I'm Esa Koskinen, working as a volunteer director of AI Safety Tokyo.
Epistemic status: estimates by an interested party; I run one of the organizations discussed. The model behind the numbers is public, and I'll edit corrections into the post.
Lately I've spent ~100 hours investigating financial information and grant applications made by AI safety organizations and independent researchers in Japan. I was somewhat disappointed by the results, and wanted to share a summary of my findings with the rest of the world.
I estimate that Japan has a pretty urgent funding gap of ~2.1 million USD, could employ AI safety researchers about 1.8-2.3x cheaper than the US, and in general has many neglected opportunities for impact. (Source spreadsheet; yen converted using 1-month average to smooth fluctuation.)
In 2027, I expect Japanese AI safety organizations could hire ~35 researchers and ~10 staff members with good marginal effectiveness. However, despite lots of interest and demand, the organizations don't have enough funding to make it happen.
If you want to help out, please leave a comment, share this post with your friends, or drop me a line personally at esakoskin2@gmail.com. I don't expect LW readers to be familiar with AI safety in Japan and the situation has changed recently, so a quick overview might be useful.
AI safety gatherings in Japan started in 2018, but became much more regular and structured after AI Safety Tokyo was established in Jan 2023. Since then, multiple organizations have spun up for doing education, governance, and research.
In order of size, along with a short summary of their recent approach and challenges:
If you look at the current challenges above, there is a common theme: Japanese AI safety organizations tend to lack funding. Most groups are run by volunteers, many of which would love to work on AI safety full-time. However, having to run on volunteers leads to constant strain from operational bottlenecks, and most orgs have to discard promising opportunities due to lack of time. Despite this, Japan has had ~80 AI safety related events in the past year, ran the ARENA curriculum, and brought many new people into the field. We have also run annual international conferences since 2023:
The 2026 relocation cuts both ways: the conference born in Tokyo was good enough that Noeon Research was able to pitch it in Oxford, and Tokyo didn’t have the funded ecosystem to keep it. That is one of the failure modes this post is about.
At the moment, the community is facing four main bottlenecks, each of which is a neglected opportunity for impact:
1. Funding: During the past two years, we in Japan have put significant effort into creating a pipeline for scaling up the field and getting people to do more useful research in it. However, we don't have the funding to pay salaries, restricting work to only what can be done in one's spare time, as a volunteer.
The obvious answer would be for people to apply for more grants. The reasons why this is not happening and what could be done about it are elaborated in the following sections.2. Retention: Many people come to Japan from overseas, do independent AI safety research for 6 months, realize they can't get funding, and leave the field. My personal contacts include about a dozen such people. I've had discussions with roughly a third of the AI safety community here, so I expect the real number to be 2-4x higher.
This can be super damaging: some of these people have done MATS/ARENA, taken part in a fellowship, or completed other high-quality programs like SPAR or TARA. The loss of their talent and knowledge presents a huge sunk cost, as well as a wasted opportunity. It's also demotivating for the rest of the community.
Japan as a country has a lot of appeal, and I think we could plausibly become Asia's largest AI safety hub. I've spoken to dozens of people in the field who would prefer to live and work here instead of the US or London, if only the pathway for being funded was clearer and more realistic.3. Awareness: For politicians and the general public, AI risks - especially existential ones - are almost unheard of. Japanese politicians are under-educated on AI, and rarely push back on issues regarding it. Many decision-makers have a willingness to learn more, but don't have enough Japanese-language material to do so.
In general, Japanese culture tends to value safety and cautious use of technology, despite also being techno-optimistic and interested in machine consciousness. I think the tension between these facets could be resolved in favor of AI safety with relatively small effort compared to most Western countries.**4. Japan AISI: **They have a massive budget, yet struggle with enabling talented people to make use of it. As a result, nobody outside their industry partners can leverage the budget. The reasons are complex, so I wrote a separate section on this.
When asked why people's grant applications fall through or what prevents them from applying in the first place, we've heard a variety of reasons. The five most common, in order of prominence:
In past years, most successful research grants have only covered expenses for 6-12 months at a time. Some independent researchers have concluded that the lack of job security feels too harsh, when compared to working at traditional companies or well-paid startups.
This year, people close to me have submitted about 20-25 grant applications, and only tiny ones for event venue fees have come through.
As a result, people end up working as part-time volunteers, burning through personal financial runway, or leaving the field. None of which seem ideal for growing a vibrant and productive community.
Establishing an NPO for regranting could be an effective solution. Here's why:
The idea would be to get enough funding to pay the current would-be-employees of Japanese orgs in 2027, then build a buffer that can be used to guarantee job security and enable further growth.
We already have people doing the paperwork for setting all of this up. The next step would be to actually secure funding commitments to cover the funding gap ($2.1M). I think the most realistic pathways for this fall into two buckets:
The first approach is, to the best of my ability, addressed by this post. The second has been attempted without success; see the next section for more details.
A common question that comes up: Japan AISI has $56M of funding yet the other organizations have almost zero. What's up with that?
The answer, as usual, is that things are more complex than they seem.
More details on Japan AISI's organizational structure
Here's their organizational chart, first in relation to the cabinet office, and then within Japan AISI itself. The cabinet office is headed by the Prime Minister, and contains various agencies and ministries. The METI and Digital Agency collaborate on guiding IPA, whose mission is to broadly enhance Japan's IT capabilities.
For AI, in practice it means (among other things) reducing the friction of hyperscalers and frontier labs to do business and model training in Japan. Within Japan AISI, there are several teams and committees, each headed by a manager and consisting of 1-10 research / staff members. There is a lot of communication between Japan AISI's teams and other agencies. Each team is tasked with making ~annual reports that are published on the Japan AISI website.
More background and source materials:
J-AISI launched February 14, 2024, with its secretariat hosted inside IPA under METI and the Digital Agency. As of April 2026 it had 31 staff including concurrent appointments, a near-term target of 60, and a Prime Ministerial instruction to aim for a UK-scale institute of about 200; its funding rose from ¥3.8億 (~$2.4M) in the FY2024 supplementary budget to ¥88億 (~$56M) in the FY2025 supplementary budget. The December 2025 AI Basic Plan orders an immediate doubling of staff, and much of the secretariat is seconded from companies and ministries.
Published output includes evaluation and red-teaming guides, a data-quality guidebook, NIST AI RMF crosswalks, an open-source evaluation tool, and a cooperation memorandum with Anthropic; it has published no frontier-model evaluations, and it runs no grant programs. External safety R&D money flows through NEDO commissioned projects and GENIAC compute subsidies, which are built for companies and consortia rather than nonprofits. The wider legal frame: the AI Promotion Act passed May 28, 2025 with no penalty provisions.
Given all of this, it's no wonder Japan AISI's primary activities fall under:
While these are generally a nice thing for someone to work on, none of them budge my models of AI safety outcomes by much.[1] Having talked to people at Japan AISI, as well as people who have good connections with them, it seems like it would be infeasible for them to fund the non-profit arm of Japan's AI safety community. Not only is the overhead a sticking point, many of their employees are overwhelmed with busywork and don't have time to focus on the bigger picture.
As such, I am pessimistic about our chances of getting funding through them. However, it is an avenue we are actively investigating, and could improve a lot if both sides had people with the necessary bandwidth. In the meantime, our collaboration is mostly limited to running events together.
Convincing Japan AISI or the cabinet office would take lots of effort and 1-on-1 networking. Nobody I know wants to do it as a volunteer, but we have people who would work on this if they were being paid.
I made a detailed spreadsheet with the best known estimates for each organization's financials, number of potential employees, and funding gaps for 2026-2027. It was refined with multiple people who understand different areas of the community - big thank you to everyone who helped out! You can check it out here:
https://docs.google.com/spreadsheets/d/1EZBlPASzwtE3MvWeA9YxHTdCQkykk6LUANacDSLGBqs/edit?usp=sharing
In total, excluding Japan AISI, our estimates are:
If the $2.1M is not reached or we hit some unexpected costs, I'd cut from the two research grants first. This is because we could address Japan's unique opportunities more effectively with a portfolio of approaches, than with a larger research bet that deprives other organizations of impact. Our estimate for the bare minimum needed to resolve all operational bottlenecks is $426k. However, I expect that we would probably proceed with a different plan if only the minimum is reached, for three reasons:
The current plan seems to be most effective in the $1.5-3.2M funding range for 2027.
The average salary of a mid-career researcher in Japan is around 6.5M yen, or $40k per year. The average overhead of employment (taxes, insurances, accounting, etc.) is close to 50% on top of that, bringing the total cost of employment to $60k / person.
Browsing the jobs at 80,000 Hours and elsewhere, it seems like the current range for entry-level research positions is around $80-110k. Even for this lower-end range, the base salaries are 2.0-2.8x higher.
However, employer overhead is lower in the US, averaging at 30%. Thus the total cost of employment is $105-140k, or at least 1.8-2.3x higher than in Japan.
Beyond that, living costs in Japan tend to be much cheaper than in the US or UK, enabling a pleasant lifestyle despite the lower salary.
All in all, $2.1M doesn't seem like a big sum for enabling ~10 orgs to go from perpetual operational bottlenecks to doing their best work. I expect we could improve Japan's impact by at least an order of magnitude with it, while growing the field by 40 professionals or more.
If you're a funder or someone who could connect us with one, please share this post with relevant friends or drop me a line at esakoskin2@gmail.com. I'd be excited to resolve the situation and help put Japan on the map! The yen is cheap, and our people dedicated. Let's make the most out of it.
However, I do admit that they have done a good job at establishing Japan as a neutral party in the Asian and international AI safety governance space. It's plausible that some of their work could pave the way towards more robust US-China collaboration in the future.
By effective funding gap, I mean the gap between each org's current expected budget, and their ideal, most effective impact scenario for FY27.
In my model of organizational impact, marginal effectiveness improves during the first few employees, before it starts hitting diminishing returns. This is mainly because organizations have operational overhead that doesn't directly translate to impact, but which doesn't scale linearly based on the number of employees either. Larger teams can also specialize more.