of personal data used to develop the generative AI model;
  • how that data is used to train or fine-tuned the generative AI model (e.g. recognising speech patterns); and
  • how individuals can decline or withdraw consent to the use of personal data for AI training (e.g. an easily accessible opt-out mechanism).

General notifications describing only the broad purpose of processing are insufficient.

Generative AI deployment involves model providers, system providers, and system deployers. Each of these players has their own distinct data protection obligations:

• Model Providers: must comply with all PDPA obligations, including:
  • as organisations, retention limitation obligation. If the training data retained by model providers includes personal data, they should develop and make available a data retention policy that includes the rationale for retaining data for a longer period.
  • as data intermediaries, implement safeguards to protect personal data. In particular, it is good practice to document and make available the measures they have taken to safeguard personal data from downstream sources (e.g. data access controls, data residency and retention policies). 
• System Providers: may act as organisations or data intermediaries depending on their role. As data intermediaries, they are expected to manage evolving data protection and security risks (e.g. prompt injection). They should also share information on system-level safeguards with downstream deployers. 
• System Deployers: bear primary responsibility for PDPA compliance. These include: 
  • assessing upstream safeguards to conduct a holistic assessment; 
  • limiting data use to appropriate purposes; 
  • implementing and maintaining safeguards for new data sources (e.g. prompts, outputs, agent activity data); 
  • educating end users on types of data that should be input; and 
  • establishing clear written governance policies and making them pre-emptively available, and regularly reviewing risks.  

In the context of generative AI, the PDPC recognised that it may be challenging for organisations to comply with their Access and Correction Obligations under the PDPA (i.e. granting individuals the right to request access to and correction of their personal data) due to the massive amounts of data used and the nature of generative AI models (e.g. training data stored as embeddings, user data temporarily held in context windows). 

Notwithstanding these difficulties, organisations are expected to adopt best practices in processing access and correction requests. Some examples of best practices include: 

• adopting upstream data handling measures (e.g. verifying data accuracy at collection and maintaining data provenance records); and 
• tracking maturity of and adopting technical measures to remove inaccurate personal data (e.g. machine unlearning).