AI Safety Hub Free Agency: Creating New AI Safety Hubs Around the World

By CalEA @ 2026-08-27T14:14 (+8)

Summary

I am developing a program which I call AI Safety Hub Free Agency, a program to seed new AI safety hubs by sending highly agentic and motivated individuals (Free Agents) with proven experience in organizational founding to cities that are lacking in AI safety communities, connecting them with local talent who act as president of the new organization, and giving them the proper funding to create and fund a new organization. Free Agents will run a playbook on how to create an AI safety hub based on those which have been created before and use their own expertise in organizational founding to catalyze the creation of a new AI safety hub. The Free Agent stays for 3-6 months and acts as a permanent mentor to the project after they leave. Free Agents can seed 3-4 AI safety hubs per year this way. Having more AI safety hubs increases the reach of the global AI safety community, provides space to run events and courses and inspires local talent to take on roles in AI safety and start their own AI safety projects.

Introduction

AI safety hubs are a valuable resource for the development of safe AI. Places like the London Institute for Safe AI (LISA) (where I wrote this from) and Constellation, provide a space to share ideas, do research and receive mentorship to help grow the field and bring projects to life. They often serve as the first point of contact for those who may be interested in AI safety but have no context or direction in how to get started. I am not the first person to say that we need more AI safety hubs. I want to help create more AI safety hubs.

My model for doing so is the Lebron James model. When Lebron James left Cleveland in 2010 to take his talents to South Beach, he went on to the NBA finals four years in a row and won two championships for Miami. When he returned to Cleveland a few years later, he brought them their first championship. After that he went to the Los Angeles Lakers and won yet another championship. Even though he is the GOAT, he didn’t stay in one place but rather spread his talent around. He was able to fit into different systems in different cities and bring immediate success.

In AI safety, highly capable individuals like this exist as well. These people might have already founded their own local AI safety hub where they’re from, be the president of some AI safety organization or have had tangible success in other ways. Similarly, there are cities that might have all of the foundations necessary to have an AI safety hub, but just need their Lebron James to come and tip the scales to turn an idea into a reality. I call these individuals AI Safety Free Agents. Free Agents are highly agentic, motivated and willing to go abroad and serve as the necessary catalyst to create an AI safety hub.

The Plan in a Nutshell: Send highly skilled, proven founders to a city where there is no AI safety hub, give them a budget, a playbook on founding a safety hub, and 3-6 months and connect them with a local contact who speaks the local language and has insights into the culture and lay of the land and together they found an AI safety hub for their community. The local is the owner of the project while the Free Agent serves as mentor and does the heavy lifting. Free Agents give permanent ongoing support for the project even after they’ve left the city.

If we want more AI safety hubs, we have to be proactive about it. It’s nice to write about it and try to rally locals to do it themselves, but if we want to make changes in a world of short timelines, we need to be aggressive and jumpstart AI safety hubs wherever we think they might have an impact. 

Currently, no such project or organization like this exists, as far as I’m aware. There are a number of initiatives to help seed AI safety communities, but they typically rely on the potential founders making the effort themselves and seeking out funding and mentorship. Kairos Pathfinder, for example, helps to get AI safety university groups started. Their project mostly focuses on American universities and founders have to make the initiative themselves by applying for the program.

The Free Agency program would serve as a more proactive approach. Rather than waiting for people to hear about projects to help them start their own organization, we would scope out potentially high value targets and drop the tools and infrastructure directly to them. This would allow us to target highly valuable and impactful cities and countries rather than relying on agentic individuals to make themselves known. Creating an AI safety community in a country where there is none seems potentially more valuable than creating yet another university group in the American Northeast in which dozens of organizations already exist. 

There is already a proven model of success. The founder of the Singapore AI Safety Hub (SASH) is not from Singapore, but was able to find a local contact and create an AI safety hub there. The Lisbon AI Safety Hub (LAISH) was founded by a non-local (me) who noticed a gap in AI Safety in Portugal and wanted to fill it. I found local people who have helped me make it a reality. My own impression from creating LAISH is that there was already a high amount of latency but the country lacked the foundations. Once I had created it, I received a quite significant amount of interest and people who wanted to get involved. These people were already there, they just needed somewhere to go.  There are other cities where it should be possible to copy this model. 

With this model, it’s plausible that we could create more AI safety hubs in the next year in key underserved places. If there is enough interest and individuals willing to go and create a hub, this could grow to double digits. I hypothesize that it would take a Free Agent three to six months to found an AI safety hub to the point where it is self-sustaining and they no longer have to be physically present on the ground. 

Note: It’s possible that there are only 4-6 highly valuable cities in which a new safety hub makes sense. If we create this many new hubs, I would still consider this a success. A few hubs can make a lot of impact and grow to become a central location for regional talent.

Note 2: The timelines are rough and some places will need far less help and time from the Free Agent while others will need longer term on-the-ground support. It depends a lot on the local AI safety environment and the local talent we can find.

Since there are already a lot of existing communities, the roadmap to creating an AI safety hub is already there. Those who have founded AI safety hubs before should create a unified playbook with all of the steps to creating a hub including setting up a website and email, branding, sourcing venues, coursework and what types of events to run. This would significantly reduce the friction in starting a hub and accelerate the process so that it doesn’t get stuck in development hell. There is a SPAR project to improve resources and playbooks for university groups. There are some aspects that are particular to university groups, but a lot of the lessons should be broadly applicable to non-university groups as well. Other individuals I have talked to are working on resource packages for founding new AI safety communities, but no formalized group is working on it that I know of. Facilitating coordination on this could be meaningful.

Lens Academy also develops course material based on the AI Safety Atlas (as well as other courses) which I will be using for LAISH’s first fellowship and can be easily adapted to different safety hub fellowships. This is a significant part of the playbook and takes away the need to recreate coursework for every new hub which is inefficient and not a good use of time when the material already exists. 

The playbook is something I am already working on but if you have founded your own AI safety hub and want to help contribute, please reach out to me on LinkedIn: https://www.linkedin.com/in/calebeliprice/. I respond to every message and connection request. I also plan to serve as an on-call guide to those who go to start an AI safety hub and think that this would be a useful role for others who have founded or are running AI safety hubs.  

Finding a local contact is a critically important step. BlueDot has a spreadsheet with every person who has ever said they are in a particular city and wants to work on AI safety. This is a useful resource for sourcing local talent. It’s essential that the local be capable and willing to take on the project. In order to keep them working on it full time, they can receive a competitive local salary for as long as they work on the project. The local should be the owner of the project from the beginning. The Free Agent just serves as the mentor and implements the playbook, doing the heavy lifting when it comes to setting up the organization and mapping out the steps. This way, when the Free Agent leaves, the local is the president and takes full control without there ever needing to be a hand-off. The local gives legitimacy to the organization and keeps it from feeling like something being imposed on the local community. 

Highly agentic locals may eventually be poached by organizations or want to move to London or SF if the opportunity presents itself, but hubs are meant to be living organisms and it’s not impossible to find a successor to take over in this situation, as has been the case in other safety hubs.

The Free Agent will act as a long-term mentor, checking in weekly or bi-weekly with the president of the new AI safety hub to make sure that everything is going smoothly and fixing issues from abroad as they go. This way, the Free Agent can move on to another place and start another AI safety hub. 

Note: All of the above is also applicable to places where an AI safety hub already exists, but it is struggling or disorganized. Free Agents can also go to places like this and help right the ship and get them back running at full capacity. An example of this is Japan where an AI safety community exists but hasn’t materialized into a continuous and meaningful hub. This might even be easier as the groundwork is already done. In this case, the people running the already existing but struggling AI safety hub would need to welcome outside help, which is not guaranteed. Cities where there are no AI safety hubs or where the existing hub welcomes help should be prioritized to maintain friendly relations and eliminate the feeling of competition.

Candidate Cities

I have used the AI Safety communities map to look for places that are missing AI safety hubs. All of Africa has only seven such communities, and only three of these even have their own website. The Netherlands alone has more communities than all of Africa combined. The Arabic speaking world only has two communities. South Korea, which is one of the most critical parts of the AI development pipeline, has no representation.

This is a list of the potential criteria to assess whether a city is the right fit, which came up when discussing this with colleagues:

• Country or city importance, including “middle power” relevance.
• Quality and scale of local technical talent.
• Existing AI or EA community.
• Current hub quality or absence.
• Ability to find a strong local owner.
• Tractability of working with existing organizations.
• Potential links to government and policy networks.
• Cost of living and salary in the local area.
• Population size.

In addition to the above, it may be optimal to focus on EU countries that lack an AI safety community and fill in the gaps until all 27 members have their own AI safety groups. This is for a number of reasons:

• Being a member of the EU gives member states an outsized impact on global geopolitics. Every member state has local politicians who are members of the EU Parliament, where they can propose ideas and share the concerns of their constituents in one of the most important political forums in the world. Influence from local grassroots organizations can therefore spread, even from very small EU member states like Malta or Luxembourg.
• The EU has already served as one of the initiators of political change in AI safety, with the EU AI Act, and has a governing body which influences AI safety regulations in the EU AI Office. Laws and regulations from the EU tend to have a multiplicative effect outside the EU due to the “Brussels Effect”.
• After the US and the UK, Europe has the most robust network of EA and AI safety organizations and people. Existing organizations like Safe AI Netherlands and Safe AI Germany could more easily lend their talent and resources to neighboring countries.
• Free Agents from the EU may have more ability and interest in moving to an EU country to help start a hub due to visa constraints (ability to stay in a Schengen country indefinitely, which is not the case in other countries), cultural similarities, and proximity to their home country. 

The EU countries that currently lack AI safety communities are the following: Cyprus, Estonia, Greece, Lithuania, Luxembourg, Malta and Romania. Of these candidates, Estonia and Romania seem the most likely targets, due to the above criteria.

Potential candidates below. If AI safety hubs already exist in these locations that I don’t know about, please let me know. I would also suggest adding it to the AI Safety communities map. If you are already in one of these locations and want to start an AI safety hub, also reach out and I can try to help and advise. These are presented in no particular order. 

Seoul

• Metro Population: 26 Million
• The de facto capital of everything that happens in South Korea
• South Korea plays an instrumental role in the AI supply chain with SK Hynix and Samsung both being big players
• Strong ally of the West with geographic proximity to China
• Strong tech sector and highly educated public

Kyiv

• Metro Population: 3.5 Million
• Currently involved in the Russo-Ukrainian War, in which AI has played a massive role. Many people consider the war as the testing ground for new AI military technologies. This is a relatively overlooked aspect of AI safety that could have hugely catastrophic results. Many people in Ukraine are working in AI but relatively few are focused on the safety aspects, especially the EA related outcomes AI military proliferation
• Eastern Europe is largely unrepresented in AI safety, having only one each in Latvia and Bulgaria

Jakarta

• Capital of the fourth largest country in the world
• An AI safety community already exists there, but exists only as a WhatsApp group. The individuals in the group could be leveraged to start a much more significant group which does more than just chat.
• Indonesia is a member of both ASEAN and BRICS. The current AI safety landscape is relegated mostly to the “Global North”, meaning that “Global South” countries have little voice in the direction of AI safety. China tends to ally more with “Global South” countries meaning that influence in this realm could have a significant impact.

Other Candidates (some more likely than others):

• Cairo: Large population, center of the Arab world, Arab speaking world has extremely little AI safety, except for Morocco. BRICS member. Egypt has very short visa stays for foreigners though (15 days unless you’re from the Arab Gulf)
• Addis Ababa: Very large population, Africa is extremely underrepresented in AI safety, low costs. BRICS member. 30 day tourist visas can be extended by up to 90 days.
• Kigali: Rwanda is a tech hub in Africa and a desirable destination, could partner with AI Safety East Africa, based in Kenya. Relatively short visa-free stays (30 days) except for certain citizens.
• Lima: Large population, can be part of the growing Latin American AI Safety network. Visitors can stay between 90-183 days visa-free.
• Tbilisi: Large expat community, easy to live there as a foreigner with relatively little restrictions to reside there for a year, foreigners can stay for up to a full year without any visa which is plenty of time to start a hub. Tbilisi is relatively small compared to the other options.
• Other Latin American countries like Panama, Costa Rica, Guatemala and Bolivia could be targeted by existing Spanish speaking communities to help “fill in the gaps” as proposed above with the EU. Organizations that already exist in Argentina, Mexico, Colombia and Chile could help to facilitate the creation of new Spanish-speaking hubs.

I am open to ideas about where an AI safety hub is most needed and where starting one might be most possible. I am not connected with every local community or aware of all local regulations and laws regarding starting organizations, so recommendations are subject to change.

Successes and Failures

Here is a list of some criticisms I have received:

• Why not just focus on cities that are essential to AI development already? Are middle powers relevant?
  • I strongly believe that middle powers are relevant and I am not the only one. Even if these countries don’t play a huge role in AI development or policy, the talent still exists there. If we don’t have local representation, it’s possible that we are missing out on a huge amount of talent. Many people aren’t able to enter AI safety at all because they lack the economic power to move or a strong enough passport to reside long term in the current hubs of AI safety (US, UK and EU). That doesn’t mean they don’t have something meaningful to contribute.
• Shouldn’t locals just start their own communities? Why send someone else?
  • In some places, this is already the case. Highly motivated individuals exist all around the world and some of them have already started their own AI safety hubs. But the gaps on the map show that this isn’t the case everywhere. It’s really difficult to make a broad call to action. But when people are given a task and pushed to do it, they can do it. They just need that starting push.
• Where are you going to find these highly agentic people to move abroad for six months?
  • This is a legitimate concern and one of the biggest bottlenecks. I can be one of those people but I want to find others as well. I have some interest from people I’ve met in the community already. If you are a person who matches my description or know someone who would be willing to be a Free Agent, please let me know. People are willing to uproot their lives to move to London or SF already, so why not Seoul or Bucharest if the theory of change is strong enough?
• What if the safety hubs you create just fail?
  • This is possible. The local appetite might not be there or the organization may just peter out over time. We won’t know if that’s the case until we try and even if only half survive, I feel that the venture will have been worth it.
• 6 months is too long. Free Agents might not be willing to relocate for that long.
  • A legitimate criticism and worth considering. I think 6 months is around the optimal time to create a safety hub and stay there to make sure it’s running smoothly. 3 months is potentially enough as long as the Free Agent is in regular communication with the local and they found the right person to manage the hub after they leave. Others have suggested as little as a few weeks as necessary, but this feels too short to me but may be possible in some places.

Main Risks:

• Not finding Free Agents or local contacts
• The safety hub we create does not make a meaningful impact
• The new safety hub drifts in value alignment once they become disconnected from support from the main core of AI safety
• Local AI safety hubs turn out to not be a strong theory of change and generally don’t make much of a difference in field-building

Action Plan:

• Finish the AI Safety Hub Playbook so that it can be handed off to the Free Agent and implemented wherever they go
• Finalize cities for the pilots
• Find Free Agents who are willing to relocate for a period of 3-6 months
• Send the first Free Agents to cities that are most likely to succeed
• Create and run AI safety hubs around the world

Metrics of Success:

• Creation of at least four new AI safety hubs in one year
• The AI safety hubs function independently with continued guidance from the Free Agent
• The AI safety hubs run fellowships that have at least 10 graduates and local community events that are attended by at least 10 people regularly
• At least 5 people from each new AI safety hub city have been accepted into further fellowships or AI safety positions thanks to connections and training from their new local AI safety hub

If you want to be involved in the project, are a potential Free Agent or know a potential Free Agent, reach out to me on LinkedIn or via email: calebeliprice@gmail.com

Special thanks: My colleagues at Gen Stream 1 including Yaqoob Ahmad, Julien Sireau, Zuza Topolska. Elias Schlie and Mark Kagach from Lens Academy. Sam Smith. Oliver Kurilov at LISA.