I recently wrote a comment focused on the "AIS gig economy" that I think is relevant here. In particular for mid-career transitioners who will, on balance, have less patience and life slack to absorb the effects of participating in it.
A tentative comparison of the early AIS workforce to early-career academia suggests that a gig work culture could be a factor in/source of at-scale churn. It is in academia.
I think focusing on the real trajectories of people transitioning into AIS work, specifically how funding is currently materializing into early employment opportunities and how entrants experience navigating that path, seems high impact and low effort to evaluate (surveys).
And finding evidence supporting the thesis that early AIS and academic work are similar would be a good: infra-level interventions on the workforce funnel and good industry/community practices are easy to adopt early, hard to adopt later, and a structural problem has a "one-point" intervention, whereas individual-worker-level interventions and support structures that are needed to deal with its downstream effects scale O(n).
I'd like to see more (and more critical) assessments of the field's onboarding realities along these lines.
I agree that academia has generally done a poor job addressing its mental health crisis (I certainly had that experience, personally). At the least, though, their problem is older, well quantified, and their intervention outcomes are well studied, so, to the extent that we can exapt lessons from there, AI safety's approach to mental health can be (or begin) better informed.
I’d like to examine here how much the early career academia pattern is mirrored in the AIS workforce. Specifically, I think this premise is of large consequence to get wrong:
Academia, and even more so, early career academia, is notorious for low mental health and this also ties in with bad pay, oversupply of PhDs, and no job security. This is quite different for AI Safety with it being comparatively well-funded and without those same structural stressors
Scoping what constitutes the AIS workforce seems like a good first step, because arguably the workforce is much larger than those whose conditions can be characterized as “comparatively well-funded and without those same structural stressors.”
Surprisingly, there doesn't seem to be a real census of the AI safety workforce that distinguishes stable employment from fixed-term employment, fellowships, independent grant-funded research, PhDs, unemployment/job-search, etc. 80,000 Hours has been collecting something resembling a census since 2024, but as far as I can tell it hasn’t published aggregate employment-status results.[1] So I don't think we actually know what fraction of people seriously pursuing AI safety careers have stable jobs.
The evidence I can find suggests a fairly substantial AI safety gig economy. One 2025 estimate found about 1,100 FTEs at self-identified AI safety organizations[2] (likely a few thousand undercount given the qualifier, according to 80k). Meanwhile, a 2026 projection from people running AI safety talent programs expects 2,000-2,500 research-fellowship seats this year, plus ~300 non-research fellowship seats.[3] These aren't directly comparable since fellowship seats are annual flow rather than workforce size, and many are short or part-time, but they make it hard to regard temporary/soft-money work as a marginal feature of the field.
The outcome data too looks pretty similar to early career academia for these folks. A dataset of 647 alumni from nine relatively late-stage AI safety/governance fellowships found that 21.5% had done another such fellowship somewhere in their trajectory, and 11.1% did another one afterward.[4]
MATS acceptance rate was ~15% circa 2024 (more recently 4-7%), with its co-director saying many rejected applicants are proficient researchers without better options. 63% of all past scholars applied for independent-researcher funding immediately post-program to continue 4+ months on stipends he describes as LTFF-equivalent - "effectively self-employed."[5] Since then that funnel widened, with roughly 75% of fellows now continuing into a 6-12-month extension,[6] entry to which required (in 2024) securing an external grant, typically from LTFF or Open Phil.[7] In the one cohort MATS has tracked post-extension (n=18 though), five months out 33% were doing independent research, 22% were in PhDs, ~17% at established organizations, and 17% unemployed, a rate MATS attributes partly to a "gap between available safety positions and number of interested researchers".[7:1]
On the demand side, MATS's Q4 2025 interviews with hiring managers and funders describe organizations forced into hyper-selectivity, hiring only near-autonomous senior workers, while labs largely fill roles internally.[8] MATS also reports ~80% of alumni remain in AI safety[6:1], which, in light of the earlier numbers, suggests retention means circulation among temporary positions more than stable placement.
This is all really close to the postdoc pattern.
I'd qualify the bad pay part of the academic analogy since some AI safety fellowships pay pretty well. As I think is apparent above, the stronger parallel is with precarity, since a lot of the top-of-funnel positions are in 9-12 week or 3-5 month programs, followed by another application for an extension, grant, fellowship, job. In that respect the academia comparison might even understate the instability: postdocs are insecure, but generally operate on year-scale rather than month-scale funding timelines.
So I don't think the comparison lets us cleanly say that academia has bad mental health because it has oversupply, insecure employment and career competition, whereas AI safety lacks those stressors and therefore similar outcomes point toward the special psychological burden of x-risk. The first premise looks wrong if “AIS workforce” is interpreted honestly. AI safety appears to share quite a few of the same ordinary occupational stressors - scarce stable positions, repeated re-selection, soft-money research, status competition and uncertain career progression - even if the field as a whole is well-funded.
That doesn't imply x-risk contributes little, or that tailored support isn't useful. It just makes me less confident that the similarity in mental-health outcomes is evidence for something distinctive about confronting catastrophic risk rather than, at least in significant part, a familiar labor-market structure.
For what it's worth, I personally think stepping out of the mould of academia's interventions is high-value, which is why I'm focusing on understanding what the labour-market morphism looks like. To the degree there's a match, we have good information about what's been tried already.
I think it’s useful to ground these results by looking at another setting in which cultural and institutional norms put ambitious, intellectually selected young people into a similar context, which I think can be reasonable characterized as high-stakes, status-competitive, weak-boundary work environments.
Here are some comparisons GPT found to people in early-career academia:
Survey here
Early academia
Difficulty switching off from work
68%
78% ranked difficulty maintaining work–life balance among their top five concerns. 49% reported a long-hours culture at their university, explicitly including sometimes working through the night.[1]
Imposter syndrome
60%
42% experienced frequent imposter thoughts. More than 26% experiencing serious/intense imposter thoughts.[2]
39.3% reported anxiety-disorder symptoms.[4] A different study found 17% had clinically significant symptoms of anxiety.[5]
Hopelessness or depression
29%
26.5% depressive-disorder symptoms.[4] A different study found 24% had clinically significant symptoms of depression.[5]
Loneliness or isolation
32%
24% severely or very severely lonely (note this was collected during COVID).[6]
Considered cutting back or leaving
45%
51% considered leaving because of work-related mental health concerns.[4]
It seems worthwhile to me to question the premise that AI safety/x-risk work is unusually psychologically harmful. I’m not in the field, but I think there is a somewhat romantic story available here: that these problems arise because people are grappling with x-risk, carrying a moral burden, and working on something unusually consequential. That story is surely partly true, but it’s also flattering. Ingroup dynamics give us some reason to suspect that attributing distress to the exceptional significance of the work may be a motivated explanation, when a less glamorous one is that AI safety has reproduced a familiar high-pressure knowledge-work environment: long hours, poor boundaries, status competition, perfectionism, career uncertainty, people who are willing to sacrifice other parts of their lives to succeed.
The practical downside of viewing the problem as exceptional is that solutions become less available, need to be built bespoke. But if a large part of the variance is coming from more ordinary occupational dynamics, then a much larger institution (academia) has spent a long time running into the same failure modes and trying to intervene. There may be a helpful update here that upweights boring and well-established interventions.
I recently wrote a comment focused on the "AIS gig economy" that I think is relevant here. In particular for mid-career transitioners who will, on balance, have less patience and life slack to absorb the effects of participating in it.
A tentative comparison of the early AIS workforce to early-career academia suggests that a gig work culture could be a factor in/source of at-scale churn. It is in academia.
I think focusing on the real trajectories of people transitioning into AIS work, specifically how funding is currently materializing into early employment opportunities and how entrants experience navigating that path, seems high impact and low effort to evaluate (surveys).
And finding evidence supporting the thesis that early AIS and academic work are similar would be a good: infra-level interventions on the workforce funnel and good industry/community practices are easy to adopt early, hard to adopt later, and a structural problem has a "one-point" intervention, whereas individual-worker-level interventions and support structures that are needed to deal with its downstream effects scale O(n).
I'd like to see more (and more critical) assessments of the field's onboarding realities along these lines.
I agree that academia has generally done a poor job addressing its mental health crisis (I certainly had that experience, personally). At the least, though, their problem is older, well quantified, and their intervention outcomes are well studied, so, to the extent that we can exapt lessons from there, AI safety's approach to mental health can be (or begin) better informed.
I’d like to examine here how much the early career academia pattern is mirrored in the AIS workforce. Specifically, I think this premise is of large consequence to get wrong:
Scoping what constitutes the AIS workforce seems like a good first step, because arguably the workforce is much larger than those whose conditions can be characterized as “comparatively well-funded and without those same structural stressors.”
Surprisingly, there doesn't seem to be a real census of the AI safety workforce that distinguishes stable employment from fixed-term employment, fellowships, independent grant-funded research, PhDs, unemployment/job-search, etc. 80,000 Hours has been collecting something resembling a census since 2024, but as far as I can tell it hasn’t published aggregate employment-status results.[1] So I don't think we actually know what fraction of people seriously pursuing AI safety careers have stable jobs.
The evidence I can find suggests a fairly substantial AI safety gig economy. One 2025 estimate found about 1,100 FTEs at self-identified AI safety organizations[2] (likely a few thousand undercount given the qualifier, according to 80k). Meanwhile, a 2026 projection from people running AI safety talent programs expects 2,000-2,500 research-fellowship seats this year, plus ~300 non-research fellowship seats.[3] These aren't directly comparable since fellowship seats are annual flow rather than workforce size, and many are short or part-time, but they make it hard to regard temporary/soft-money work as a marginal feature of the field.
The outcome data too looks pretty similar to early career academia for these folks. A dataset of 647 alumni from nine relatively late-stage AI safety/governance fellowships found that 21.5% had done another such fellowship somewhere in their trajectory, and 11.1% did another one afterward.[4]
MATS acceptance rate was ~15% circa 2024 (more recently 4-7%), with its co-director saying many rejected applicants are proficient researchers without better options. 63% of all past scholars applied for independent-researcher funding immediately post-program to continue 4+ months on stipends he describes as LTFF-equivalent - "effectively self-employed."[5] Since then that funnel widened, with roughly 75% of fellows now continuing into a 6-12-month extension,[6] entry to which required (in 2024) securing an external grant, typically from LTFF or Open Phil.[7] In the one cohort MATS has tracked post-extension (n=18 though), five months out 33% were doing independent research, 22% were in PhDs, ~17% at established organizations, and 17% unemployed, a rate MATS attributes partly to a "gap between available safety positions and number of interested researchers".[7:1]
On the demand side, MATS's Q4 2025 interviews with hiring managers and funders describe organizations forced into hyper-selectivity, hiring only near-autonomous senior workers, while labs largely fill roles internally.[8] MATS also reports ~80% of alumni remain in AI safety[6:1], which, in light of the earlier numbers, suggests retention means circulation among temporary positions more than stable placement.
This is all really close to the postdoc pattern.
I'd qualify the bad pay part of the academic analogy since some AI safety fellowships pay pretty well. As I think is apparent above, the stronger parallel is with precarity, since a lot of the top-of-funnel positions are in 9-12 week or 3-5 month programs, followed by another application for an extension, grant, fellowship, job. In that respect the academia comparison might even understate the instability: postdocs are insecure, but generally operate on year-scale rather than month-scale funding timelines.
So I don't think the comparison lets us cleanly say that academia has bad mental health because it has oversupply, insecure employment and career competition, whereas AI safety lacks those stressors and therefore similar outcomes point toward the special psychological burden of x-risk. The first premise looks wrong if “AIS workforce” is interpreted honestly. AI safety appears to share quite a few of the same ordinary occupational stressors - scarce stable positions, repeated re-selection, soft-money research, status competition and uncertain career progression - even if the field as a whole is well-funded.
That doesn't imply x-risk contributes little, or that tailored support isn't useful. It just makes me less confident that the similarity in mental-health outcomes is evidence for something distinctive about confronting catastrophic risk rather than, at least in significant part, a familiar labor-market structure.
For what it's worth, I personally think stepping out of the mould of academia's interventions is high-value, which is why I'm focusing on understanding what the labour-market morphism looks like. To the degree there's a match, we have good information about what's been tried already.
80,000 Hours' AI-risk career census initiative. ↩︎
Bottom-up estimate of roughly 1,100 AI-safety FTEs. ↩︎
2026 estimate of AI-safety fellowship throughput. ↩︎
Career-path analysis of 647 fellowship alumni. ↩︎
Ryan Kidd’s comment on Arb Research’s Impact Assessment of AI Safety Camp post. ↩︎
MATS program overview with scale and retention figures. ↩︎ ↩︎
MATS follow-up data on employment outcomes and funding stability. ↩︎ ↩︎
2026 interview study of AI-safety hiring constraints. ↩︎
I think it’s useful to ground these results by looking at another setting in which cultural and institutional norms put ambitious, intellectually selected young people into a similar context, which I think can be reasonable characterized as high-stakes, status-competitive, weak-boundary work environments.
Here are some comparisons GPT found to people in early-career academia:
It seems worthwhile to me to question the premise that AI safety/x-risk work is unusually psychologically harmful. I’m not in the field, but I think there is a somewhat romantic story available here: that these problems arise because people are grappling with x-risk, carrying a moral burden, and working on something unusually consequential. That story is surely partly true, but it’s also flattering. Ingroup dynamics give us some reason to suspect that attributing distress to the exceptional significance of the work may be a motivated explanation, when a less glamorous one is that AI safety has reproduced a familiar high-pressure knowledge-work environment: long hours, poor boundaries, status competition, perfectionism, career uncertainty, people who are willing to sacrifice other parts of their lives to succeed.
The practical downside of viewing the problem as exceptional is that solutions become less available, need to be built bespoke. But if a large part of the variance is coming from more ordinary occupational dynamics, then a much larger institution (academia) has spent a long time running into the same failure modes and trying to intervene. There may be a helpful update here that upweights boring and well-established interventions.
Nature's 2019 global survey of 6,320 PhD students
A study of 302 Australian PhD students. Also in there, Imposter thoughts predicted symptoms of depression, anxiety and suicidality.
A study of 62 master's and doctoral psychology students at a Canadian university.
Nature's international survey of 7,600 postdocs.
A meta-analysis of nine studies, 15,626 PhD students.
A study of 222 doctoral researchers in the Berlin area.