Singapore legalised machine reading in 2021. It is now asking what the machine gets to keep.
The Ministry of Law and the Intellectual Property Office of Singapore put that question to anyone who wants to answer it on 26 August, in a consultation paper split into a copyright half and a patent half. Submissions close at five in the afternoon on 22 October, through two FormSG forms and one ministry inbox.
Fifty-one days remain.
Earlier rounds were run with rights holders, technology companies, creative-industry representatives, lawyers and academics, which is to say with everyone who was already in the room. This round is open, and the room will mostly stay the same.
The copyright half will draw the submissions. The patent half is where the money sits.
Start with what Singapore already did.
Sections 243 and 244 of the Copyright Act 2021 give the country one of the widest text-and-data-mining exceptions anywhere: commercial training is covered, not merely research, and no licence agreement can contract the exception away. The condition the whole thing balances on is lawful access to the copy you learned from.
The consultation now asks whether that condition is clear enough, which is a polite way of asking a question with a lot of money attached. An operator fine-tuning on three years of customer transcripts under a model provider’s terms of service is relying on an answer nobody has given them.
Lawful access is four syllables that decide whether a training run was a licence or an infringement.
The patent questions are quieter and they travel further. One asks how inventorship should work across the range of human-AI interaction: framing the problem, choosing among machine-generated candidates, modifying a machine-generated solution. The other asks what happens when AI-generated technical disclosures are published at industrial volume.
That second question is the sleeper.
Prior art is a commons, and a commons can be flooded.
Novelty is assessed against everything already disclosed, so a flood of plausible machine-written technical documents raises the wall for every human filing that comes after it, whether or not the filer has ever touched a model. Prior art is a commons, and a commons can be flooded.
A Singapore hardware or diagnostics founder with a filing strategy has a stake in that paragraph and probably does not know the paragraph exists.
Lawful access is four syllables that decide whether a training run was a licence or an infringement.
Geography is doing real work here.
Beijing settled the first-order question in court rather than in a paper. In Li v Liu the Beijing Internet Court held that a Stable Diffusion image carried copyright, because the plaintiff’s prompts and parameter settings amounted to personalised choice and arrangement, and that the author was the human user rather than the developer or the model.1
A Chinese operator has had a rule to plan against since late 2023. A Singapore operator has a consultation.
The slower method produces the better rule, on the assumption that the people it will govern actually write in.
They usually do not.
Consultations of this kind are answered by trade associations, law firms with an IP practice, and two or three platforms with government-affairs teams whose job is precisely this. The submissions are good and they are also a particular kind of good: written by counsel for clients whose exposure is litigation rather than product. The founder whose exposure is the training set, the studio whose exposure is the output, and the engineer whose exposure is the prior-art wall are the parties least represented and most affected.
Fifty-one days, two forms, one inbox. The paper will close either way.
Footnotes
-
The Guangzhou Internet Court went the other direction on outputs a few months later, holding a platform liable when its model reproduced Ultraman. Chinese courts have now answered both halves of the question Singapore is still asking, which is either reassuring or not, depending on which half you are exposed to. ↩