Japan vs Singapore AI regulations: An Open-book Exam Meets Driving Test

Japan and Singapore want AI regulation that protects people without slowing innovation, but they go about it very differently. Since joining AvePoint and helping drive its AI strategy, I’ve had to work within both systems and the contrast comes down to one metaphor: 

Japan gives you an open-book exam; Singapore gives you a driving test.

	Japan	Singapore
Approach	Soft guidelines	Testing and toolkits
Wants to see	Your reasoning	Your results
A mistake costs you	A public warning	A financial penalty
Public trust in AI	Less than 30%	50%
Copyright test	Intent	Lawful analysis
Best for startups	Training on data	Safety testing

How do Japan and Singapore regulate AI differently?

In Japan, the guidelines from the Ministry of Economy, Trade and Industry (METI) and the Ministry of Internal Affairs and Communications (MIC) ask whether you’ve thought it through properly, whether you have a process and whether you can explain your decisions later if someone asks.

It’s not a punitive approach, since there’s no fine attached. Even Japan’s own AI law, which came into force last year, has no penalty clause. Since you’re not really at risk of penalties, companies spend a lot of time building a paper trail. A big enterprise customer or a government partner will want to see that paper trail before they’ll work with you. 

Meanwhile, in Singapore, the Infocomm Media Development Authority (IMDA) has a free tool called AI Verify. You run your model through it, and it produces a report on bias, explainability and other AI-focused metrics. These sandboxes now let you put a finished product in front of real evaluators before it goes anywhere near a customer. Because of this, Singaporean AI companies spend their days in the lab calculating and simulating to present clients with data.

In a nutshell: both countries want proof. The difference is that Japan wants to see your reasoning, but Singapore expects to see your results.

Which approach builds more public trust?

One number surprised me here. KPMG and the University of Melbourne ran a global study with a participant base of over 48,000 people, asking whether they trusted AI.

• In Singapore: 50% said yes.
• In Japan: Less than 30% said yes. The report highlighted that Japanese respondents were more worried than optimistic.

This seemed counterintuitive at first. On paper, Japan’s approach is the gentler one, with less friction and basically no real penalties anywhere. But on further reflection, I realized that this can be interpreted in a far more cynical way. What I was seeing as hands-off oversight could be read by someone nervous as a system with no checks and balances.

In contrast, Singapore’s approach is visible, from the toolkits to the published sandbox results. The public has something to check, analyze and use as a basis to understand exactly what’s going on. To those wary of the advent of AI, a transparent system is what instills trust, and that’s why Singapore’s confidence rate was stronger.

How do they handle copyright and training data?

This is one of my favorite parts of the comparison. The two countries get to almost the same place through completely different logic.

In Japan, if a machine uses a copyrighted work for analysis, and no human consumes that work in the process, that’s generally fine without permission. By regulatory standards, this is one of the more permissive laws around the world, and it’s probably part of why global AI labs pay so much attention to Japan.

Meanwhile in Singapore, a specific exception to copyright law covers computational data analysis. Here’s the detail I found particularly interesting: A company can’t contract it away. The exception survives even if a website’s terms and conditions say otherwise, as long as you got the content legally in the first place.

Japan’s test is basically about intent. Are you appreciating this like a human or just crunching it like a machine? Singapore’s test is more mechanical: Was this lawful analysis, yes or no?

The end point is that both sets of laws leave content far more open to AI training than anything in Europe.

A key point here just to highlight how permissive both countries are: these laws also cover personal data. Both countries let you train or improve AI on people’s information without going back for fresh consent every time, as long as it’s reasonable and you’ve anonymized where you can.

Okay, but then what happens when a company misuses personal data?

The enforcement style mirrors everything else about these two countries.

When ChatGPT first blew up, Japan’s privacy regulator sent OpenAI guidance: Don’t collect sensitive personal data for training without consent, and delete any you already have. That’s very Japanese, to me. It reminds me of  getting called into the principal’s office.

Singapore is much stricter. Expect some financial penalties when you fall afoul of their regulatory agencies:

• Marina Bay Sands: Fined more than 300,000 Singapore dollars after a system migration error exposed the data of hundreds of thousands of customers.
• SingHealth: A breach exposed the personal data of more than 1.5 million patients. SingHealth is Singapore’s largest group of public health care institutions. The regulator fined both it and IHiS, the IT vendor that ran its systems.

So if you make a mistake in Japan today, you probably get a public warning and some reputational heat, but in Singapore, there’s a good chance it costs you money. However, that gap might close soon. Japan has an amendment working its way through its legislative branches that would add financial penalties to its privacy law for the first time — bringing enforcement of AI offences closer to Singapore.

How do they compare against the EU AI Act?

Both governments have made the comparison directly, which made my job easier. There was no need to read between the lines.

For Japan, the government says it wants to be the most AI-friendly country in the world. Its own officials describe the approach as deliberately lighter than Europe’s.

In Singapore, the approach is more nuanced, and I find it more interesting personally. Last year, our Minister for Digital Development made a public statement that the law cannot regulate every instance of how autonomous AI agents behave. Rather, the smart move is to build trust through assurance and give up on controlling every outcome in advance.

That’s a real departure from how both Brussels and Tokyo tend to think about this.

I like using a football analogy for this: Europe’s playing defense. It’s telling you: here’s what you can’t do. Japan and Singapore are both playing offense, but in different formations. Japan’s pitch is a recruiting ad: here’s all the reasons you should give us a shot. And Singapore’s is closer to a coaching manual, straddling both lines: try this out, but we’ve still got some rules to follow.

Where should an AI startup test a data-heavy product?

Right now, I’d tell a startup to start in Singapore, for one very practical reason: The infrastructure is already built.

• Several AI sandboxes already exist. One recent pilot paired 16 specialist testers with 17 companies deploying generative AI across finance, health care and other sectors.
• The government’s own numbers from an earlier sandbox for small and medium-sized businesses showed close to 80% kept using what they built after the trial ended. That’s a real signal.

While Japan isn’t lacking for sandboxes, they were built for an earlier generation of problems: fintech, self-driving cars and blockchain settlement. When I went through the list of companies that have used them, I found no flagship AI cases. The infrastructure exists, but it hasn’t been pointed at AI yet, which is a real issue if you’re coming with an expectation of ready-made tools waiting for you.

That being said, if your real bottleneck is data, Japan is the place you’ll want to start. As mentioned already, its copyright law is one of the most generous in the world for training on existing content. Depending on your needs, this could outweigh the outdated sandbox issues.

So if I’m being straight with a founder, I’d say run your safety testing in Singapore, and if you need to train on a mountain of Japanese language content, do that piece in Japan.

But a final word of warning to round out my thoughts. For the future-conscious: keep an eye on the lawsuit two of Japan’s biggest newspapers just filed against an AI company. That case is going to draw the actual line on how far that generosity really goes.