Image generated using openart.ai

This page last updated 2025-04-03

We introduced some key ideas around responsible use of AI earlier in this site, touching upon ideas of ethics, academic integrity, research security, and privacy. Let's have a look at these in a bit more detail and consider how to discuss these ideas with our colleagues and students.

The University Intranet also has information about acting responsibly, summarised as our guardrails.

Privacy

We need to be careful about personally identifiable information or other sensitive or protected information about individuals when using generative AI. Apart from the risks for this information to enter training datasets or be leaked through cybercrime, there may be legal implications of providing private information to AI tools. When using generative AI, do not submit personal information about other people. Individuals should make an informed decision about whether or not they wish to enter their own information into generative AI tools. Where available, individuals should make a personal decision around opting out of data collection for AI training.

Current University guardrails state that we cannot provide mass-market publicly-available generative AI tools with confidential, personal, proprietary, or otherwise sensitive information. This includes a person's name, identifying numbers, other personal or health information, etc.

Some further reading:

Guidance on privacy and the use of commercially available AI products from the Office of the Australian Information Commissioner
Privacy in an AI Era: How Do We Protect Our Personal Information?

AI ethics

Bias

Because AI models are trained on data that humans have generated, this data can often reflect existing societal biases and prejudices. This may lead to AI models then reflecting human biases back when we interact with them. For example, AI models may privilege (or indeed de-privilege) certain races, genders, or perspectives on certain topics. Earlier on, AI bias could be compellingly demonstrated by asking image generation models to produce pictures of doctors or CEOs, who would be rendered predominantly as white and male. AI companies and labs are working to reduce bias in AI training data and AI outputs, but it's still important to recognise the bias inherent in systems trained on human data. It's also important to emphasise the importance of taking a critical lens to AI responses, and consider multiple sources and perspectives when working with AI.

Some further reading:

Intellectual property

Many of the more capable AI models are trained on huge data sources that contain trillions of words of text and countless images, videos, and audio clips. Many AI companies do not reveal the sources of their training data, and some are being taken to court by content creators for allegedly training on unlicensed works. Beyond training, there are live questions around what individual users are able to upload into generative AI tools. For example, is one allowed to upload a journal article to AI for summarisation? There are intellectual property laws to consider, which vary by jurisdiction, as well as licences with publishers to understand. Whether it is ethical to use AI models that have been potentially trained on unlicensed material, whether one should provide to AI tools material that is not our own, and also who holds the rights to AI-generated material - these are all important considerations when engaging with generative AI responsibly.

Current University guardrails state that we cannot provide mass-market publicly-available generative AI tools with copyright materials, such as the Library's eResources, the University's proprietary teaching resources, and student assessments.

Some further reading:

AI and Copyright in Australia from the Copyright Agency
Generative AI Has an Intellectual Property Problem

Environment

Generative AI models demand substantial computational power both when they are trained and when they are used for 'inference' (i.e. when generating a response to a user's request). The data centres where training and inference of AI models occur use considerable amounts of energy to run, and water for cooling. Many AI companies are looking into ways to make their AI models more efficient as this increases response speed and also reduces environmental impact. Often, newer AI models, or smaller AI models, use less energy and water during inference. At an individual level, we may wish to reduce non-essential use of AI tools, or opt to use slightly less powerful AI models if those will suffice. It's important to discuss the environmental challenges that AI poses, perhaps framing it in forward-looking and empowering ways that promote human agency and look towards practical solutions.

Some further reading:

Truth

The ability of generative AI tools, especially multimodal ones, to generate life-like content presents challenges around truth and veracity. Many scholars recognise that we are moving into a 'post-truth' society, where mistrust over public information is fuelled by the ease with which it is to generate fake content using generative AI. Deepfakes are an oftentimes nefarious manifestation of this - where generative AI is used to create fictional narratives, imagery, and audio of individuals. In these times, it is even more important to develop our critical thinking skills, and seek out reputable sources of information and advice.

Some further reading:

Academic integrity

Generative AI presents a complex challenge to academic integrity. Its increasing ability to create human-like multimodal content that achieves decent grades poses a risk to the integrity of our award programs. Beyond the assurance of learning angle, an over-reliance on generative AI may negatively impact students' development of key skills including critical thinking, creativity, and communication. It's clear from many studies, and a consideration of the ubiquity of generative AI tools, that banning, restriction, or detection are not the answer. Instead, the two-lane approach to assessment aims to preserve assessment validity (whether an assessment is actually measuring what we think it is measuring) whilst also ensuring that our programs remain relevant and help students develop disciplinary knowledge, skills, and dispositions alongside generative AI.

Some further reading:

AI Detectors Don’t Work. Here’s What to Do Instead.
The educational integrity enforcement pyramid: a new framework for challenging and responding to student cheating
Validity matters more than cheating

Research and data security

Some AI companies give users the option to opt out (or opt in) to having their inputs and outputs when using AI systems be retained for use as training data for future AI models. That is, your interactions with AI tools may - depending on the tool you use - be used to train AI. Moreover, the information you provide is likely stored on servers that may be outside the control or purview of your institution. This presents obvious concerns when it comes to the security of confidential data such as unpublished research findings, commercially sensitive information, and more. For example, it is tempting to upload interview transcripts to generative AI for summarisation and to identify emergent themes, or upload a spreadsheet to AI for quick exploratory analysis. Further, some funding bodies or outlets may stipulate restrictions or recommendations on how generative AI tools are used (or not).

It's therefore important when using AI tools to consider whether the information that you are providing to them are sensitive or protected in some way, and keep each other accountable for these potential risks to information security. It's also important to consider contextual guidelines around generative AI usage and the expectations from different stakeholders.

Some further reading:

What are the guidelines around research education?
Using generative AI in research: 10 tips