13 people with an average of 15 years of professional experience in technology, policy, operations, law, consulting, and communications described entering AI safety as the most complicated career transition they had undertaken. While part of it can be theoretically explained by the features of the domain itself and of the roles the transitioners were pursuing, we want to explore if we could learn from their experience. In this article, we look at AI safety talent pipelines through the eyes of transitioners who have spent decades close to hiring, talent development, people management, and making their previous career transitions in industries outside the field, set alongside the perspective of organisation leaders and talent practitioners in AI safety.
After publishing our first article on the talent pipelines, we had 25 new interviews, 18 with experienced professionals making deliberate and substantial efforts to join AI safety work and 7 with organisation leaders and talent practitioners. Both perspectives are included, though the balance naturally reflects the sample: more depth on the transitioner experience, with the organisational side providing context where available. The transitioner interviews mostly came from the outreach the first article generated. The analysis methodology follows that described in the first article.
For at least a year, the AI safety ecosystem has been broadcasting a clear message: the field needs more people, including from other domains (as we observed in our first article). The field has been actively preparing for the response: launching pipeline programs and championing talent matching initiatives. That message is reaching the addressees, and the professionals from other areas, at least those we talked to, have taken that invitation seriously.
The professionals we interviewed bring, on average, 15 years of experience (some up to 30+), spanning technology, policy, law, consulting, government, healthcare, communications, operations, and education.
The response is growing. Two independent accounts, both 25+ years of experience, one in international policy and another in government, describe pools of colleagues who consider moving to AI safety. This can be indirectly confirmed by the capacity challenges reported by talent-matching organisations we spoke with. It also aligns with what five interviewees described as a broader pattern: growing public discourse about AI risk reaching professionals well beyond the existing community, combined with recent disruptions in government, nonprofit and big-tech employment releasing experienced people who are looking for mission-driven work into the market.
The motivations our transition-side interviewees describe, for themselves and among their professional networks, are worth noting: concern about existential and catastrophic risk, a desire to apply hard-won institutional experience to problems that matter, and a sense of responsibility sharpened by parenthood, career maturity, or proximity to the technology itself.
The commitment is not abstract. At least three interviewees left stable, well-compensated positions to pursue this transition, simultaneously bearing responsibility for their families. Others are investing hundreds of hours alongside full-time work. One interviewee has been applying for months while writing regularly about AI safety; another built evaluation tools at night and on weekends without compensation; most of the people we've talked to report hundreds of hours of dedicated transition activity spent on applications, courses, fellowships, conferences and mentoring programs.
From the organisational side, the response is visible as well. Hiring managers we spoke to describe receiving hundreds of applications for single roles, with one organisation reporting 670 applications for a single special projects position. Consideration time varies from 30 seconds (account by an interviewee) per CV to ~5 minutes per candidate, mentioned in this article.
On the quality side, one org leader described being "so impressed" by the richness of operational backgrounds in their applicant pool. Even amid the volume, hiring managers report a non-negligible proportion of genuinely strong candidates. One described routinely identifying at least 20 candidates per round who appeared immediately qualified to advance.
On the other hand, quality-wise, other hiring teams report challenges with application quality. We have received more accounts confirming that AI-generated applications have become a growing problem. Counterintuitively, though, expecting more reports on low-quality AI content, we have instead learnt that it has also become the problem that candidates submit resumes so precisely tailored to role descriptions that they could not possibly be authentic. One recruiter described encountering applicants who, once on a call, could not speak to the experience listed on their own resumes. Others noted that even among genuine applicants, many lacked the practical ability to perform the work when tested. One organisation found that candidates with impressive operational resumes were "surprisingly underperformant" in work trials, while a candidate with a less obviously matching background excelled by demonstrating prioritization, thoroughness, and independent judgment.
Besides, multiple interviewees pointed to the discrepancy produced by evaluating the following 4 facts together:
The coexistence of these seemingly contradictory facts could point to something specific.
What emerges from the data may not be a simple story of good talent being turned away or an absence of quality within the pool. It might be something more structurally complex. Let’s explore that.
After all the interviews (30+) for both articles, with org leaders and transitioning experienced professionals, work in AI safety to us is increasingly resembling mounted archery. The target (AI going well[1]) is constantly moving, the steeds are moving (the orgs[2]), the archers are moving (people doing the work[3]), the bows and arrows are moving (intervention uncertainties) and the steppe is on fire (the world around). If this metaphor is at least half accurate, it explains the particular talent expectations as well as the ways the talent pipelines are being built. A particular type of talent might be needed and the hiring error might be costly.
Seven interviewees from the transitioning side pointed to noticeable and understandable conservatism in AIS hiring practice, however raised the question whether these practices still serve the field well. This relates to the duration of some hiring cycles, spanning multiple months and accounts of explicit preference to have an unfilled position over taking risk to hire the wrong person. They noted a contrast between this cautious approach to staffing and the field's otherwise high risk tolerance in areas like grantmaking. Some observed a general hesitation to onboard external talent, particularly individuals with highly established backgrounds. One of the seasoned professionals we talked to accounted for offering volunteer support and still facing high barriers to entry.
One senior professional interviewee went through an eight-stage process over eight months for a role managing a small team, including two work tests totalling eight hours, multiple leadership interviews, and a work trial, more than they would experience for a higher-level position in another industry. Another observed that when work tests involve multi-day in-person participation, candidates outside a small number of geographies face practical constraints around travel, visas, and time away from existing commitments. A third, after completing weeks of applications and work tests, found that the final interviewer was not familiar with their earlier submission. The same interviewee encountered application forms with time estimates of half an hour that contained multiple essay questions.
Multi-stage hiring pipelines reflect a genuine desire to be thorough, but in many cases the investment required from candidates has grown disproportionate to the scope of the role. Processes spanning many months with extensive written components can place a heavy burden on applicants. When popular programmes receive thousands of applications, the aggregate time candidates invest is substantial. A back-of-the-envelope calculation reveals thousands of hours high-competence candidates spend on applications that produce no or little value for themselves or the community. [4]
Is it possible to re-evaluate how organisations approach the hiring risks? When organisations hold hiring cycles or hold positions open for extended periods in search of a perfect fit, they are expressing a genuine, but perhaps outsized commitment to quality. The opportunity might be in pairing that commitment with a fuller accounting of what prolonged vacancies cost. It is true that a wrong hire is not neutral to an organisation’s work and can cause harm. It is also feasible to recognise, that under the compressed timelines the field takes seriously, every month a role remains unfilled is a month of organisational capacity waiting to be unlocked. Recognising this means applying the same rational risk reasoning to hiring that the field may already be applying in other domains.
It might be helpful to experiment with some practices our interviewees noticed in other industries. E.g. One technology professional who participated in multiple hiring processes in big-tech companies, described the most effective processes as ruthlessly efficient, objective, and fast. A central practice was to make an offer immediately to the first candidate who checked all the boxes. Another technique would be to primarily filter by what an applicant would add that a team does not yet have. We have not got enough data to evaluate how common these approaches are already within the AIS field, and also wonder if there are many more risk-related techniques worth learning.
Other questions left open:
There are clear opportunities to build better infrastructure across talent evaluation, role definition, hiring process design, and application feedback.
Developing domain expertise in talent evaluation. As organisations that began as research-focused expand into operations, communications, and programme management, there is a natural learning curve in designing evaluations for unfamiliar functions. Eight accounts highlighted moments where experienced candidates encountered interviewers still building their assessment skills in these domains. Some interviews relied heavily on AI-generated questions or were conducted by evaluators early in their own careers. Solutions could include shared capability frameworks, a common vocabulary for non-research competencies or distribution of target competencies between interviewers.
Designing roles with greater intentionality. Five interviewees from non-AIS domains described roles assembled by gathering unresolved organisational needs into a single listing, producing job descriptions that bundle several distinct functions together. Four described terminology that does not map to standard industry equivalents, making listings difficult for domain specialists to find or interpret. A smaller number identified specific professional categories, including product management and strategic communications, that have not yet been recognised as distinct functions within the field, despite being well-established disciplines in adjacent sectors. What might work is experimenting with role descriptions around more “outside-world” terminology and functions split.
Building a culture of useful feedback. Seven interviewees pointed to closing the feedback gap as one of the field's most significant opportunities. Currently, many candidates at all stages receive little or no information about why they were not selected, which makes it difficult for individuals to improve and for the system as a whole to learn. When feedback is provided, it sometimes prioritises caution over clarity. Candidates are often left unable to distinguish between a genuine skill gap, a communication mismatch, a timing issue, or a decision driven by other factors. Some promising models already exist: at least one organisation has experimented with group-themed feedback organised by rejection stage. We wonder how doable it is to scale this approach and whether it would provide a reasonable learning ROI to the above-mentioned application costs for transitioners and the community.
Feedback from candidates to organisations represents another untapped resource. An interviewee noted that they hold back feedback in settings where the same organisation also serves as a reference and is in close contact with others they may later approach. In a tightly networked field, candid input carries considerations that make honest feedback difficult to offer while these functions remain combined. One of the interviewees described a practice used in adjacent industries where an applicant would be assigned with an “applicant buddy” who would liaise communication and feedback both-ways, being outside of the hiring process. We are not confident how feasible it is for AIS organisation at this stage, however we wonder about other potential intermediaries. Could an intermediary, collecting feedback and returning it to organisations in aggregate, unlock this as a source of institutional learning?
Building a picture of the field's own workforce. Two interviewees, independently (one from the field-building and another from the transitioning professionals side), pointed out that having public taxonomy on the talent and skills distribution within AIS orgs and skill-specific gaps would help both organisations and transitioning professionals make more informed decisions. We have not yet conducted in-depth research to verify the need for those and would welcome some directions to the existing solutions.
As one interviewee put it, AI safety hiring runs on a signalling system built on its EA and academic origins, while professionals from outside the field carry their own signalling system, built on decades of practice in different domains. Other transitioning side interviewees echoed this observation. Candidates are screened for familiarity with cultural frameworks, participation in recognised programmes, or fluency in a particular idiom. The friction is compounded by the order in which filters are applied.
This showed up in how interviewees described their own processes. One interlocutor from the talent-matching side described being asked to screen candidates on qualities like taste, judgment, and unique online presence, without any shared definition of what those meant. Five hiring-side interviewees independently described values-screening questions calibrated to community familiarity rather than professional competence: event attendance, programme participation, and vocabulary fluency, signals primarily acquired through participation in the community itself, rather than through professional experience elsewhere. Three organisations report difficulty reading credentials from outside the ecosystem, and nine candidates report not knowing which signals matter, since those signals remain unstandardised and often unarticulated even within the organisations relying on them.
One transitioning-side interviewee questioned whether recruiting organisations are capable of reading the signals that mid-career and senior professionals produce. Another, with roughly fifteen years of senior digital product experience across industry, described organisations being unable to identify any way they could use their skill set once they arrived. A third with a decade of product experience described being advised to demonstrate agency by building something, but found that nobody could specify what kind of project would constitute a valid signal, or for whom. They observed: if the field screens for the kind of extreme self-direction that would lead someone to build independently, it might be selecting for people who would not seek employment in the first place. Another described systematically testing different framings of the same experience depending on whether an organisation appeared to prioritize alignment or competence, because the operative evaluation paradigm was not visible from the outside.
The same gap does not stop at job postings and early screening. In Article 1, we flagged, as an open question, whether work-sample test design reliably measures what a role actually needs, rather than what happens to be easy to test. One account from this round suggests the gap can reach further into the process too: a work-sample test built around one skill set was used to evaluate a role whose own description centred on a different one.
Six interviewees framed this as a solvable design problem: could evaluation criteria be developed that match the breadth of talent the field says it needs for specific roles, rather than the breadth it currently screens for? At least four interviewees from various sides independently proposed approaches, from profession-specific guidance mapping industry roles to AI safety equivalents, to role taxonomies organized by professional discipline rather than cause area.
Other questions left open:
The practical costs of the current pathway shape who is able to participate. Unpaid and lightly compensated entry work favours candidates with financial reserves, household support, or independent wealth. Three interviewees noted they could personally afford the transition, but pointed to others they knew attempting it who could not. But sustained contribution eventually requires financial sustainability, regardless of how impact driven the motivation is. Three interviewees were successful in receiving career transition grants or securing paid fellowships. At this sample scale, these conversion numbers might not be unusual, but the pathways that do exist might be not universally visible or accessible. As one interviewee who successfully transitioned put it, the career transition grants ideas, while valuable in principle, are opaque and difficult to navigate from the outside. Interviewees described encountering similar dynamics in grant application processes to those described elsewhere in this piece for job applications.
Financial strain is one reason some people do not make it through. Another, and sometimes complementing, barrier is not knowing whether trying is worth it at all. Among the transitioning professionals interviewed, seven described reaching a point where continued investment in the transition no longer felt sustainable. Alumni of transition programs report that meaningful numbers of their cohort members, including people with deep experience in technology, entrepreneurship, and financial operations, concluded the pathway was not open to them and redirected their careers. As one put it: "there are gates and keys and locks that you don't understand, you don't know where they are, and you don't know who holds the key, and you don't know how to talk to them.". Others were advised by leaders of the very pipeline organisations that they should consider other directions.
Each departure represents not only an individual decision but the loss of months or years of context-building, relationship development, and skill investment the ecosystem helped produce. Better feedback, of the kind described in the Opportunities for stronger talent infrastructure section, is one possible retention lever, offering some return on the investment (learning) even where there is no impact or income yet. Are there mechanism that could help the field hold onto the people it's already spent months or years bringing in?
Open question:
Among transitioning interviewees who invested substantial time pursuing recognised entry pathways, five have channelled their energy into building independently, and three are actively exploring that direction.
After months spent navigating applications, courses, fellowships, and volunteer opportunities, unsatisfied with the input-to-return ratio in terms of learning and impact, many described arriving at a natural turning point. Some decided to build specific initiatives rather than wait for institutional entry, registering legal entities, launching consultancies, or creating paid services. Others found that simply stepping back from the application cycle and redirecting energy toward independent projects and conversations was itself what allowed them to start contributing.
However, one interviewee raised a strategic question as a counterweight to this pattern: while celebrating founder mindsets has real merit, does it also risk producing a wide landscape of disconnected efforts, when the greater need may be helping promising existing projects scale? The same interviewee also offered a counterweight to their own concern: with no guaranteed pathway to making AI safety go well, trying many things in parallel may be the best available strategy, not a failure to coordinate.
Other questions left open:
The patterns above have led to more questions than answers. We hope they will find their way into discussion, and we would welcome being part of it. With this piece, and with the one before it, we want to help move things toward organisations converting more of the talent already arriving into capacity, and transitioners converting more of their effort into roles where that capacity gets used to produce impact. In the meantime, working through the interviews and this article, one line of inquiry led us to a problem-solution hypothesis we are now validating separately, outside the scope of what is written here.
There is also one more thing we are curious to explore:
One interviewee pointed to the US federal government's former Presidential Management Fellowship, created out of a structural gap in civil service entry, as a model worth learning from. It ran on a single, nationwide assessment open to any graduate field, chemistry to public policy, PhDs included, testing what the interviewee called minimum standards: can you write, analyse a problem, come up with recommendations, and communicate well with others. Passing gave candidates what they called a 'free ticket,' recognised across every federal agency, then two paid years to find out whether the fit ran both ways. As they put it, the underlying premise was that "the preceding experience, educational or professional, is sufficient to do the work; there just needs to be a program that brings people from that world into the world that needs the help." We recognise this solution cannot be adopted verbatim for a variety of differences, for instance, PMF fed into one employer, the federal government, while AI safety is a multi-actor ecosystem with no single entity to place people into. However, we wonder if AIS could draw some separate practices from this example.
We are curious whether other fields have faced something similar. For example, biotech and how they covered the shortage of science-fluent managers, and open-source software's path to formalising generalist roles, both seem like they might rhyme with what we're describing here, but we do not know enough to say. If you do, we would like to hear from you.
You can contribute by leaving a comment, contacting us directly (DM or [email protected] ), or filling in our short form (~5 min) to share your preferred means of communication with some background. Thank you for anything you're willing to share!
We appreciate the time our interviewees invested in sharing their experience and thinking with us, and for being open to helping us further. We are grateful to our mentor Martyna Wielopolska, who has stayed involved since the Sentient Futures Project Incubator, continuing to offer sharp questions and candid feedback on our work. And we are thankful to Kevin Flick, PhD, for rigour and honest challenge on the draft.
Most data collection, analysis, and drafting were done manually. We used AI notetakers during interviews, with consent, to support human-driven thematic analysis. For drafting, we used an LLM for structural editing and to help catch unsupported claims and logic gaps. All findings, claims, and conclusions were reviewed and validated manually.
AI progress
Uncertainties with timelines, funding, impact, talent etc.
Working, learning, fighting FOMO
Five interviewees explicitly mentioned that the expected application times are often significantly deflated in comparison with the real time to produce a quality application. The following is a deliberately conservative, illustrative estimate. If we take a 4-hour application process spanning 500 candidates: that's 2,000 hours of collective effort, of which only 4 hours belong to the one candidate who gets the job. The other 1,996 hours produce no outcome for the applicant who spent them. Even assuming 90% of applications are AI slop, and so took little genuine effort to produce, the remaining 10% of high-quality applications still account for 200 hours of quality effort with no outcome, more than a full working month (at ~40 hours a week) spent on a single position. It takes just 10 such positions to reach thousands of hours of this genuine, no-ROI effort. The calculation does not account for multi-stage processes, which would push the count higher still.
