At the beginning of May, I taught at an ML4Good bootcamp and gave a session on starting and running AI Safety initiatives. I wanted to write about this too for a while, but this session finally motivated me to do so. Over the past year I have started AI Safety Saarland, helped launch Safe AI Germany, and mentored groups around the world through Pathfinder and independently. These are 10 lessons – not necessarily in a specific order, and not particularly polished – from my past year of field-building that have already been valuable to my mentees and that I hope can be valuable for many more organisers.
When I started AI Safety Saarland, my greatest blocker was that I did not know anyone around who was knowledgable about AI Safety and could help me with the initiative. I put out posters on campus saying "Co-founder wanted" to see if anyone would pick it up. Within two weeks, I had 20 emails in my inbox from people interested in co-founding. While this clearly showed interest, I realised that none of them really knew what AI Safety was about. I decided to nevertheless try to put a team together with some of them as not all tasks require a deep understanding of AI Safety and I could try to upskill them on the go. My main reasoning behind this was also that if I gave people responsibility in the group while they take an intro course, it may increase their chances to stick with it because they already see how they can contribute, which closes the learning-to-action gap. So I organised a crash-course teaching them the essentials of AI Safety in two afternoons and motivated three of them to take the AI Safety Collab course, while keeping all strategic decisions on my end to ensure the group was focussed on AI Safety and not just trustworthy AI or AI ethics.
How did it go? Well, I think reasonably well. Two of the people who took the AI Safety Collab course organised the intro fellowship straight away, got into Pathfinder in the Spring semester and now lead the initiative. But this strategy does not necessarily work for all people and I would not say that the core team of AI Safety Saarland is as strong as it could be if all organisers were deep into AI Safety. To make it work, it requires a lot of dedication to upskill and invest in the organisers. But even if they are the only ones going into AI Safety later on, this was probably worth it.
On another point, I think that whether to go with such a strategy or not depends heavily on your goals and ambitions for the group. If you want to scale quickly, offer many different types of activities and events, you will need more people to just handle it all on the logistics side and then it might not matter as much how deep your volunteers are into AI Safety for many tasks. If you just want to create a very focussed community you will likely want to adapt a different strategy. But even then, I think it is very important to ensure organisers keep upskilling – they are likely the most valuable people in your community.
Many parts of the strategy of your initiative depend heavily on what your goal is and from experience I think not enough people think about this clearly in the beginning or make this explicit. For me, I wanted to start the dialog on campus about AI Safety, get faculty and many students involved and create a pipeline for people to enter the field. For other groups I have advised, it was much more that the organisers wanted to create a tight community for themselves to read and discuss topics in AI Safety. Both goals and everything in between are very legitimate, but it helps a lot to be clear about it because how you do marketing, what you offer and what work you as organiser are doing very much depend on it.
What your goal for the group is also depends a lot on what your motivation and goal as an organiser is. If you want to upskill yourself content-wise without much organising overhead, creating a reading and discussion group in which you participate yourself might be the better strategy. If you want to build pipelines and are more about getting other people into the field, then you will need to work much more strategically about who you want to reach, how you reach them and what you offer them. This might then look like organising an intro course, where you will likely be the facilitator and might not upskill yourself a lot in terms of content (in terms of many other skills you will!).
So I highly recommend sitting down for a moment before you start planning and write your motivation and goal behind the group out. This will make it much easier to plan, market and run your group.
Before I started, I had many people suggesting that I start a local initiative. I did not really feel like I wanted to do it for a long time because I mostly felt that I would not enjoy it much. It turned out that I was pretty wrong about it and in fact I pivoted from wanting to become a technical AI Safety researcher to becoming an AI Safety generalist. But I think this is actually a very important point even if it sounds somewhat egoistic: if you start or run a fieldbuilding initiative, please don't see only how this helps others, but also keep in mind how this helps you.
For me starting AISS was also a (somewhat) cheap test to explore and test my fit for a different kind of career path in AI Safety and it became something that gave me rich experience and led to many more opportunities. Organising without seeing where the benefit for you personally is, is dangerous for your motivation and capacity. Running a local initiative gives you career capital and you will also want to move on at some point. If you are strategic about what you want to test out and gain experience in while organising, it will not only be much more fun for you, but might also have a much higher downstream impact. As selfish as it may sound, even if AISS died tomorrow (which I very much hope it does not!) and nobody there picked up AI Safety (which also does not at all seem to be the case because I do care a lot about bringing people into the field), organising AISS was likely one of the counterfactually best things I did for my own career and downstream impact (at least with SAIGE, but hopefully also at the Generator Residency). So please don't forget about yourself and use your local initiative as a playground to run cheap tests on things you'd love to organise and learn about!
Leveraging what people at your place are interested in can increase your reach by a lot as I found in the three examples below:
Even though you might imagine your group in a certain way, I think it is very worth taking a step back every now and then to reflect on which specific interest groups at your place you can leverage to work on some subfield of AI Safety. You might not feel confident enough about some of those fields (like I did about AI security), but you can very likely find people in the broader community that know about this and can give you the information to get started. For example, when I went to EAG SF in February, I reached out to some AI security people and asked if they had a reading list that could be used for the subchapter I was starting. You don't need to be an expert in everything before you start an initiative on it. This brings me to the next point:
When I started the group, there was a lot of work to do before we went official. For example, I wanted to organise this research fellowship and for that had to reach out to potential mentors. Why would some researcher or even professor agree to be a mentor for an initiative that did not yet exist and was organised by a random undergrad student? I still don't know how I managed to get 14 mentors, but I think it comes a lot down to how you represent yourself. If you give the impression to others that this is serious, you know what you are doing (even if you don't) and act professionally throughout, you can build trust.
For me it helped a lot to have the domain, website and email address for the group. I never told people that I was an undergrad, but that I was starting a serious org to bring AI Safety on campus. It feels very uncomfortable to work like this, but as with life in general, if you do not build the confidence in yourself that you can manage this, how should other people trust you?
Something else I noticed that likely helped me a lot, was that I focused much more on maximising success and not minimising failure. Organising a local group is pretty low stakes and even if the group fails, you probably learned something for yourself. If you take that as the baseline success, it can only get better! Hence, even if it feels in the beginning like you are faking a lot, this is necessary to eventually make it. Small things like telling people that I am organising this fellowship or putting up a professional website very much boosted my confidence in the group and made it a reality.
I mentioned it already in the points above, but I think experimentation is highly under-appreciated by many organisers starting out. Yes, there are certain types of events and activities that have proven to work well for many local AI Safety groups around the world. But this does not mean that you always have to do everything the same way. University AI Safety groups are a fantastic playground to try out different formats and ideas, but also every university is different and not everything works equally well at all places.
It helped me a lot to see all activities we organised and all the processes around them as experiments and opportunities to find out how I can make things even better.
I will come to marketing in another point below in more detail, but as an example: when you are starting out, you will very likely not have much information on what marketing strategies for your events will work well at your place. You may have some information from other groups, but each place is still somewhat different. You will very likely not have the optimal marketing campaign at your very first event, so instead of trying to find this out beforehand, it is much easier to just try out many different strategies and evaluate on the go and afterwards how it went.
For AISS' Kick-off, we had posters everywhere on campus, flyers on cafeteria tables, LinkedIn ads, Instagram posts, repeated messages in large student WhatsApp groups, people going to lectures to pitch our event there and tabled in the cafeteria during lunch time with chocolate for people who signed up. We had a question on the luma sign-up asking where people heard about the event. As registrations came in, we realised that chocolate and LinkedIn did not bring many sign-ups, while posters, flyers and WhatsApp messages were bringing many more. So we doubled down on those channels for this and future events, bringing us consistently high numbers of attendees. Experimentation is key to figure this out, but you also need to set things up in a way that allows you to reflect on them. Without the extra question in the sign-up, we would not have known what worked best.
The same is true for event types. Just because there is no other group that has tried out a certain format, it does not mean that you can't do it. My research incubator was an experiment, and so were the various public events we organised. For example, we tried to run one social event per month, but for whatever reason, barely anyone came no matter how much food we offered. I still don't understand why people were not interested in socials, but well, we went on organising more talks with pizza afterwards and by doing so reached many more people. If something does not work, don't be discouraged – experiments can fail, but are a datapoint more that you can use in future planning.
I can't emphasise marketing enough and yet, this is likely the most common failure mode I observed when organising myself or mentoring others. If you don't tell people that you are organising a certain event or have applications open for your fellowship, how should they know about it? This seems very trivial, but for some reason it isn't for many groups. I might dedicate a separate post to this topic in the future, but here are the main three lessons to do marketing for your group better:
I really think, with just a little experimentation and effort in marketing, you can increase your reach a lot which makes all your downstream activities more fun to run. To stress-test whether your group is visible enough (which sometimes is easy to assume when you are spending a lot of time with it), ask your seat neighbour (assuming it's not your co-organiser or friend), whether they have heard about the AI Safety initiative on campus. If they haven't, you can do more!
I started AISS partially because I wanted to have a local community of AI Safety people after meeting so many amazing AI Safety people at EAG(x)s, ML4Good, GCP etc.. Now, when I created this community and people got interested in AI Safety via this community, they may have found the local community unconsciously sufficient and weren't necessarily seeking as many external connections. Similarly, for the organiser team, as long as I was there, external relations were covered by me and hence, it might have seemed less necessary for other organisers to grow their network outside the group.
But for a group to have an impact or to sustain itself when you leave, you will need people to connect and build a network outside your local community. I found this particularly challenging as in the beginning there were not many other groups in Germany that connected and at least coordinated together, but also because many of my co-organisers or fellows needed a visa to go beyond Germany which makes EAG(x)s etc. challenging to attend. Motivating people to take BlueDot courses to meet more people can be one step, but actively connecting and introducing people can also be very helpful. National coordination, especially when based in Europe, appears to work wonders too and can even help to start many more groups. SAIGE is bringing together German AI Safety group organisers for weekends and since we launched, several new chapters have started. Safe AI Netherlands is even more a network of local groups that now work together to streamline processes and help each other out. I would be very excited to see more collaborations between local AI Safety initiatives like those. I think it often motivates organisers a lot if they feel part of something bigger and have a network of support beyond their city.
I knew from the beginning of AISS that I would only be able to run the group one, at most two semesters. Nevertheless, in the beginning and especially because many of my co-organisers did not have much AI Safety background yet, I kept all decisions to myself. I had a rather precise vision of how I wanted posters, events and processes designed. But as we scaled quickly and it became clear that I would leave the city only four months later, I realised that the only way to keep the group alive after I left was to radically delegate to other people even if that meant they were doing things slightly differently from what I imagined. But delegation is also a very valuable skill to learn and can empower your co-organisers a lot. It is also an investment in them, a sign of trust, and if you approach it that way, you can still steer major decisions and their development into AI Safety in the way you want. So yes, delegation can feel scary at first, but may ultimately make your group much stronger and yourself more replaceable.
With everything I wrote above, I want to encourage you to be ambitious. So many times over the past year, I had doubts that I could do certain things, but then often realised that really the only limiting factor was my own belief that it wasn't possible. AI Safety very much needs people that approach capacity building with an agentic and ambitious mindset. There might not be time to first do a pilot and then grow a little bit each semester. We need ambitious efforts now and there are many people in the community that will very likely be happy to support you if you want to start something. You can really just do things and by this make a difference!