Introduction
As we are keenly aware, many regulators throughout the world are facing a sharp increase in complaints and requests coming to the door. The OAIC is no different. In the last financial year, we had a 60% increase in privacy complaints and 35% more freedom of information complaints. While we continue to increase the numbers we finalise, it’s impossible to keep pace in the current circumstances.
For the OAIC, complainants’ use of artificial intelligence to formulate their complaints has so far not been a significant issue. But elsewhere we continually hear of how AI is being used by the public to aid their requests, leading to complaints that may be spurious or rely on incorrect information, that “clog” the system. A phenomenon that has been called “agentic flooding”.
This can be overwhelming for regulators. And it’s part of the reason why regulators are increasingly looking at AI tools and other measures that can reduce backlogs and aid productivity.
The OAIC is also exploring the AI resources it can harness. As a regulator of privacy and information access, we are especially aware of the risks AI can pose to the provision of those rights if not designed and implemented correctly.
For example, if we offer a chatbot mechanism that empowers the individual when making a complaint, then those gains are tarnished if fundamental rights of privacy and information access are sacrificed at the same time.
It might be instructive to target a use case – in this case a chatbot in the context of complaint handling to really explore these issues, starting with a chat bot as the first-tier public interface for complaints handling.
There are real benefits in undertaking the process of deploying a chatbot e.g. the benefits of amalgamating all public facing records. But there are also risks. In a government context we have obligations e.g. an Ex Gratia or CDDA scheme; administrative law considerations, record keeping obligations. One way to mitigate these risks is to consider a design that facilitates data collection in response.
In developing the best response, we need to take a holistic view. It is not just about getting the queue of complaints down.
Ultimately our goal is for people to engage with regulators, raise their concerns and be dealt with efficiently, and the regulator to achieve its mandate for the community. A better-informed community and better rights preservation practices by respondent entities is the ultimate ambition.
Implementing artificial intelligence in the provision of government services, and mitigating risks, raises a number of issues. The preservation of rights in the deployment of AI is at the core of our current work program.
We are currently finalising a Statement of Principles for the APS to assist in achieving the productive and beneficial goals of AI whilst ensuring the preservation of the fundamental human rights of information access and privacy.
This comes against our 2026 report on FOI obligations and ADM which will complement the guidance we are developing to support the introduction of privacy disclosure requirements for entities using ADM which comes into effect on 10 December 2026. We have also published significant guidance on AI including training models and its direct application in the health care setting in the Harrison AI case.
This work is framing a structured way of approaching AI to ensure the preservation of rights. There are 4 elements that we are currently applying to inform consideration of right preserving approaches:
- Designing
- Procuring
- Deploying
- Harnessing
Design
I will start with talking about the design of AI solutions.
Chatbots, as my colleagues will attest, can be useful methods of triaging inquiries, can offer useful self-service items and multilingual capabilities. They can empower complainants, tailor complaints, and help the organisation respond.
And they are a prompt to designing internal systems, which may rely on AI, that are used by agencies to develop their responses. Internal systems can draw upon publicly available records, jurisdictional information, and the data provided by the customer, to ensure responses are soundly based. If not, we need to review our practices – I’ll deal with that further under deployment.
The questions arise over what information is required from the customer and what information is relied upon in the response. Some high-level considerations:
- Provenance – at least a 2-way street
- Accuracy/Currency – an ongoing obligation
- Content – the nature of information provided and received, sensitive/personal/policies
- Control – records/access/updates/risks/reviews/revision
As a first step, deploying agencies should develop public facing policies and internal procedures for the use of AI systems to facilitate transparency and ensure good privacy governance through evaluation including a privacy impact assessment.
A chatbot is providing information – but it is also receiving information – is it receiving information from an individual or another chat bot? What is your jurisdiction? Are the rights being exercised only available to a legal person – is that person being assisted – how do you know/record that? What preconditions are necessary in your jurisdiction and how has the system you implement to exercise a function, e.g. advice reflected any jurisdictional considerations.
The accuracy and currency of the information you share is also relevant to your jurisdiction and your obligations. This is an ongoing obligation, and it has relevance to the information provided by both the organisation and the individual.
Regarding content - what information should you collect? If your organisation collects personal information through public facing AI systems, such as through the use of AI chatbots, this collection must comply with the requirements of Australian Privacy Principles 3 (the collection of personal information to be reasonably necessary for your entity’s functions or activities, and to be carried out by lawful and fair means); APP 5 (what notification requirements have you included) and then there is APP 12 - access to personal information.
If your organisation is using a proprietary AI system rather than a publicly available chatbot, for example, and has protections in place to ensure that information entered into the system will not be disclosed outside the organisation, such as to the system developer, this will constitute a use rather than a disclosure of personal information.
The key question is: how does your organisation plan to use the AI product? What is the potential privacy impact of these uses and how is that information access by the individual, the organisation and any third party?
Procurement
These are all crucial issues when it comes to procuring AI solutions.
The Australian Information Commissioner Act, which I oversee, impacts on the systems, policies and practices of government agencies.
Privacy regulation clearly extends to the private sector, as you would know, but freedom of information (FOI) legislation also has an impact. It reaches through into private sector contracts. This provides visibility of technology, and risks and harms in a public and private setting. As I will mention more than once, transparency is critical to the use of AI by government agencies.
While the issues are more obvious with third-party systems than proprietary technology, all agencies should examine the impacts on privacy and information access when it comes to implementing systems, and these should be built into procurement contracts.
There are a number of questions agencies should ask when using third-party systems.
- What legal obligations does a respondent entity to preserve rights – FOI Act and Privacy Act jurisdictional parameters?
- What contractual provisions are in place to preserve information access and privacy rights?
- What data has the product been trained and tested on? Is this training data diverse and relevant to your organisation, and what is the risk of bias in its outputs?
- What are the limitations of the system, and has the developer put in place safeguards to address these?
- What safeguards including notification of adverse harm of the use case more broadly are secured in your contract?
- What is the intended operating environment for the product? Are there any known or likely security risks?
- What are the data flows in the AI product? Will the developer or other third parties have access to the data which your organisation inputs or generates when using the AI product? Can you turn these features off or take other steps to comply with privacy obligations?
AI can play an important role in ensuring greater efficiencies in Australia’s FOI system, which we also regulate, but it’s important that good record-keeping is part of the system.
AI-generated and related records may be subject to FOI requests and record-keeping obligations. Agencies should have traceability mechanisms that document the inputs, outputs, training datasets of AI systems alongside administrative records for their development, implementation and management.
The other major issue to be consider is security of data and risk of data breach. This can occur from human error if the appropriate safeguards are not in check. But as we have recently seen, through Open AI’s breach of Services Australia, that the threat can come from AI itself. What safeguards are in place to prevent agentic AI from initiating breaches or releasing personal information?
Deploy
It’s clear from what I have said so far is that the use of AI complaint handling tools is not a set and forget situation. When these tools are deployed there is a constant need to monitor all systems to ensure they are delivering to stakeholders and risks are being mitigated.
An added responsibility is that customer-facing chatbots are often used by vulnerable members of the public. This may lead to both positive and negative consequences for customers, however, it also raises further sensitivities about the security of personal information.
The chatbot interface may also appear to be working seamlessly but further interrogation may reveal previously hidden problems.
Record-keeping of AI usage is also important. Staff should be empowered to report concerns or issues with the use of AI, including concerns about the accuracy or reliability of AI-generated outputs (including AI hallucinations or incidents) or any unintended use of AI that is not consistent with internal guidelines.
Taking this a step further, agencies should also have complaint handling and dispute resolution mechanisms to resolve privacy enquiries and complaints around their use of AI systems.
There needs to be clear guidelines about what are no-go areas. This can go to issues of data security, privacy and information access. But it also goes to more fundamental issues such as how the agency fulfils its function as a complaint-handling organisation, and how it embodies APS values.
For example, here is a scenario we have discussed at the OAIC – it is subject and jurisdiction specific so it may not be so clear cut in your environment.
One of our staff is drafting a preliminary decision letter on a complaint received by the OAIC. The complainant has made 8 separate submissions that need to be referend in the decision. Can they use generative AI to create a summary of these submissions and then use this summary as my basis of their decision?
Our response is no, they cannot. We are exercising a significant external review power in FOI. While an AI summary can be used to help understanding, the staff member needs to read all relevant documents to ensure they fully understand the matter. Furthermore, the direct inclusion of summary content produced by generative AI in a decision letter is not an approved use of AI, as the content may directly inform a statutory decision and may constitute automated decision-making (ADM).
Harness
There are significant benefits to be harnessed. These will be unavailable to us if we don’t have a system that involves governance. Data capture and review will enable us to identify errors, mis and disinformation, unintended consequences and importantly trends. As complaints handlers our predictive power is largely driven by lag indicators – e.g. seeing a trend line that identifies new and emerging harms.
If we don’t review the chatbots operation at an aggregated level we will miss a most important source of intelligence. We also need to look at activities that can lessen the need for complaints in the first place, such as proactive publication of government information on agency websites.
Additionally, given the challenges presented by AI, and the speed at which the technology has developed, processes and systems will continually need to be tested and revised as part of a continuous evolution.
Accountability and reporting lines will need to be strengthened, and any data that will inform processes needs to be updated and remain verifiable.
There will be situations where individuals will act under Scheme for Compensation for Detriment caused by Defective Administration, and that’s where strict accountability processes will be vital.
Our work
Now I’d like to turn to some of our other activities that are relevant to this work.
The OAIC is very close to finalising a Statement of Principles for the Australian Public Service to assist in achieving the productive and beneficial goals of AI, while ensuring the preservation of the fundamental human rights of information access and privacy.
As an integrity agency, the OAIC believes that good privacy practices, timely and transparent access to information, and robust information governance will be vital to these objectives. These principles recognise the significant trust placed by the Australian people on the APS for its responsible stewardship and governance of AI to act in the interests of the public.
We will be publishing factsheets that provide operational guidance on key privacy, FOI and information management obligations that agencies must consider when using AI.
Keep an eye out for it – we think it’s a very timely and important initiative.
In regard to transparency, we have also done some important work on ensuring that the use of automated decision making is captured under the FOI Act.
Earlier this year we conducted a desktop review of 23 government agencies’ websites who are authorised to use ADM. The final report looked into whether agencies authorised to use ADM are being transparent about their use through the public reporting requirements contained in the FOI Act – and what regulatory guidance we can provide in response.
We found that information published about ADM use was limited, with only 17% disclosing its use. Information about decision-making and the exercise of agencies’ functions is important for the Australian community.
It improves integrity, accountability and trust. The benefits of utilising ADM technology in government will only be realised if risks are appropriately mitigated and trust is built with the Australian community.
A key enabler of trust is transparency, and by leading the way in transparency, the government can build public trust in the technology and ensure its benefits are shared widely across society. Wherever we turn in this arena, we see a demand for transparency and reinforcement of the need for human-centred decision making.
On the privacy front, it’s not long until the ADM transparency obligations come into effect in the Privacy Act.
Our 2026 Australian Community Attitude to Privacy Survey, or ACAPS, found that Australians remain cautious about the use of AI in decision-making that may affect them, with nearly all saying some conditions should be in place before it is used.
Australians expect organisations to be open about how AI is used and how decisions are made, with particular concern about decisions that may impact eligibility for financial support.
This helps explain the need for the Automated Decision-Making (ADM) transparency obligation in the Privacy Act 1988, which come into effect on 10 December. We recently reliance guidance to assist organisations to make the transition.
Privacy policies will need to contain information about substantially automated decisions, which significantly affect individuals’ rights or interests, including the kinds of decisions and kinds of personal information used.
This transparency is important because it helps stop bias and discrimination, enables contestability and consumers deserve to know when ADM is used instead of a human decision-maker.
Again, there is that emphasis on the human factor. We can never lose sight of the importance of human-centred values to drive the systems we design.
This is a very rich topic, which is central to our ability to deliver for the Australian public, and I look forward to further discussing it with you today.
