Copyright and AI model training: The balancing act as seen from Singapore

Should our intellectual property laws evolve further?

A month in the artificial intelligence (AI) world can feel like a decade, so I wonder whether anyone still remembers OpenAI’s inspired Studio Ghibli trend way back in March 2025?  

It might be forgotten now under the sediment of the rapidly flowing river of AI development, but the undercurrents and tensions between the creative industry and technology industry remains strong.

A “tech-culture war”?
There is a fertile ground for the ongoing “tech-culture” war (for want of a better term) between the AI evangelists and human-centric creatives.

Within each “side,” there is a spectrum of views, and there can be something of a middle common ground, so at the risk of oversimplifying and, as a thumbnail or caricature the extremists of each “side,” you can see arguments lined up like so: extremists on the human-centric creative side argue generative AI is too rapidly displacing economic models that have kept human creativity alive, and that its effects include generating “AI-slop” that poisons the waters of human discourse, cheapening and, ultimately, bankrupting social and spiritual life.

Extremists on the AI evangelists side argue generative AIis no different from the camera replacing paint and palette, that new avenues of expression are being  pioneered through this technology, whilst augmenting and democratising creativity everywhere, freeing artists from the drudgery of pointless “manual” labour.

Singapore’s intellectual propertly laws
But if the point of any intellectual property law is to encourage the development and responsible adoption of technology whilst balancing the interests of content owners and their fair share of the fruit of their labour, then the law must align to how the technology works.

Into this breach stepped Singapore with its “AI-friendly” copyright framework as evidenced by the Computational Data Analysis provisions in the Copyright Act 2021 (“CDA Provisions”). Let’s review briefly.

To begin with, let us remind ourselves that any court would pay special attention to the language used and to assess whether the facts of any case can fit the exception and this will vary according to the specific case.

First, we need to know what “computational data analysis” means. This is defined in section 243 in the CDA Provisions and, for the sake of argument, we can accept that this covers the identification, extraction and analysis you see in text data mining, and even, to a certain extent at least, the various steps in the preparation of a work for training a generative AI model – e.g. tokenisation, embedding, etc. So, for this article, “CDA” will be our shorthand for this process.

Now, let us consider the wording of the CDA Provisions to see what is permitted. The CDA Provisions permit anyone to “make a copy” of a work or a recording of a protected performance for the purpose of CDA or preparing the work or recording for CDA to happen (section 241(1)(a) of the Act); to “communicate” a work or recording of performance in accordance with the CDA Provisions so as to verify the results or as part of collaborative research (section 241(4) of the Act).

The act of executing the CDA itself is not mentioned explicitly as a permitted act in the language.

We know that, in the large language model (LLM) training process, developers do more than merely “copy” a work when they are preparing the LLM. Tokenisation, for example, is a key step. Broadly, the process of tokenisation, is to render the work in a form which can be trained on.

At first glance, it seems that tokenisation is more than merely “copying” and merely “communicating”. It may resemble the making of an “adaptation” (see s17(c) of the Act) - i.e. creating “a version of the work (whether or not in the language, code or notation in which the work was originally expressed) that is not a copy of the work”.

This article should not be read as concluding on this point – after all, we would need a technical expert to review and opine, and a court to confirm. But what if this is correct? Would the CDA Provisions be sufficient to cover the world of generative AI?

Were the CDA Provisions not intended for generative AI?
We should remember that the CDA provisions were originally passed the year before generative AI was globally popularised.

Whilst it may well have been that generative AI was in the mind of policy makers and legislative draftsmen, perhaps it would help to be clearer as to whether the CDA Provisions are intended to cover all acts of CDA itself, and not merely “copying” and “communicating.”

It might be also the case that a court would look at the broader context and consider that the CDA Provisions are indeed sufficient. A court could do so by taking a purposive approach and not confine itself to just the sanctioned acts described in the verbs “copy” and “communicate”.

But is there any harm in “catching up” the language and making it clearer?

The conundrum
This takes us back to the actual underlying problem. Developments in AI move fast. Can any country’s legislation keep up?

Are there any potential “gaps” that are currently emerging in a way that the intended policy outcome and the legislative intent as expressed in statutes are not aligned?

Does this mean we need to rethink how laws are reviewed and edited, if only in the AI space? Would there be ways to do this faster and be more nimble?

Should our intellectual property laws evolve further?

Singapore is famous for its practical, efficient and effective policy-to-law making track record. If there is one country that can get it done, it is Singapore.

But I do think our work is not done.

Indeed, nothing in AI, which changes daily, or the law, which must develop alongside it, is ever done.