TECHNICAL AI SAFETY // FIELD MAP
← field map
UNIT 6 · START CONTRIBUTINGchapter 6 · reading

Next steps: Policy fellowships

BlueDot Impact · Technical AI Safety · unit 6, chapter 6
Course pageNext steps: Policy fellowships — BlueDot ↗
TL;DR — If the lever you want to pull is regulation rather than a loss function, the standard on-ramp is a stipended three-to-six-month fellowship that puts you in the rooms where things get decided. The course names five: GovAI (research, London/DC), IAPS (US and compute policy, DC-or-remote), Horizon (a real placement into the US executive branch or Congress), Talos (the Brussels equivalent), and BlueDot's governance course as the zero-cost first step. The binding constraint is not your ML background — it is citizenship, work authorisation, and clearance eligibility. Sort by jurisdiction first, topic second.

Unit 6 has spent five chapters assuming you will contribute by building something — an eval, an interp result, a red-team finding. But much of what determines whether frontier systems ship safely is decided by people writing rules rather than code, and those rooms are short of anyone who can read a model card without a translator. What follows is a directory, not an argument: six routes into a policy career that will take you seriously because of the technical work, not in spite of it.

Why the technical background is the asset, not the detour

Policy questions are usually empirical underneath. Does a compute threshold at 1026 FLOP still catch the systems we care about? Is a reported eval result a measurement or a marketing artifact? Can a safety commitment be verified from outside the lab? A generalist analyst learns to ask those; someone who has trained a model and written an eval harness can answer them, and can tell a real technical objection from a rhetorical one. The catch is that the output format is memos with a recommendation on top — so these programs are largely a supervised environment for turning engineering fluency into writing.

Four different things all called "fellowship"

Confusing them wastes applications:

Undecided? Take the cheapest move — governance course, then a directory scan — not a competitive application.

The eligibility filter people discover too late

Every one of these is bounded by a passport. Horizon needs US work authorisation and DC residence, sponsors no visas, and its clearance-requiring placements are closed to non-citizens. Talos's preference is structural, not snobbish:

"We have a strong preference for candidates with EU citizenship since many of the most important roles in EU institutions are only open to EU citizens."— Talos Fellowship, Talos Network

IAPS is the loosest on this axis — a remote option open worldwide around a mandatory two-week DC residency — making it the most accessible entry point from outside the US or EU. Sort by the passport you hold before the research agenda you like; a perfect topical fit you are ineligible for is worth nothing. The related surprise runs the other way, and IAPS states it plainly:

"The Fellowship is designed to train professionals from a variety of backgrounds and no technical expertise is expected."— AI Policy Fellowship FAQ, IAPS

Read that as calibration, not permission to skip the technical work: you compete not on ML depth but on demonstrated interest in a specific policy question and on the quality of your writing.

Timing quietly disqualifies people too

These run annual or semi-annual cycles with long lead times — several close nearly a year before the work starts. Horizon's 2027 cohort closed in August 2026, selected late 2026, training from January 2027: eighteen months from "I should look into this" to a paycheck. Talos runs spring and autumn cohorts; GovAI's winter fellowships close the preceding autumn. Reading this in the wrong month, the plan is: governance course now, one published piece of writing, apply next cycle with something to point at.

The one thing to carry away: write one public, technically-grounded policy artifact before you apply. A 1,500-word memo taking a concrete position — why a compute threshold is or isn't the right trigger, what a third party would need to verify a safety case, where an eval standard measures the wrong thing — does more than another line of ML experience. It is direct evidence of the skill these programs select for, and the cheapest test of whether you enjoy the work.

Readings, linked

The course gives no time budget here — this is a directory chapter, and the intended time cost is however long it takes to shortlist. Start with the BlueDot governance course if you are undecided, and with the fellowship matching your jurisdiction if you are not.

Exercises

This chapter ships no exercises — it is a resource list, and the course expects you to act on it rather than write about it. Two field map extras, because a directory you skim is a directory you forget:

  1. The eligibility-first shortlist (field map extra) — Before reading a single research agenda, build a five-row table: one row per program in the readings, columns for work authorisation required, physical presence required, next deadline, stipend, output you would produce. Fill it from the linked pages only — no guessing. Then strike every row you are ineligible for and rank what remains by whether you would rather write a paper or hold a job. What a good answer has: at most two surviving rows, a named next deadline for each with the year attached, and one sentence on what you would do in the twelve months before that deadline. The exercise fails if it produces a list of five things you are "interested in" — the point is elimination, not enthusiasm.
  2. Deadline tracker over the fellowship database code (field map extra) — Turn the 70+ row Airtable directory into something that pings you instead of something you forget. What a good answer has: a script that produces a dated, sorted shortlist filtered to programs you are actually eligible for, plus an .ics file with a reminder six weeks before each deadline. Start here: (1) open the shared Airtable view in a browser and export the visible table to CSV — the share link is a client-rendered SPA, so scraping it headlessly is more work than the export button; (2) load it with pandas.read_csv and normalise the deadline column with dateutil.parser, keeping a parse_failed flag rather than dropping bad rows; (3) write an eligibility predicate as an explicit function of your citizenship and location, and log every row it rejects so you can eyeball the false negatives; (4) generate the calendar with the ics package (pip install ics pandas python-dateutil); (5) as a stretch, add a weekly requests + BeautifulSoup check on each program's page that diffs the text against a stored copy and alerts on change — most of these sites announce a new cycle by editing one paragraph, with no feed and no email.

Go deeper

Next: Next steps: Other fellowships · Back to the map.