A tough lesson I learned over the past few years is that it’s pretty stressful to work in the “Please Don’t Build the Torment Nexus” ecosystem, especially as it increasingly seems like the world is hell-bent on building the Torment Nexus.
It’s hard to be productive when you’re stressed and anxious. And setting aside productivity, it’s also just plain bad if people working on AI safety — My peers! My people! — are feeling distraught.
Anxiety about AI risk really fucking sucks. If you feel this way, you aren’t alone. I’m right by your side, biting my nails, whitening my knuckles.
But even besides me, a lot of people seem to feel this way.
How big of a problem is this?
Survey data1 described in Spencer Greenberg’s recent 80k episode implies that mental health challenges for people working on AI safety are fairly common:
Most respondents experienced psychological challenges — the median person reported six different problems.
More than half (52%) said these challenges at least moderately hurt their effectiveness, while 45% had considered cutting back or leaving.
The most common problems were:
Difficulty switching off from work: 68%
Imposter syndrome: 60%
Burnout: 59%
Guilt: 56%
Anxiety or worry: 54%
Less common — though still significant — were loneliness or isolation (32%), hopelessness or depression (29%), interpersonal difficulties (28%), and fear of death (22%).
Few appeared to be experiencing clinical anxiety or depressive disorders. The more typical problem was chronic strain: people remained functional, but struggled to rest, sustain motivation, and maintain boundaries.
Somewhat ironically, I asked Fable 5 to estimate BOTEC how much productivity is being lost due to mental health challenges in the AI safety ecosystem. It estimated a loss of roughly 6–7% of total workforce output, with an 80% confidence interval of ~3 to 13%. I only quickly skimmed the report it put together, but the conclusion seems fairly reasonable to me. If anything, my all-things-considered guess is that the real productivity hit is larger.
If I’m right, this seems like a big problem. And as a corollary, if tractable interventions exist here, even modest improvements would be worth a lot.2
What is being done about it?
There’s a bunch of existing stuff I’m aware of related to mental health and AI safety:
EA Mental Health Navigator and MentNav, which provide some mental health resources for effective altruists.
Rethink Wellbeing, which seems to run 8-week guided online programs, group-based peer support formats led by trained laypeople.
I’ve seen a bunch of surveys about mental health run over the years, e.g., this, this, and this (which I referenced above).
The EA forum has a topic on self-care and wellbeing.
Many people (myself included) like this 80k podcast episode. There’s also this one, which inspired this blog post.
Spencer talks about his book. I haven’t (yet) read it, but maybe it’s good!
What else should be done about it on the margin?
Honestly, I don’t know. I’m not sure how tractable this is, and I don’t claim to have particularly great ideas. With this post, I’m mostly just trying to point at a problem, say “oh man, this seems like a big problem”, and hope that others have ideas for how to fix it.3
That said, here are a few things that might be good:
A vetted network of therapists who “get it”. I think it would be good if there was a curated, actively maintained, network of clinicians who have experience working with AI safety clients. As mentioned above, there’s the EA Mental Health Navigator — and ACX has the Psychiat-List — but maybe more can be done? When I was searching for a therapist, I didn’t find these existing resources super helpful.
Better norms in the community. The survey implies that chronic strain (e.g., issues with boundaries, rest, and motivation) is more common than clinical disorders like anxiety or depression. So maybe a lot of the fix has to do with improving the norms that shape the day-to-day experience of working on AI safety. For instance, not socially rewarding hardcoreness as a virtue so strongly.
Expand the group-program model. Rethink Wellbeing’s peer-guided CBT/IFS groups seem useful, at first glance. Maybe someone could do an AI-safety-specific version?
Generous mental health benefits at AI safety orgs. My sense is that many AI safety orgs are good at providing mental health benefits for their employees, but I might be wrong. If the org you run doesn’t offer this, you probably should fix that.
Peer support circles. Lightly facilitated, informal-ish groups of people working on AI safety who meet on some regular cadence to support each other emotionally.
Manager upskilling. Maybe someone (a high-AI-safety-context therapist?) could create resources or offer training sessions on how managers can provide better support to their direct reports.
If you’re interested in starting a project on something like this (or something else aimed at improving mental health in the AI safety space), this RFP from Coefficient Giving might be a good place to get funding. You also might want to check out the EA infrastructure fund.
If you’re applying for funding for some other project/org, consider requesting additional funding specifically to support yourself/your employees with mental health related challenges.
The survey is of 80 people working on existential risk, AI safety, animal welfare, and other high-impact causes. There’s obviously some self-selection bias here, but even after accounting for that, the rates of mental health challenges seem quite high to me. And while the survey population is broader than AI safety specifically, I'm assuming the AI safety subset looks similar (if anything, plausibly worse), but I haven't checked.
The AI safety ecosystem has a high willingness-to-pay for interventions that bring new people into the field. In theory, the willingness-to-pay for interventions that help AI safety people who are struggling with mental health challenges (and are thus less productive than they otherwise could be) should also be high.
After doing some fake math, I'd guess that the ecosystem should be willing to pay at least millions of dollars (maybe tens of millions?!) for something that reduces the productivity drag by a percentage point.
Ironically, one of the hallmarks of AI safety work is that problem diagnosis is enticingly easy, whereas solution generation is much harder.
I'm glad you wrote this! I have some thoughts:
Finding a therapist with deep EA/x-risk/AI safety context, that won’t freak out when you talk about p(doom), is hard, because the majority of therapists that exist come from different intellectual backgrounds to ‘rationalist’, i.e. therapists often, and in fact probably should, value ‘proofs’, truth, efficiency etc less than EAs, and as a result are far less likely to be engaged in the relevant literature/communities. It’s a rare combo of traits for one person to embody, to value both EA values like ‘truth-seeking’ etc, and also believe in or value the principles underpinning many therapy modalities. Which leads me on to my next point…
If somebody does have high EA etc context and these 'therapist qualities', they might be compelled to not become, or leave a career as, a therapist or coach, on the basis that it’s not as impactful as another career is. Therefore, EA rhetoric might have implicitly, unfortunately, encouraged people with the perfect combo of skill sets out of these roles.
Another thought I had is that, with the idea of x-risk/AI safety emotional support groups, I’m always wary, and think reasonably so. EA/AI safety is already a very tight knit professional community, with huge personal overlaps. I think introducing another layer of interaction, whereby people get to know each other in counselling contexts, is probably not a good idea (even if in my ideal world we can all interact with all parts of each other and this functions perfectly. Unfortunately currently I think power dynamics are too integral to the way our society/ies function, society here being s at large and s within EA/AI safety world, and this makes a further dissolution of boundaries hard to advocate for).
This is not to say that group formats shouldn’t be something EAs/AI safety people should consider, since I think peer support is so so valuable, but that we’d want to be careful about trying to ensure no conflicts of interest and preventing other problems etc.
I’m also just sceptical about your suggestion of ‘better norms’. Like, I totally agree that people are working very hard, likely too hard (in the sense that it is burning them out), often driven by terror/panic of impending doom etc, but I’m not sure that most people would actually want to sign off on a cultural shift towards more relaxed working patterns, since most people are genuinely very compelled about the urgency of the things we’re working on. Though, worth noting that even a decade ago lots of early AI safety EAs felt there to be urgency, so potentially we aren’t well calibrated i.e. we always think the situation is dire and immediate. I know you're advocating for an ideal, but it seems important to think concretely/practically about whether we would actually want to implement the ideal. Maybe we could calculate, in collaboration with them, each person's sustainable max output for a unit of time e.g. one week, and help them stay around that mark, instead of advocate for a looser, broader, 'work less, relax more' approach, because different people can work different amounts. I think this then could be used in conjunction with your good management idea, where the manager helps the person stay working within that guideline. I for example know that I can sustain focused creative 'thinky' work time for a max of 3-4 hours a day, where some people can only do 1-2 hours, and some can do more at 5-7.
Which leads me onto a bigger broader potential point of intervention which is helping people reflect more deeply about the type of roles that align with their natural rhythm and capacity. Some people with technical AI safety knowledge/context/excitement might be well placed in a research post, but some others, who realise they don't enjoy constant reading, writing/running code (idk, I'm not a super technical person lol), might be better fitted in an RM position, or as a grantmaker, etc. Not to be too hippy woo woo, but I think some of the problems we are seeing are a result of a more root disconnect from ourselves and reflecting on how we want to spend our 80,000 hours. Instead of deciding just from the head, as in where can I be most maximally effective, we could instead do some combo (the guiding principles for which I am not sure I can elucidate robustly nor compellingly) of 'what am I excited about' and 'what is important', i.e. incorporating the heart too. I guess maybe some EAs think they're doing this and/or just are doing this, but my read is that lots of people actually get swept up in 'what is the hive mind telling me I should do', and whilst I like the idea of the community guiding talent into different positions (even if I'm not convinced it's been done this entirely well in the past), I also want people to retain autonomy over their decisions, and retain connection to the more fluffy, intangible, hard to convey, hard to codify feelings about what they want to do for 40+ hours a week. My reasoning here, which I think has long been buried, is that if more people took a more holistic approach, then perhaps burnout, depression, anxiety, imposter syndrome etc, could be reduced.
I liked reading about this and I'm curious. I've noticed a recent change to Claude's thinking display — it went from first-person relational thinking to impersonal fragments that resemble an old time telegram. Do you know which team typically makes those kinds of decisions? Is it safety?