Part of the AI Regulation News hub.
Singapore's Ministry of Law and IPOS state that the computational data analysis exception already permits generative AI training including for commercial purposes, but does not reach infringing output, and are consulting on both
Most governments asking whether their copyright law permits AI training are asking an open question. Singapore is not. It says the answer is already clear, tells you what it is, and then asks whether anyone needs it spelled out more plainly.
Bottom line: A public consultation, not a legislative proposal. Its most quotable feature is a statement of the government's existing legal position: the computational data analysis exception may be used to develop generative AI models, with no express limitation on purpose, including commercial purposes.
Who this affects: Anyone training or fine-tuning models on copyright works with a Singapore nexus, AI deployers assessing output-side infringement risk, patent applicants whose inventive process involves AI, and rights owners in the region.
Issue date: Consultation period 26 August to 22 October 2026.
What changed: No law changed. What is new is the position being stated on the record, together with an equally explicit statement of where the exception stops: it does not reach the output phase.
Analysis: Note the asymmetry this creates. Training is treated as settled and permitted; output is treated as unresolved and outside the exception. That is close to the opposite of the framing in most jurisdictions, where training is the contested question and output infringement is handled by ordinary principles. It also puts Singapore in visible contrast with Australia, whose Government has said it is not considering a text and data mining exception.
Primary sources: Public Consultation on AI and Singapore's Intellectual Property Regime, consultation paper
- Instrument
- Public Consultation on Artificial Intelligence and Singapore's Intellectual Property Regime
- Authority
- Ministry of Law and the Intellectual Property Office of Singapore
- Jurisdiction
- Singapore
- Status
- Open consultation. Period 26 August to 22 October 2026.
- Bindingness
- None. A consultation paper. The statement about the computational data analysis exception describes the Government's reading of existing law, not a new rule, and it is not a judicial determination.
- Scope
- Copyright in three areas and patents in two. Copyright: certainty and accountability in AI training; risk management in AI deployment and use; and the nature of human creativity. Patents: inventorship across human-AI interaction, and the effect of large-scale AI-generated technical disclosure on prior art.
- How to respond
- Via FormSG at go.gov.sg/ai-ip2026-copyright for copyright and go.gov.sg/ai-ip2026-patents for patents, or by email to MinLaw.
- Editorial Note
- Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.
- Primary source
- https://www.mlaw.gov.sg/files/Public_Consultation_on_AI_and_IP_20260826.pdf
The statement on training
Singapore's Copyright Act contains an exception for computational data analysis, commonly abbreviated to CDA, which predates the generative AI boom. The question everywhere else has been whether such provisions reach the training of generative models, and whether they reach commercial training.
The paper answers both. It records that the legal position on those aspects is clear: the CDA exception may be used to develop generative AI models, and there is no express limitation on the purpose for which CDA is undertaken, including for commercial purposes.
It reaches that reading by a route worth noting, describing a purposive interpretation consistent with the design and policy intent to support AI and data-driven innovation in Singapore, and invoking the established principle that a statute should be interpreted taking into account changes since enactment, including technological developments that could not reasonably have been contemplated at the time.
Having said the position is clear, the Government nonetheless asks whether greater clarity is needed on the exception's scope and applications. That is the consultation's actual posture on training: not what should the law be, but does the law need to say more plainly what it already means.
Where the exception stops
The same passage marks the boundary, and it is the part a deployer should read first.
The paper identifies issues at the output phase, where there may be further uses of copyright works other than AI training, arising when models are deployed and produce infringing output pursuant to users' prompts. Its conclusion is direct: the exception would not apply here, as such uses do not involve computational data analysis.
So the protection is tied to the activity, not to the actor or the product. Training on works is CDA and is covered. A deployed model reproducing a work in response to a prompt is not CDA, and nothing in the exception reaches it.
The consultation follows that boundary with the questions it implies: how existing legal principles should apply when AI-generated output infringes copyright, how responsibility should be assessed as between AI developers, deployers and end users, and what technical measures might minimise output infringement risk in a way that is proportionate and commercially viable.
The patents half, which is the less obvious one
The copyright material will attract the attention, but the patents part asks two questions that are harder and less discussed.
The first is inventorship. The paper asks how existing inventorship principles should apply across the spectrum of human-AI interactions in the inventive process, and it names the specific points on that spectrum: problem formulation, selection from AI-generated outputs, and human modification of AI-generated technical solutions.
That framing is more useful than the familiar binary about whether an AI can be named as an inventor. It treats invention as a process with identifiable human contributions and asks which of them suffice, which is the question that actually arises in prosecution.
The second is prior art, and it is the sleeper issue. The paper asks how the large-scale publication of AI-generated technical disclosures may affect the prior art landscape, including patent searches, examination, patentability assessments and incentives for innovation as AI-generated material becomes more publicly available. If disclosures can be generated and published at scale, the state of the art can be moved deliberately, and every downstream patentability assessment inherits the consequences.
How to use this outside Singapore
The direct effect is limited to Singapore, and the statement is the Government's reading of its own statute rather than a court's. It should be cited for what it is.
Its comparative value is higher. Singapore is stating that an existing computational analysis exception, drafted before generative AI, already accommodates commercial model training on a purposive reading. Australia's Government has said it is not considering a text and data mining exception, while continuing to consult on other copyright and AI questions, and other jurisdictions are still litigating the point. Anyone arguing about the reach of a similar provision now has a government stating the permissive reading plainly. The contrast is between an enacted exception read permissively and a jurisdiction declining to introduce one, which is narrower than a rejection of reform.
For a business, the practical division is clean and worth adopting internally even where Singapore law does not apply: treat training exposure and output exposure as separate risks with separate controls, because this paper treats them as separate legal questions with different answers.
Responses are due by 22 October 2026, through separate channels for copyright and patents. An organisation with views on both should note that the paper splits the submission routes, which suggests the two halves will be analysed separately.
The sentence to keep is the Government's own: the computational data analysis exception may be used to develop generative AI models, and there is no express limitation on the purpose for which CDA is undertaken, including for commercial purposes. That is a state saying its pre-existing exception already permits commercial generative AI training, arrived at by purposive interpretation, and it stands in visible contrast with jurisdictions that declined to legislate an equivalent. The boundary is stated just as plainly: the exception would not apply at the output phase, where a deployed model produces infringing output in response to prompts, because such uses do not involve computational data analysis. Treat training risk and output risk as two separate problems. And do not overlook the patents half, which asks how inventorship principles apply to problem formulation, selection from AI outputs and human modification, and how large-scale publication of AI-generated technical disclosure will distort the prior art landscape. Consultation closes 22 October 2026.
Source File
https://www.mlaw.gov.sg/files/Public_Consultation_on_AI_and_IP_20260826.pdf
Open the consultation paper and confirm four things: the cover stating a consultation period of 26 Aug to 22 Oct 2026 and naming the Ministry of Law and IPOS; the executive summary setting out three copyright areas and two patent areas; in Part III, the statement that the legal position is clear, that the CDA exception may be used to develop generative AI models and that there is no express limitation on the purpose for which CDA is undertaken including for commercial purposes; and, in the same discussion, the statement that the exception would not apply at the output phase because such uses do not involve CDA.
The legal position on the above two aspects is clear - the CDA exception may be used to develop generative AI models, and there is no express limitation on the purpose for which CDA is undertaken (including for commercial purposes). Ministry of Law and IPOS, Public Consultation on AI and Singapore's Intellectual Property Regime, 26 August 2026
FAQ
Does Singapore permit training generative AI on copyright works?
The Government's stated position in this paper is that the computational data analysis exception may be used to develop generative AI models, with no express limitation on purpose, including commercial purposes. That is its reading of the existing Copyright Act, not a new rule and not a court ruling.
Does the exception cover infringing output?
No. The paper states that the exception would not apply at the output phase, where a deployed model produces infringing output pursuant to users' prompts, because such uses do not involve computational data analysis.
Is this a change in the law?
No. It is a consultation paper describing the Government's reading of existing law and asking whether greater clarity is needed on the exception's scope and applications, among other questions.
What does it ask about patents?
Two things: how existing inventorship principles should apply across human-AI interaction in the inventive process, naming problem formulation, selection from AI-generated outputs and human modification of AI-generated technical solutions; and how large-scale publication of AI-generated technical disclosures may affect the prior art landscape, including searches, examination and patentability assessments.
How was the permissive reading reached?
By purposive interpretation. The paper points to the design and policy intent of supporting AI and data-driven innovation, and to the principle that a statute should be interpreted taking account of changes since enactment, including technological developments that could not reasonably have been foreseen.
When and how do I respond?
By 22 October 2026, via FormSG at go.gov.sg/ai-ip2026-copyright for copyright and go.gov.sg/ai-ip2026-patents for patents, or by email to MinLaw. The two halves use separate submission routes.
Related briefings
Sponsored Training
Practical AI training for regulated professionals, built around verification, documentation and a defensible process. See the courses.