a guide to entry points into AI safety governance
where you are where the work is

The Feeder MapWhich Nearby Roles Might Lead Toward AI Safety Governance?

  • Singapore
  • Malaysia
  • Philippines
  • Thailand
  • Indonesia
  • Vietnam

The work that motivates this map is the governance of risks from advanced AI, usually called AI safety governance, and most of it is advertised in London, Washington, Brussels, San Francisco, and a small number of international institutions. The map reads 284 job postings line by line and asks which entry-level roles in Southeast Asia appear to build the abilities those destinations ask for, so that a first job can be weighed by where it may lead as well as by what it is.

The map stands on a corpus of 284 job postings, 129 destination roles at the frontier and 155 starting roles across Singapore, Malaysia, the Philippines, Thailand, Indonesia, and Vietnam. Reading them line by line produced 2,357 coded requirement lines, and 234 of the postings are live advertisements, each linked to its source.

wave four corpus, collected 2026

01Start Here

This section sets out the decision the map is meant to inform, the field it has in view, and the limits of what a reading of job postings can say about either.

Plate I. A first job as a fork in the path, which the map tries to make a little more legible.

The map begins from a practical difficulty. Graduates in Manila, Jakarta, Kuala Lumpur, Bangkok, Ho Chi Minh City, and Singapore who hope to work, eventually, on the governance of advanced AI find that the organisations doing that work, safety and security institutes, the policy and safety teams of frontier developers, and the research groups and funders around them, mostly hire in a few distant cities and seldom at entry level. The roles available nearby tend instead to be compliance seats, project officer posts, and research assistant positions at institutions whose work touches AI only obliquely, if at all. The question is whether one of these roles is likely to move a person toward the field or away from it, and how one might tell before committing.

Job titles are a poor guide to that difference. A title compresses many particular demands into a few words, and two roles with similar titles can ask for quite different things. The map therefore sets titles aside and reads the requirement lines themselves. Its working premise, which later sections examine rather than assume, is that a role which asks for many of the abilities a destination asks for is likely to develop them, and so may serve as a first step toward it.

We read 2,357 requirement lines from 284 postings, 155 of them entry roles in six Southeast Asian markets and 129 of them destination roles. For each entry role we estimated how much of each destination’s demand it appears to supply, adjusted that estimate for how open the role is to a new graduate, and kept every line linked to its source so that the reading can be checked. The result is perhaps best treated as a structured way of comparing options, and not as a verdict on any of them.

Because the scope of the field shapes everything that follows, the next section defines what we mean by AI safety governance before turning to the idea of a feeder role on which the measurement rests.

02The Field and the Idea

This section defines the field the map has in view, describes what we mean by a feeder role, and explains why the collection included roles adjacent to the field as well as roles inside it.

builds towardlittle in return
Plate II. The map frames each comparison from the near bank toward the far one, since a role may build toward a destination that builds little in return.

What We Mean by AI Safety Governance

AI governance, broadly understood, covers a wide range of concerns, among them data protection, fairness in automated lending or hiring, content moderation on platforms, and the management of model risk in banks. AI safety governance is a narrower and more recent part of that field. It is concerned principally with highly capable, general-purpose AI systems and with the possibility that they cause severe or large-scale harm, whether through deliberate misuse, as in assistance with biological or cyber attacks, through failures of control, in which a system pursues ends its developers did not intend, or through wider disruptions that no single actor chooses.

Its instruments are correspondingly specific. They include the evaluation and red-teaming of frontier models before release, the safety frameworks in which developers commit to precautions tied to capability thresholds, the national safety and security institutes established since 2023, standards and third-party assurance for advanced systems, incident reporting, the governance of compute, and the international coordination pursued through the summit series that began at Bletchley Park and through the International AI Safety Report. Southeast Asia is not absent from this work. Singapore hosted the scientific exchange behind the Singapore Consensus on Global AI Safety Research Priorities in 2025, and the corpus records Singapore postings from several frontier developers and safety research organisations.

This is the field the map has in view over the long term, and it is worth being plain that the destination corpus is wider than the field so defined. Alongside roles at safety institutes, in the policy and safety teams of frontier developers, and at safety research organisations, it includes postings in platform integrity, financial model validation, enterprise AI governance consulting, and the digital programmes of development agencies. These were included because they call for much the same kinds of work, policy analysis, assurance, programme delivery, communication, and research support, and because they are plausible intermediate destinations in their own right. The five families into which the destinations are grouped describe kinds of work, not degrees of proximity to safety, and the map does not yet weight destinations by that proximity. A high score is therefore evidence that a role builds toward a kind of work which safety governance also requires, rather than evidence that it leads to safety governance in particular. Tagging each destination by its relation to safety, and recomputing the scores on that basis, would be a natural next revision.

The Stepping Stone

The image behind the map is a river crossed on stepping stones. A destination role in London or San Francisco stands on the far bank, an entry role in Ho Chi Minh City or Kuala Lumpur is a stone in the water, and the question asked of each stone is how far toward the far bank it might carry someone who stands on it. The image simplifies, since careers are seldom linear, but it captures the feature that matters most for the measurement, which is direction.

A feeder role is one that builds toward a destination, and the relation need not hold in reverse. Standards work at a certification body in Kuala Lumpur may develop much of what an EU AI Act assurance role in Brussels asks for, while the Brussels role, as advertised, develops comparatively little of what the Kuala Lumpur role requires. The map therefore frames every comparison in one direction, from the entry role toward a destination family, and assesses reachability for the entry role alone. It should be said that the overlap calculation on its own is symmetric, as the methods page explains, so the direction in the scores comes from how the comparison is framed rather than from the arithmetic of overlap. Measures that build the asymmetry into the overlap itself exist in labour economics, and the methods page describes them as a possible refinement.

The collection also looked beyond roles with governance in their titles to roles of similar seniority in adjacent fields, among them privacy and internal audit, technology risk in banks, management consulting, public sector programme work, and platform trust and safety. In the current corpus several of these adjacent roles score as highly as roles explicitly concerned with AI governance. We take this as suggestive rather than settled. Each individual score rests on a single posting, and the more defensible observation is the general one, that in this sample the resemblance of titles and the resemblance of demands often diverge.

The near side of the map holds 155 entry roles across Singapore, Malaysia, the Philippines, Thailand, Indonesia, and Vietnam, and the far side holds 129 destination roles grouped into five families. Of the 284 postings, 234 are live advertisements collected over two days in October 2026, each linked to its source. The remainder are archetypes, composite descriptions each modelled on a real posting.

With the field and the direction of travel defined, the comparison becomes a question of measurement, and the next section describes what the reading counted.

03What We Counted

This section explains how the postings were read, what the 26 abilities are, and why both sides of the map use the same list.

one postingtwenty six abilities
Plate III. Each requirement line is read into a main ability, shaded deep, and up to two supporting ones, shaded light.

A job posting can be read as a list of demands. Behind a line such as supports the development of AI governance frameworks lies an ability, and behind the next line lies another. We read each posting line by line and coded 2,357 lines, assigning each a main ability and, where the line plainly carries more, up to two supporting abilities at a lower weight. The coding is slow and involves judgement at the margins, but it has a particular advantage. Once every role is expressed in the same list of abilities, two roles can be compared directly, and the comparison, though it inherits the coder’s judgements, no longer depends on the impression a title makes.

The abilities come from a list of 26, assembled from how international frameworks such as the NIST AI Risk Management Framework, the OECD AI Principles, and the EU AI Act describe this kind of work. They fall into six groups, which differ considerably in character, from analytical work to delivery to work with people.

GroupAbilitiesWhat it covers
Governing and policy thinking7 How the rules for AI are made, applied and checked, from risk frameworks to standards to a working understanding of the technology itself.
Analysis and research5 Studying problems with care, from policy research and writing to quantitative analysis to monitoring what is coming next.
Stakeholders and engagement4 Working with governments, firms, and communities, including across Southeast Asian languages, institutions, and regional bodies.
Delivery and operations5 Making things actually happen, programmes and projects, budgets and vendors, events, grants, and internal systems.
Communications4 Explaining and persuading, from communication strategy to public explainers to editing and publishing.
People and management1 Supervising, mentoring, and growing the people who do all of the above.

Both sides of the map are coded with the same 26 abilities under the same rules, so a role in Bangkok and a role in London can be set side by side ability by ability. Every coded line also retains its source text, so a surprising number can be traced to the sentence that produced it.

With both sides expressed in the same terms, the next section builds the score in two steps, beginning with how much of a destination’s demand a role appears to supply.

04How the Scoring Works

This section builds the score in two steps. Match strength estimates how much of a destination’s demand a role appears to supply, and reachability estimates whether a new graduate could enter the role and remain in it long enough to benefit.

entryrunwayportabilityadjacency
Plate IV. The four doors of the role used in the worked example. Entry is half open and the others fully open, giving a reachability of 0.88.

The first measurement can be pictured as two overlapping circles. One contains what an entry role appears to build, drawn from its coded requirement lines, and the other contains what a destination family asks for, drawn from the coded lines of every posting in that family. Match strength is the overlap between them. Abilities the role builds but the family does not request contribute nothing toward that family, and demands the role does not address are likewise left uncounted, so the measure credits only what the two have in common.

the overlap is the match what the starting role builds what the destination asks for
Figure 1. Match strength is the shared part of the two circles, counted ability by ability. What the role builds but the family never asks for is ignored, and so is what the family wants but the role never supplies.

An example makes the arithmetic concrete. We use the AI Governance Analyst role at Mass Markets, a digital consumer group in Vietnam, because it currently has the highest score in the corpus and so shows each step at work. It serves here as an illustration of the method rather than as a finding, since its position rests on a single posting and could change with the next round of collection. The posting is described by ten coded requirement lines, one of which reads “Help identify and manage risks associated with AI technologies, conducting AI risk assessments and impact analyses”. The family it feeds most strongly is Standards, Assurance & Compliance, whose heaviest demands are AI risk frameworks, at about 24 percent of everything the family asks for, then standards and conformity assessment at about 15 percent, and AI policy analysis at about 7 percent.

Set against the family’s demands, the ten lines supply about 0.73 of what the family asks for. The remaining 0.27 can be read as a rough syllabus, the abilities a holder of the role would still need to develop or demonstrate. Match strength says nothing, however, about whether a graduate could obtain the role at all, which is the concern of the second measurement.

Reachability poses four questions about the role and the path beyond it. Each is scored 0 to 2, and the four scores are averaged into a value between 0 and 1. Because these four judgements enter every score on the map, each is defined below.

A1Entry doorHow open the posting is to someone new. A door scores 2 when the posting welcomes fresh graduates or asks for no prior experience, 1 when it expects a couple of relevant years, and 0 when it is drawn for senior or heavily specialised applicants.
A2Career runwayWhether the pay lets you stay in the role long enough for the stone to carry you anywhere. A role can build excellent abilities and still be a poor first job if it pays below the cost of staying in the city, because most people cannot afford to linger there.
A3Skill portabilityWhether what you build there travels. Abilities, credentials, and demonstrated work that other employers, sectors, and countries recognise raise this door, while routines that only make sense inside one organisation lower it.
A4Domain adjacencyHow close the role already sits to the governance field. Working alongside the regulators, programmes, and research networks that this field relies on, including through the remote-first ways the field itself operates, raises the door, while unrelated territory lowers it.

The example role scores 0.88. Its entry door scores 1 rather than 2, since the posting asks for around three years of experience, and the other three doors score at the maximum.

The Whole Calculation, in Four Steps

  1. Step 1, read the role. AI Governance Analyst at Mass Markets in Vietnam. Ten requirement lines, each coded into abilities. Another line reads “Support compliance reviews and regulatory assessments; monitor emerging AI regulations and industry standards”.
  2. Step 2, measure the overlap. Against the Standards, Assurance & Compliance family the role supplies 0.73 of what the family asks for.
  3. Step 3, test reachability. Entry door 1 of 2, career runway 2 of 2, skill portability 2 of 2, domain adjacency 2 of 2, which averages to 0.88.
  4. Step 4, combine. 0.73 × 0.88 × 100 lands near 64. The table reports 63.5 because it carries the unrounded digits through. The difference arises from rounding and alters no ranking.

Scores of 50 and above are labelled strong, 40 to 50 moderate, and lower scores emerging. The bands are deliberately coarse, because the method does not support fine distinctions between nearby scores. The score is best understood as a relative and directional estimate resting on three approximations, that postings stand in for jobs, that coded lines stand in for the abilities a job develops, and that the destinations as advertised stand in for the work as practised. It is not a probability of employment, and it has not yet been tested against the careers people actually follow, a limitation taken up in section 07 and on the methods page.

The score, then, consists of two estimates multiplied together. The next section describes how the map can be used to explore it.

05How to Use the Map

This section describes three ways of using the map, each of which takes only a moment.

SG, 44 live postings and 6 archetypesMY, 23 live postings and 4 archetypesID, 21 live postings and 4 archetypesPH, 14 live postings and 4 archetypesVN, 15 live postings and 3 archetypesTH, 13 live postings and 4 archetypesSG50MY27ID25PH18VN18TH17

Point at a bar to read its counts.

Plate V. Starting roles by market, live postings solid and archetypes hatched. The Philippine layer is the thinnest and is being expanded.
  1. 1Find yourself. Filter by your market and search for the kind of role you are weighing. Every live posting in the map is tied to one of thirty-four broader starting archetypes, so one employer’s vacancy stands in for a whole pattern of similar jobs across the market.
  2. 2Choose a direction. Activate a destination family and the entire ranking re-orders by overlap with that family alone. Which entry role looks strongest depends on the destination in view, which is why the map offers no single overall answer.
  3. 3Open the evidence. Every role card lists the requirement lines behind its numbers, the four reachability doors and, for live postings, a link to the original text. A decision is better made with the posting in view than with the number alone.
One filter deserves particular mention. Open to early applicants keeps only roles whose entry door is fully open, which we use as a proxy for roles open to recent graduates. Across the live corpus there are 73 such postings, 55 of them entry roles. For a reader applying directly from a degree, this is likely to be the most useful setting.
The last section of the reading turns to the destinations themselves and to what each of the five families asks for.

06What the Destinations Want

The five destination families below are built from 129 hub postings, 25 archetypes, and 104 live roles collected from San Francisco, Washington, London, Brussels, Oxford, Singapore, and boards that hire remotely from anywhere.

F2 Standards, Assurance and Compliance, 31 destination postingsF3 Program and Operations, 29 destination postingsF1 Policy and Governance Analysis, 25 destination postingsF5 Research Management and Operations, 24 destination postingsF4 Strategic Communications, 20 destination postingsStandardsF2   31ProgramsF3   29PolicyF1   25ResearchF5   24CommsF4   20

Point at a room to read its family.

Plate VI. Five rooms, each drawn to the number of destination postings in its family, 129 in all.

The destinations are grouped into five families because the work they describe takes several distinct forms, each asking for a different mix of abilities. The families describe kinds of work rather than degrees of concern with safety, and roles at safety-focused organisations appear in all five. Each family is summarised in a sentence below, using the labels the rankings and filters use.

F1Policy & Governance AnalysisThe people who think, brief, and advise on what the rules should be.
F2Standards, Assurance & ComplianceThe people who check that organisations actually follow the rules.
F3Program & OperationsThe people who make the work happen, on time and on budget.
F4Strategic CommunicationsThe people who explain the work to the world and win support for it.
F5Research Management & OpsThe people who keep researchers funded, organised, and moving.

Each card below shows the abilities the family asks for most, weighted across everything its postings demand, and the three strongest stepping stones our corpus offers that family. The profiles differ appreciably. F1 is led by policy analysis and research writing, F2 by risk frameworks and standards, F3 by operations and delivery, F4 by communications abilities altogether, and F5 by operations in service of research.

07Can You Trust It

This section reports two tests of the map’s internal stability, considers what those tests can and cannot establish, and notes where the evidence is thinnest.

L02 AI Governance & Testing Programme Officer, rank 3 then 3L08 Associate Risk Analyst, Technology Supervision, rank 4 then 4L10 Graduate Research Assistant, AI Governance Lab, rank 24 then 24L18 Programme Assistant, Digital Development, rank 10 then 10L19 Digital Infrastructure Policy Analyst, rank 16 then 16L20 Researcher, Economic & Policy Studies, rank 25 then 25L01 Tech Risk & AI Governance Associate, rank 6 then 7L03 Tech Policy Analyst, rank 1 then 2L07 Digital Economy Policy Officer, rank 12 then 13L11 Project Development Officer, ICT Bureau, rank 20 then 19L12 Digital Policy Consultant, rank 14 then 15L23 Research Officer, ICT & Society Policy, rank 13 then 12L25 Research Assistant, AI & Society Group, rank 22 then 23L04 Research Officer, Digital Politics & Policy, rank 18 then 20L13 Supervising Research Analyst, rank 15 then 17L14 Regulatory Affairs Associate, rank 23 then 21L15 AI Governance Programme Officer, rank 8 then 6L09 Research Analyst, rank 19 then 22L17 Fintech & AI Policy Analyst, rank 11 then 8L22 Community & Programme Manager, rank 21 then 18L16 Researcher, AI Governance Center, rank 9 then 5L21 AI Ethics & Governance Junior Adviser, rank 5 then 1L24 Digital Policy Officer, rank 17 then 11L05 Public Policy Associate (SEA), rank 2 then 9L06 Programme Coordinator, rank 7 then 14pilotscaled up125

Point at a line to follow one archetype.

Plate VII. Ranks of the 25 original archetypes before and after the corpus grew twelvefold. Most lines run flat, rank steadiness is 0.917, and the three that moved five places or more are shown in ochre.

A map built from a small pilot invites the suspicion that its pattern is an accident of the sample. To test this, we expanded the corpus from 200 coded lines to 2,357, almost twelvefold, rebuilt every destination profile, and rescored the 25 original entry archetypes against the new profiles.

The ordering of those 25 archetypes correlates at 0.917 with the original, on a scale where 1.00 would indicate an identical order, and average overlap moved by less than five hundredths. Three archetypes changed their strongest destination, and all three had sat close to a family boundary from the outset. This suggests that the broad ordering is not an artefact of the pilot’s size, though it does not by itself show that the ordering is correct, and the test is not fully independent, since the same codebook and coding practice produced both corpora.

The second test asks whether the five families are distinguishable. Each live destination posting had its family label withheld and was assigned to the nearest family profile by its abilities alone. The assignment matched the original label 82 times in 100, against about 20 that chance would give, and the mismatches fell mostly between research operations and the neighbouring families of standards work and programme delivery, where real organisations also tend to combine functions.

Both tests concern consistency. They show that the map behaves stably as the sample grows and that its categories are reasonably coherent. Neither tests the premise on which the scores depend, that working in a high-scoring entry role makes a later move into a destination role more likely. Testing it would require evidence about the careers people actually follow, for instance the earlier roles of people now working in AI safety governance, and gathering such evidence seems to us the most important next step for the map.

The Three That Moved

IDStarting archetypeMovement

Anchored in Real Postings

The 25 entry archetypes are composites, and each was modelled on a real posting. The twelve listed below were collected while the pilot was being built. In the expanded corpus, 234 of the 284 postings are live advertisements, each linked to its source.

TagLive postingOrganisationMarketWhat it anchors
What might change these conclusions? A differently constructed ability list, a change in what destinations ask for, evidence from actual career paths, or fuller coverage of the markets reached only thinly. Thailand and the Philippines are the clearest cases of thin coverage. Thai-language boards yield fewer policy-adjacent entry roles, and junior postings in the Philippines often sit behind application portals that resist collection, so entry-level roles in both markets are probably underrepresented. The methods page records the full list of limitations.
The Data Room

The reading ends here, and everything below is the dataset itself, arranged four ways. The rankings list every starting role with its score, the grid draws the same result as a picture, the evidence section opens the requirement lines behind every number, and the sources table lists every destination posting. Each part opens with a curated slice and a button expands it to the full set, so the page stays readable while nothing stays hidden.

08The Rankings

The table lists every starting role with its Feeder score, strongest first. Choosing a destination family re-ranks the whole table by overlap with that family alone, because direction changes the answer. Clicking any column sorts by it, and stacking the filters above narrows the view. The table opens with the ten strongest roles, and the button below it unfolds the rest.

# ID Starting role Source Market Destination Best overlap Reach Band Feeder score

No roles match the current filter combination.

09The Full Grid

This is the same result drawn as a picture. Each row is a starting role, each column a destination family, and the darker the cell, the more of that destination the role builds. The amber ring marks each role’s strongest destination. Clicking a column header activates that destination across the whole map, and clicking any cell opens that role’s evidence card below. Rows dim when a market filter is active. The grid opens with ten rows, and the button below it unfolds the rest.

10The Evidence Behind Each Role

Every number in this map rests on readable evidence. Opening any role shows the exact requirement lines we read, the abilities each line carries, the four reachability doors, and for live roles a link to the original posting so you can verify the reading yourself. The small chips name the abilities, and hovering a chip shows its plain definition. The cards open with the ten strongest roles, and the button below unfolds the rest.

No roles match the current filter combination.

11Every Destination Posting

This is the full supply side of the map, all 129 destination postings with their organisations and hubs, each linked to its source where one exists, and published so that anyone can retrace or challenge the reading. The table opens with the first ten rows, and the button below unfolds the rest.

IDFamilySourceRoleOrganisationHubLink