AI Q&A - Copyright Agency

About Generative AI tools

What is a generative artificial intelligence (GenAI) tool?

A GenAI tool is software that can produce new content (e.g. writing, an image, music) that is similar to human-authored content. For example, a person may ask (‘prompt’) a GenAI tool to write a short description of outback Australia, or an image of an avocado doing yoga on the beach.

How is GenAI different to other types of AI?

Other forms of AI can perform tasks like analysis of large datasets based on ‘machine learning’ but do not generate new content that emulates content created by human authors.

Some Generative AI terminology

What is a Large Language Model (LLM)?

An LLM is software that can recognise text and generate new text. This enables the software to ‘understand’ a request, and to respond to the request with newly generated content (rather than ‘retrieved’ pre-existing content).

LLMs can perform these tasks because they were ‘trained’ on vast amounts of material, mostly from the internet. The ‘training’ enables the software to recognise and generate patterns based on probabilities in human language.

What is a hallucination?

In response to a prompt, an LLM can generate outputs that are factually incorrect, nonsensical, or unrelated to the given context, despite sounding plausible. These are referred to as ‘hallucinations’.

What is ‘pre-training’?

People use the term ‘pre-training’ to describe the development (‘training’) of a language model that can be enhanced or customised for specific purposes or applications.

Language models can be customised in different ways. They include ‘retrieval augmented generation’ (RAG) and ‘fine-tuning’.

What is Retrieval Augmented Generation (RAG)?

RAG is a process that uses a specialised knowledge base, such as a database of company policy documents, or a dataset of peer-reviewed journal articles on a particular topic. When a person makes a request of an GenAI tool, the request is ‘contextualised’ with reference to the knowledge base, so that the response (output) is more likely to be relevant.

AI tools that use RAG can provide information in outputs, such as citations, about the information that the output was drawn from. RAG can reduce ‘hallucinations’.

What is fine-tuning?

Fine-tuning is a process of customising a language model to produce more useful or relevant outputs for a particular type of users, by further ‘training’ of the model using a specialised dataset.

For example, a Local Health Area could fine-tune a language model to better interpret requests (prompts) from staff, and to generate more appropriate outputs. It would do this by curating a high-quality dataset containing medical records, patient histories, symptoms, lab results, and diagnoses. This dataset is carefully annotated by medical experts to ensure accuracy. The LLM is further trained on this medical dataset. It learns to recognise patterns, understand medical terminology, and associate symptoms with likely diagnoses.

What is the difference between a language model and a GenAI tool?

A GenAI tool uses a language model to interpret and generate text.

For example, ChatGPT is a general-purpose AI tool that uses OpenAI’s language model to interpret and respond to questions from the general public. Microsoft’s Copilot software also uses OpenAI’s language model, but in a more customised way.

Gen AI tools can generate various types of content, including text, images, audio, video, and 3D models.

What is a ‘diffusion model’?

A ‘diffusion model’ is software that underpins many GenAI image-generation tools. Like LLMs, diffusion models have been developed using a vast amount of content, mostly images from the internet.

Different ways that copyright material can be used for generative ai

What are the different ways that copyright material can be used for GenAI?

Very broadly, there are three different areas in which content is being used in connection with GenAI:

General-purpose GenAI software (LLMs and image-generation software) that have been developed using vast amounts of content, mostly from the internet
Smaller, targeted development using smaller, specialised datasets, for example for RAG and fine-tuning
Users of GenAI tools including copyright material in requests (prompts) to GenAI tools: e.g. ‘please summarise this journal article’.

How does copyright apply to the use of other people’s material for GenAI?

Generally, copyright laws apply on a country-by-country basis, depending on where the activity occurred.

Various types of general-purpose GenAI software, using vast amounts of content, have been developed in the United States. Generally, this means that US law applies, and nearly all the court cases involving GenAI development are taking place in the US.

For a range of reasons, it can be difficult for Australians to take legal action for copyright infringement if their works were used for GenAI development in the US. Copyright Agency and others in the creative industries are therefore urging the Australian government to look at other ways for compensating Australians whose works were used without consent or compensation.

Australian copyright law applies to smaller scale development that is occurring, and is likely to continue, in Australia. The law applies to GenAI development in a similar way that it applies to any other use of other people’s material. GenAI is relatively new technology, but Australia’s copyright law is generally ‘technology neutral’, and its principles apply to this new technology as they have to other technological developments.

managing how works are used for generative ai development

What can creators do about works already used for offshore LLM development?

For a number of reasons, it can be difficult for Australian creators to take legal action against companies that used their works for training LLMs in the US.

Copyright Agency and others in Australia’s creative industries are asking the Australian government to introduce a law or mechanism that would compensate Australians whose content has been used for GenAI development offshore without their consent or compensation.

If you are interested in assisting with advocacy to the Government, please contact the Member Services team.

How can creators signal that their work should not be used for future LLM development?

This can be difficult for works that are available on the internet. Although it is possible to use a 30 year old signal (robots.txt) to request that any URL is not used for training, respecting this signal is not mandatory.

How can I be remunerated for the use of my work for AI development?

Use of Australians’ works for AI development in Australia can be licensed, for a fee, in the same way that other uses of works can be licensed for a fee.

Copyright Agency is developing licensing solutions for ethical use of other people’s material in connection with GenAI in Australia.

The use of Australians’ material in offshore GenAI development using vast amounts of content remains unresolved. Some GenAI developers have entered into some copyright licensing deals, but the past use of copyright material remains the subject of multiple lawsuits.

Will the Australian government take steps to stop social media platforms using artists’ works for AI training?

It is not yet clear what steps the Government will take.

responsible use of generative ai tools

How can I use GenAI tools to assist with my writing?

Things to consider if using GenAI tools to assist with your writing include:

disclosing that you have used a GenAI tool (e.g. to your employer, client, publisher or readers); and
making sure that you understand the functionality and settings of the AI tool that you are using so that you can avoid any unintended use of any data you provide (e.g. your prompts)

How can I make sure that material used in connection with a GenAI tool is not captured outside my workplace?

You should make sure that you understand the functionality and settings of the AI tool that you are using so that you can avoid any unintended use of any data you provide. You may need to choose a tool that enables you to meet your workplace’s policies relating to data security and confidentiality, and the copyright licensing arrangements that apply to your workplace.

How does GenAI affect teachers and teaching?

The Australian Framework for Generative AI in Schools sets out guidelines for the responsible and ethical use of generative AI tools in the school sector.

Australian teachers can use GenAI tools for teaching, consistently with the framework: see here.

How can I use AI tools to interpret copyright materials without infringing copyright?

You can do this if you are covered by copyright licensing arrangements. These currently apply to people working in the education and government sectors. Copyright Agency is developing similar arrangements for the business sector.

GENERATIVE AI OUTPUTS

Can GenAI outputs be covered by copyright?

A GenAI output can be covered by copyright if it resulted from sufficient human creative input, in a similar way that works created using other technological tools can be covered by copyright.

Who owns copyright in GenAI outputs?

As a general rule, the first owner of copyright is usually the human creator of the work. If the work was created by an employee as part of their job, the copyright will usually be owned by the employer.

first nations creators

How could GenAI affect First Nations artists?

There can be additional implications of GenAI for First Nations creators because of their Indigenous and Cultural Intellectual Property.

The Government has recognised this, and there are specific references to this issue in various reports and documents by and for the Government.

policy and regulatory settings

What is the regulatory environment for GenAI in Australia?

Currently, GenAI is covered by laws and regulations that also apply to other types of activities and technologies. However, the Government is looking at whether there should be some AI-specific laws.

Australia’s copyright legislation applies to AI-related activity as a matter of general application: there are currently no AI-specific provisions. For example, it allows some AI-related activities under statutory licence schemes for education and government, subject to fair compensation. It also allows for licensing arrangements to enable businesses, for example, to use other people’s material in connection with AI tools.

The application of Australian copyright law to AI is being reviewed by the Attorney-General’s Department’s Copyright and Artificial Intelligence Reference Group (CAIRG): see here. Copyright Agency is a member of CAIRG, and of its steering committee.

Is the copyright regime for GenAI different in the US to Australia?

Most countries, including Australia, are signatories to international copyright treaties, like the Berne Convention, that set minimum standards for copyright laws in each country. This means that Australia’s copyright law is similar in most respects to copyright laws in other countries. There are, however, some differences. For example, US copyright law has an exception for ‘fair use’, that AI developers in the US claim applies to their use of other people’s content for ‘training’. This will be a matter for the US courts to decide in the numerous court cases currently in train.

What is Copyright Agency’s policy position on AI?

Broadly, Copyright Agency’s position is:

the current copyright legislation allows the ethical development of AI in Australia, including with fair compensation to Australians whose content is used;
the Australian Government should resist calls by multinational companies to water down Australia’s copyright legislation to allow AI developers to use other people’s content without compensation; and
the Australian government should introduce a law or mechanism that would compensate Australians whose content has been used for AI development in other countries without compensation

Share Tweet