Can Artificial Intelligence Be Protected Under Intellectual Property Law?

Introduction

Artificial intelligence is rapidly becoming the gold mine of the decade and perhaps the 21st century. Companies such as Nvidia, OpenAI and Microsoft have experienced unprecedented demand. Yet developing modern AI models, including ChatGPT, has taken decades. Artificial intelligence is complex, and this complexity lies at the heart of the legal question this article explores: can AI models be protected under intellectual property law?

This discussion focuses specifically on large language models. It examines the current IP protections available for computer programmes, the complications introduced by AI and whether LLMs can be conceptualised as protectable works.

Legally Protecting Computer Programmes

Intellectual property law protects intangible creations, including software. For a programme to be protected, it must be legally conceptualisable. Under the Database Directive 1996, a computer programme must be the author's own intellectual creation.

The protectable elements of a programme include:

• preparatory design material
• the set of algorithms
• the programme's restatement in a computer language
• the translation of source code into object code

The Court of Justice confirmed in SAS Institute Inc v World Programming Ltd that preparatory material, source code and object code are protectable. The Court also held that the "choice, sequence and combination of words, figures or mathematical concepts" are the areas in which an author exercises creativity and therefore produces an intellectual creation.

How LLMs Differ from Typical Computer Programmes

Large language models differ significantly from traditional programmes.

Structure

A typical programme is a series of algorithms that execute functions when certain conditions are met. LLMs, by contrast, are neural networks composed of interconnected nodes trained on vast datasets. They operate more like human cognitive processes, recognising, predicting and generating text.

Training

Although an LLM starts with base code, the data it is trained on shapes how it functions. If a model such as ChatGPT is trained on internet‑scale datasets, most of its outputs are not derived from the author's code but from patterns learned from external data.

This raises a key legal question: does an LLM's functioning extend beyond the author's intellectual creation?

Adaptation

LLMs can adapt in response to new data and user interactions. Their outputs may vary depending on previous prompts. This suggests that some aspects of an LLM's operation may exist independently of the author's creative choices. However, one could also argue that such behaviour is possible only because the base code permits it.

The tension lies in whether the author's intellectual creation includes the model's learned behaviour or whether that behaviour falls outside copyright protection.

Can Artificial Intelligence Be Protected Under Intellectual Property Law?

There is no definitive answer.

Copyright protects only the author's intellectual creation. The extent of an author's responsibility for an LLM's operation is ambiguous. Much of an LLM's behaviour depends on user input and, more significantly, the training datasets.

Ownership of Training Data

If an author does not have permission to use the datasets on which the LLM is trained, the resulting model is unlikely to represent the author's intellectual creation. This raises complex questions about data ownership, consent and the legality of scraping publicly available content.

Uncharted Territory

LLMs introduce novel challenges for copyright doctrine. Traditional software protection assumes that the author determines the programme's behaviour. LLMs undermine this assumption by learning from external data and generating outputs that may not be traceable to the author's creative choices.

The Only Certainty

Despite legal uncertainty, developers will continue advancing AI. Lawmakers will be equally eager to shape AI regulation and determine how intellectual property rights apply to increasingly autonomous systems.

Conclusion

Whether large language models can be protected under intellectual property law remains unresolved. Current doctrine protects only the author's intellectual creation, yet LLMs' functioning depends heavily on training data and user interaction. As AI evolves, lawmakers will need to confront these challenges and decide how far copyright protection can extend into machine‑generated behaviour.