Singapore's Computational Data Analysis Copyright Exception

Singapore legislated for machine analysis of copyright works before almost anyone else, and the exception it produced is unusual in two respects: it has no opt-out, and it cannot be contracted away. Both features make it a genuinely different instrument from the European text-and-data-mining regime it is often compared to.

The provision and when it came into force

The Copyright Act 2021 replaced Singapore’s Copyright Act 1987 and commenced on 21 November 2021. Division 8 of Part 5 contains a permitted use for computational data analysis at sections 243 and 244; the parallel provision for performances sits alongside it. The consolidated statute is published by the Attorney-General’s Chambers at Singapore Statutes Online, which is the text to read rather than any summary of it, including this one.

“Computational data analysis” is defined in the Act, and the definition is deliberately broad: it covers using a computer program to identify, extract and analyse information or data from the work, and expressly contemplates using the work as an example of a type of information or data for the purpose of improving the functioning of a computer program in relation to that type. That second limb is machine learning described without naming it, drafted in 2021, and it is why the provision reaches model training cleanly rather than by analogy.

The four conditions

The permitted use is not unconditional. In substance the user must satisfy all of the following, and the exception is unavailable if any one fails:

• Purpose. The copy must be made for computational data analysis, or for preparing the work for such analysis. A copy made for some other purpose and later repurposed is not covered by the act of copying that created it.
• No onward supply. The copy must not be supplied to anyone else except for the purpose of verifying the results of the analysis, or for collaborative research or study relating to the purpose for which the copy was made. This is the condition that stops the exception being a route to redistributing a corpus.
• Lawful access. The user must have had lawful access to the work from which the copy was made. See below — this is where the exception actually bites.
• No circumvention of paywalls or access controls. A copy obtained by breaching an access restriction is not lawfully accessed, and a work obtained from an infringing source falls outside the exception even if the user did not create the infringing copy.

Lawful access is the whole of it

If you take one operational point away, take this one: the Singapore exception is a provenance rule dressed as a copyright rule. It does not ask what your model does, how transformative your use is, or whether the rightsholder objects. It asks how you got the file.

Lawful access covers works you have paid for, works made freely available by or with the authority of the rightsholder, works you access under an institutional licence, and works accessible on a site without a paywall. It does not cover works obtained by circumventing a paywall or an access control, works downloaded from a source you knew or ought to have known was infringing, or works reached in breach of a login-gated agreement. The consequence is that a shadow-library corpus is outside the exception in Singapore for the same reason it was the losing half of the US analysis in Bartz v Anthropic — see the four-factor test applied to training.

The engineering corollary is that provenance metadata must be captured at collection time, per item: source URL, date, access method, and the basis on which access was lawful. Reconstructing that later is not possible in any meaningful sense, and the burden of establishing the exception rests on the person relying on it.

Why contract cannot switch it off

The Copyright Act 2021 treats certain permitted uses as not excludable by agreement, and computational data analysis is among them: a term in a contract that purports to exclude or restrict the exception is, to that extent, of no effect. This is the most significant structural difference from the European scheme, and it cuts in the opposite direction from the rest of the world’s drift.

Under Article 4(3) of the EU’s CDSM Directive, a rightsholder can reserve TDM rights and thereby remove the general exception. Singapore has no equivalent opt-out, and it goes further by neutralising terms of service that would achieve the same result. A website’s terms saying “no automated analysis of this content” therefore do not, on their own, defeat the exception — although they may still bear on whether access was lawful in the first place if access itself was conditional on accepting them, and that interaction has not been worked through by a Singapore court.

What the exception does not do

It permits copying for analysis. It does not do any of the following, and conflating them is the most common mistake:

• It does not permit publishing or distributing the corpus, beyond the narrow verification and collaboration carve-outs.
• It does not address the model’s outputs. If a model emits a substantial part of a protected work, that is an ordinary infringement question and the analysis exception is irrelevant to it.
• It does not resolve database rights, contract claims unrelated to the copyright exception, personal data obligations under the Personal Data Protection Act, or confidentiality.
• It does not travel. Copies made on infrastructure in another jurisdiction are governed by that jurisdiction’s law, and most jurisdictions have no equivalent provision.

For the comparable Japanese provision, which is broader still in what it permits but conditioned differently, see Article 30-4 of the Japanese Copyright Act, and for Singapore’s governance regime as distinct from its copyright law, the Model AI Governance Framework.