Robot writers and angry authors: Copyright protection for AI books

On 19 September 2023, Amazon announced its online bookstore Kindle will impose a publishing limit of three books per day amid escalating concerns works using generative artificial intelligence would flood the market and obscure human writers.¹

This 'daring' move has not deterred the use of AI. Reflecting the rapid adoption of AI, since the public release of ChatGPT by OpenAI in 2022, 320 AI-related publishing startups have launched, with the aim of integrating generative AI into the writing, editing, and publishing process.²

With no sign of the use of AI in literature stopping, a critical question arises – are these AI-assisted and AI-generated works protected by copyright?

Creating an AI

Generative AI created to generate text are called large language models (LLMs) and work by being 'trained' on terabytes to petabytes of text data (such as novels and articles) to identify and encode common patterns and relationships (called 'parameters') found throughout the training data.

This information is 'tuned' by human contributors who both provide feedback on the output generated by AI and feed the LLM pre-labelled data. The data inputted at this stage contains pre-established parameters for the LLM to implement, such as a 'preferred' and 'non-preferred' output to a question.³ This extensive repeated process of training and tuning an LLM results in a neural network that is able to identify the 'best' output in a given context.

The process required to create an LLM is important in order to understand that when producing an output, generative AI does not apply independent thought or creativity like a human would. Instead, it is an amalgamation of the training data and what the algorithm, based on its tuning, determines to be the best output in response to a prompt.

The use of generative AI in literary work falls within a spectrum between using LLMs to generate the whole work, using LLMs to provide assistance in writing the work, and complete rejection of AI usage.

Copyright for literary works

While generative AI has been a major disruptor in many industries, in its current state, it is unlikely to significantly change existing copyright laws. Instead, pre-existing thresholds must be tested in the wake of this new technology.

Original literary, dramatic, musical and artistic works are protected under Part III of the Copyright Act 1968 and only protects original expressions of information from a human author.

This does not mean all works using computer programs (and by extension, LLMs) are automatically barred from copyright protection. Works that involve 'independent intellectual effort' or 'sufficient effort of a literary nature' by an individual in the creation process still enjoy a monopoly over the work.⁴

Unless AI gains consciousness, it is unlikely this requirement would be changed for generative AI, as giving protections not granted to other computer programs would be antithetical to established law.

Clear-cut cases

In Data Access Corporation v Powerflex Services Pty Ltd, ⁵ the High Court of Australia upheld the Full Federal Court of Australia's reasoning that a computer-generated work (a 'compression table') was copyright protected.

This was on the basis a human author, with 'substantial skill and judgment', coded the algorithm that created the work.⁶ This is contrasted to Telstra Corporation Ltd v Phone Directories Company Pty Ltd,⁷ where the Full Federal Court held the effort and skill applied by human contributors to find contacts to be put in a telephone directory did not result in copyright protection.

The final product (a compilation of the contact information) was generated with non-proprietary software and because information itself cannot be copyrighted, the effort by the human contributors in collecting information and making minor corrections was not materially reflected in the computer-generated expression of information.

For books that are entirely written by generative AI, like compilations, training and tuning the LLM would not fulfil the requisite human effort required to provide copyright protection. Edits and corrections that do not substantially change computer-generated work also cannot be regarded as authorship.⁸ These efforts do not include the 'creative spark' or 'requisite skill and judgement' for an original work under Part III.⁹

The use of LLMs solely for planning, brainstorming, or proofreading the human author's original work will not affect copyright protection, as this does not change the fact the final product is an original expression resulting from the author's own intellectual effort.

Uncertain cases

Where current law is indeterminate is where an author exercises both independent effort and use of LLM to write a book.

Current precedence for computer software cannot be applied to generative AI as unlike any other computer program, LLMs are able to adapt and change their output based on context.

If an LLM is used to fill gaps left by the author, it is able to train and tune itself to provide an output that mimics the author's writing style and tendencies. In this sense, there still has been material contribution by the author in the final work, as the author has written the text the LLM trains itself on. However, it remains that the author did not exercise 'substantial skill and judgment' for the sections written by an LLM.

For a copyright infringement to occur, it must use a 'substantial' part of the copyright material without permission.

If the copyright material is defined to be the entirety of the work, the consideration of individual intellectual effort will not be in relation to select sections of the work, but instead as a whole.

As a result, copyright protection may be determined based on either the proportionality or the significance of the contributions between the author and LLM. In the absence of clear legislation or case law, this will remain unclear.

Beyond books

Computer code and programs also fall within 'literary works' under Part III of the Copyright Act, making it subject to the same law as books.

For computer software, copyright protects the source code itself, not the function. In an era where a majority of developers use AI tools such as Microsoft Copilot to assist (or even replace) coding,¹⁰ any decisions made in relation to the use of AI will become vital to determine if a company will be able to protect its work under copyright.

The future

The adoption of generative AI and LLMs is increasing significantly year after year.¹¹

Legislators and courts will inevitably have to clarify aspects of copyright law that previously has never been tested. As human and AI work become more indistinguishable, it is recommended a detailed record of the contributions of the LLM be kept to distinguish AI assistance from independent intellectual effort.¹²

If you have any questions regarding copyright and the impact of AI, please contact our IP team.

(No LLMs were used in writing this blog).

Authors

Ben Coogan | Partner | +61 7 3338 7503 | bcoogan@tglaw.com.au

Jason Hong | Law Clerk

Notes

¹ https://www.kdpcommunity.com/s/article/Update-on-KDP-Title-Creation-Limits?language=en_US&forum=KDP%20Forum

² https://www.publishersweekly.com/pw/by-topic/international/Frankfurt-Book-Fair/article/96540-how-digital-innovation-both-strengthens-and-threatens-the-book-business.html

³ E.g. For an LLM being trained to be a virtual assistant, a question might be: 'How cold is it outside today?' A 'preferred' answer may be: 'Today it is not as cold as expected, with a high near 20°C and a low around 14°C', while a 'non-preferred' answer may be: 'It's not too cold today'. (Example based on https://platform.openai.com/docs/guides/fine-tuning#:~:text=for%20more%20information.-,Preference%20fine%2Dtuning,-Direct%20Preference%20Optimization)

⁴ IceTV Pty Ltd v Nine Network Australia Pty Ltd (2009) 239 CLR 458 [33], [99].

⁵ [1999] HCA 99.

⁶ Data Access Corporation v Powerflex Services Pty Ltd [1999] HCA 99 [123].

⁷ Telstra Corporation Ltd v Phone Directories Company Pty Ltd [2010] FCAFC 149 [89], [140].

⁸ Telstra Corporation Ltd v Phone Directories Company Pty Ltd [2010] FCAFC 149 [71].

⁹ Telstra Corporation Ltd v Phone Directories Company Pty Ltd [2010] FCAFC 149 [32].

¹⁰ https://www.gartner.com/en/newsroom/press-releases/2024-04-11-gartner-says-75-percent-of-enterprise-software-engineers-will-use-ai-code-assistants-by-2028; https://github.blog/news-insights/research/survey-reveals-ais-impact-on-the-developer-experience/; https://stackoverflow.blog/2024/09/23/where-developers-feel-ai-coding-tools-are-working-and-where-they-re-missing-the-mark/

¹¹ https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

¹² https://www.industry.gov.au/publications/australias-artificial-intelligence-ethics-principles/australias-ai-ethics-principles

‍

‍