Assess lawfulness of training data: publicly available does not mean usable
Severity CriticalOAIC generative AI development guidance, key points 2 and collection section
Before putting any dataset into pre-training or fine-tuning, determine whether it contains personal information — particularly where the provenance is unclear — and comply with the resulting obligations; public availability or accessibility is not itself a lawful basis. Additional steps such as deleting information may be needed. Collection must also respect the APP 3 necessity limit: collect only personal information reasonably necessary for the entity’s functions or activities, draft appropriate parameters for what is included when compiling a generative AI dataset, and implement mechanisms to filter out unnecessary personal information.
- Applies when
- Chinese AI companies collecting or compiling training data for models serving the Australian market.
- Evidence to retain
- · Assessment of the personal information content of each dataset
- · Documented inclusion parameters and filtering rules
- · Deletion or de-identification execution records
- · Due diligence conclusions on data of unclear provenance
Estimated first-time effort: 10–25 person-days (estimate, not a commitment)
Take particular care with sensitive information and with images and recordings of people
Severity CriticalOAIC generative AI development guidance, key point 3
OAIC states that sensitive information generally requires consent to be collected, and that many photographs or recordings of individuals, including artificially generated ones, contain sensitive information and therefore may not be scraped from the web or collected from a third-party dataset without establishing consent. Companies planning to train face, portrait or voice models must build the consent-chain check into dataset construction rather than remediate before launch.
- Applies when
- Chinese companies training or fine-tuning facial recognition, portrait generation, voice cloning or digital-human models for the Australian market.
- Evidence to retain
- · Sensitive-information screening reports for image and audio datasets
- · Consent-chain verification records, including consent evidence from third-party dataset suppliers
- · Exclusion or deletion records where consent could not be established
- · Legal review of the intended use of images of people
Estimated first-time effort: 10–25 person-days (estimate, not a commitment)
Complete a secondary-purpose assessment before using held personal information for AI training
Severity CriticalOAIC generative AI development guidance, key points 4 and 5
Where personal information already held is to be used to train an AI model and that was not a primary purpose of collection, confirm whether consent for that secondary AI-related purpose exists; if not, establish that the secondary use would be reasonably expected by the individual, taking particular account of their expectations at the time of collection, and that it is related — directly related for sensitive information — to the primary purpose, or that another exception applies. Where that cannot be clearly established, seek consent for the use and/or offer a meaningful and informed opt-out to avoid regulatory risk.
- Applies when
- Chinese AI companies planning to reuse historical Australian user data — conversation records, uploads, behavioural logs, support tickets — to train or fine-tune models.
- Evidence to retain
- · Secondary-purpose assessment addressing both reasonable expectation and relatedness
- · The notice given at original collection and evidence of user expectations, including historical privacy policy versions
- · Implementation and execution records for the consent or opt-out mechanism
- · A separate record of the direct-relatedness determination for sensitive information
Estimated first-time effort: 8–20 person-days (estimate, not a commitment)
APP 10 accuracy: take risk-proportionate reasonable steps and use disclaimers appropriately
Severity HighOAIC generative AI development guidance, Accuracy section; APP 10
OAIC notes that generative AI models are probabilistic, do not understand the data they handle or generate, and are known to produce inaccurate or false results. Under APP 10, developers must take reasonable steps to ensure the personal information collected, used and disclosed is accurate, commensurate with the likely increased level of risk in an AI context, including using high-quality datasets, undertaking appropriate testing and the appropriate use of disclaimers. Disclaimers should clearly communicate the limitations of accuracy, including whether the dataset only includes information up to a certain date, and should signal where careful consideration and additional safeguards are needed for high privacy risk uses, such as decisions with a legal or similarly significant effect on an individual’s rights.
- Applies when
- Chinese companies offering AI products in Australia that may generate content about identifiable individuals or feed decisions affecting them.
- Evidence to retain
- · Dataset quality control and provenance assessment records
- · Accuracy testing plan and results, including factual outputs about individuals
- · Disclaimer text and screenshots of where it appears, including the data cut-off statement
- · List of high privacy risk uses and the additional safeguards applied
Estimated first-time effort: 10–25 person-days (estimate, not a commitment)
Privacy due diligence and transparency updates when procuring and using commercial AI products
Severity HighOAIC commercially available AI products guidance, top five takeaways 1–5
When using commercially available AI products — third-party model APIs, SaaS and freely available chatbots included — OAIC’s use guidance requires five things. First, privacy obligations apply to any personal information input into an AI system and to output data generated by AI where it contains personal information, so due diligence at selection must consider whether the product has been tested for the intended uses, how human oversight can be embedded, the privacy and security risks, and who will have access to the personal information input or generated. Second, update privacy policies and notifications with clear and transparent information about the entity’s use of AI, ensure public-facing AI tools such as chatbots are clearly identified as AI to external users, and establish policies and procedures for AI use. Third, using AI to generate or infer personal information, including images, is a collection under APP 3 and must be reasonably necessary for the entity’s functions or activities and done by lawful and fair means — inferred, incorrect or artificially generated information produced by AI models, including hallucinations and deepfakes, is personal information where it is about an identified or reasonably identifiable individual. Fourth, where personal information is input into an AI system, APP 6 permits use only for the primary purpose of collection unless there is consent or the secondary use is within the individual’s reasonable expectations and related — directly related for sensitive information — to the primary purpose. Fifth, as best practice OAIC recommends that organisations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools. This dovetails with the APP 1.7 automated-decision disclosure duty applying from 10 December 2026: general AI transparency first, then the statutory disclosure item for automated decisions.
- Applies when
- Chinese companies whose Australian entity uses third-party AI tools on customer or employee personal information, and vendors delivering AI products to Australian customers who must support their due diligence.
- Evidence to retain
- · Privacy due diligence records for commercial AI vendors
- · Internal policy and training records on permitted uses and prohibited input data types
- · The AI-use paragraphs in the updated privacy policy
- · Approval records and logs for employee use of AI tools
- · Mapping to the APP 1.7 disclosure requirement and the update plan
Estimated first-time effort: 6–15 person-days (estimate, not a commitment)