Germany's Got Talent. How do we get it to work on AI safety?

By Melanie @ 2026-09-12T15:06 (+2)

Summary

I interviewed 14 people working full-time in AI safety who came through the German ecosystem to find out how they got in and what would help others to do the same. 

A couple of patterns emerged:

Epistemic status 

I spent 8 weeks interviewing 14 people working full-time in AI safety, about how they got there and what it would take to get more outstanding people living in Germany working on it too. Interviewees worked in different areas and at every seniority level. They have either lived here for a long time, were born here, or moved to Germany for a degree.

Interview questions differed since a technical researcher probably can't usefully answer questions about hiring or talent-pool strategy. Alongside the interviews, I read other EA, LessWrong and 80,000 Hours posts on talent pipelines, talked to fieldbuilders at a retreat, and spoke with hiring managers from career transition programs and research fellowships.

What follows are patterns that often came up unprompted and went uncontradicted by anyone else I spoke with. To protect interviewees' anonymity, I won't say more about who they are.

I'm treating the patterns below as hypotheses about a young field, observed through a German sample, not as settled facts about Germany specifically. 

Methodology

Germany has the largest AI workforce in the EU, roughly double the next largest, and ranks fifth globally in absolute numbers. It is not, however, a hub for AI safety orgs and, to my knowledge, has no national fellowship or incubator, besides SAIGE. Despite this, people who came through the German ecosystem are contributing at the highest level right now. So I set out to learn what we can take from their paths to help others repeat what worked and avoid what didn't.

During my interviews, I got pushback on this method, because those paths happened at least two or three years ago and may not be reproducible today. One interviewee entered when the field had been small for roughly a decade. She grew into her role because her organisation had more work than people, and she got noticed because she had volunteered within the community before. She noted that if the number of applicants was as large as it is now, she wouldn’t have been able to have her career.

My goal isn't to give instructions on how to reproduce those paths exactly, but to gather ideas for better on-ramps. I think the people I interviewed are well placed to provide those, because they know what actually moved them forward, and several of them now decide about who gets hired.

What do the people I interviewed have in common?

The EA Funnel

Twelve of my fourteen interviewees entered the AI safety space through EA channels: the 80,000 Hours career guide, an EA intro course or reading group, an EAGx conference, or a local group. Those were often the first contact, and my interviewees then went on to do ML4Good, ERA, GCP or a Talos fellowship.

Since my sample came mostly from a list of BlueDot alumni, and BlueDot is itself an EA-adjacent on-ramp, this number is not surprising, and it shouldn't be read as "most people that work on AI safety are EAs". That is not what I captured. What I think it does tell us is that this on-ramp works.

One interviewee told me that the AI safety community largely hires people they know and trust. Building trust means understanding the context, which this post describes as knowing the organisations, concepts, people, and culture. Much of that is picked up at EAGx events, or by reading and writing on the EA Forum.

Demonstrated agency

At least seven of the fourteen have a clear pattern of self-directed work preceding paid work. Rather than applying for jobs, they approached people or organisations directly. They volunteered until paid work followed, offered to build something an organisation needed, started a local initiative, or organised an event.

One of them made an early bet on a research direction most people weren't yet taking seriously, then made a point of getting to know the person most associated with it. When a mentorship opportunity came up much later, he was the one who got remembered. As he put it: "From my CV, it looks like this happened by chance. But there was a lot of work before I even started."

Why does demonstrating that you can do something without being told matter so much in this field? 

A few reasons: it demonstrates real interest, since people don't build or do something unpaid unless they actually care about the outcome; it demonstrates alignment, since what someone chooses to work on reveals whether their priorities match the field's actual goals; it demonstrates skill, since the result, however small, is real evidence of competence; and it demonstrates that you won't need much managing to begin with.

One interviewee estimates that roughly 2,000 to 3,000 full-time AI safety roles currently exist globally, against many thousands of aspirants, with many fellowships accepting below 3%. In an established field, that filtering is done by credentials. In AI safety, only very few people have a ten-year track record, so the strongest signal is a self-driven project that sets the candidate apart.

Who gets noticed, and who doesn't 

Agency only leads to opportunities if it is visible to the AI safety or EA community. With one interviewee estimating that 50–80% of roles are filled through personal networks, many talented individuals are overlooked simply because their work never reaches those already in the field.

Some backgrounds seem to be really common among EAs. One interviewee ran an internal analysis at EA Germany and found that around 20% of the people in his sample who ended up in impactful careers had come through Studienstiftung, against roughly 0.5% of German students who receive it. He also named maths olympiad participants and Germans who studied at Oxford/Cambridge or at an Ivy League school as common backgrounds among EAs. 

So someone who studied at a German university without a student group and never encountered EA might also miss the AI safety community completely, even though they might be able to contribute.

What nearly made interviewees quit

Most people who try to get a full-time position in AI safety have quit, or not succeeded yet. To find out what we can do to help people in Germany, I asked every interviewee what they were missing when they got started. 

The most common struggles were rejections from fellowships, jobs or grants, and not knowing what they could have done differently or should do next (this seems to be improving, and fellowships are adding more resources and next steps to their rejection emails).

At that point, the usual advice is to build more context, skills, and to find an opportunity to contribute agentically. I think that advice is much harder to get right than it might sound. 

Knowing which opportunity is worth your time

I noticed this during my time at SAIGE: I would spend hours researching opportunities and applying to projects, then talk to my mentor or another senior person, and be told I was looking at the wrong thing or would be much better off doing X instead. So a 10-minute conversation saved me months of doing something that probably wouldn't be the best use of my time. People without that access are doing the same work without the correction. 

Networking in Germany

Finding a mentor, or opportunities, works best in person. 10 of my interviewees named gaps such as missing coworking spaces, retreats, and the high-energy formats that exist in the US and UK, but not in Germany. Without them, the field here is hard to scale. One interviewee said many people would rather work in Europe and don't, purely for the lack of an encouraging environment.

Building context means leaving Germany

Gaining context currently means going somewhere else. Almost every programme or event my interviewees named happens outside of Germany, and often requires travelling repeatedly or for extended periods of time. 

Who this filters for

This filters for people who can afford unpaid time off and have no caring responsibilities, which is part of why the field looks the way it does. It skews heavily male, particularly on the technical side, where one interviewee estimated that some fellowships and the organisations running them are around 95% male.

Contributing from Germany

Most of my interviewees have left Germany to work on AI safety full-time. This section is about what the ones who stayed are doing, or did. Several of these paths were taken by people who have since moved away.

Working remotely

One interviewee's read of the job boards was that around 80% of roles are in person, and mostly in the Bay Area or London. That makes remote roles competitive, and also the main way to get a full-time role while living in Germany. 

Another interviewee shared his experience working remotely. He was offered a contractor role because his employer was “scared of German employment law”. He then spent about a year at one organisation on rolling three-month contracts, and eventually left Munich partly because "nobody rents a flat to someone on short-term contracts". 

This is one account from someone who now has a full-time position, and it should not be a reason to rule out this path. 

Founding 

One interviewee told me that she went to an EAGx conference, did 23 one-on-one meetings in a weekend,  and people shared their successful funding proposals, their own strategies, and their failure modes with her. 

She then founded an organisation and ran it on her own savings for months. Her first funding application was rejected, which she called entirely understandable, because there was no track record yet.

Another thing that came up independently in three interviewees was that there is no shared sense of what Germany is trying to build, or what its impact can be. I suspect that makes it harder to decide what to found here in the first place.

Germany's existing institutions

One interviewee made the point that the German government already has institutions working on things close to AI safety, and that he would like to see far more people working there. Since I spoke to him, Germany has founded its AISI, which will presumably absorb some talent.

A second interviewee reached a similar conclusion: rather than competing for one of the very few seats at the best-known orgs or think tanks, he took a role at an organisation with real industry access. Before he and his colleagues arrived, it had almost no connection to the AGI-serious policy world. It does now.

Recommendations

I think the EA channels demonstrably work. Alongside them, we should tap into the pools we are not reaching yet, and build on-ramps for the people who are already asking versions of this community's questions, but won’t find EA by themselves.

Who we are not reaching

Germany has around 117,000 AI professionals and 2,469 AI startups, counting everyone from AI-literate people in adjacent roles through to researchers building deep learning systems. Not all of them could or should work on AI safety. But the gap between people working “in AI” and the people working on making it safe is hard to miss. 

Interviewees described trying to reach people by showing them articles, which led them to conclude it was all exaggerated; students who decided the frontier labs just have good marketing, and one person who said people around him think he is one of the crazy ones. One interviewee's account of the field's own norms and narratives was that it quickly becomes a bit cult-like, and she does not know how long the field can sustain that if it wants to grow. 

I think to solve this, we should prepare some arguments beyond short timelines and doom. One interviewee described her motivation as the appeal of a field young enough that you can reach the frontier quickly and become one of the few people in the world who contribute to a narrow piece of it. 

I think the same is true for other groups worried about related questions, for example around data centres, sovereignty or surveillance. Their concerns are not identical to this community's, but they overlap. 

What we can provide

BlueDot's courses are where we send career transitioners and people showing early interest. They finish, and often cannot get into a fellowship right away, because they have no relevant, hands-on experience. The advice at that point is to do a self-directed project with short feedback loops, which is much harder to get right without a mentor or a community.

What I think would help is a lightweight accountability structure for those kinds of projects, with someone who checks in on the progress, gives feedback early enough so nobody spends months on something irrelevant, and reassures them when their efforts are on track. 

If you think this is a good idea, or if it already exists somewhere, let me know.

How to start (simplified)

Get good, and get known. Work out which skills actually matter by reading job postings for roles you are not yet qualified for. Then try to gain those skills through courses or smaller projects with tight feedback loops. 

Work in public, where hiring managers can see you.

Look wider than the ten aligned organisations; for example, at industry associations, ministries, and the federal institutes.

Objections

Isn't this whole piece biased by survivors? 

Yes, I interviewed people who made it in, so I can’t confirm that what they have in common accounts for successful entry. 

Building agency, context, and skill is, however,  what hiring managers have told me they are looking for in applicants.

Wouldn't reaching out to untapped talent pools dilute the quality of applicants? 

I’m not proposing mass outreach or encouraging people to randomly apply to fellowships. I’m suggesting building narratives for specific talent pools to get the best applicants possible from each talent pool.

Final thoughts

Germany already has outstanding people. Most of them are not working on this yet.

If you want to change this, consider field-building as your career. You might find the case for AI safety capacity-building work, AI Safety's Biggest Talent Gap Isn't Researchers. It's Generalists., a reading list for generalists, and 80,000 Hours' AI safety field-building career review interesting.

Also, if you think I got things wrong, I'd like to hear from you.

 

Thanks to Tilman Räuker for your patience while mentoring me through the process, to the SAIGE team for assigning me this project, and to everybody who took the time to answer my questions.

Crossposted on LessWrong.