Method and limits

How the data is collected and classified, how the rules were checked by hand, and what changed when the method was corrected.

The question

What share of the jobs AI safety organizations are advertising right now could someone early in their career realistically apply for? "Early in their career" means two years of relevant full-time experience or less: a strong new graduate, or someone moving over from another field.

Where the data comes from

Roles are read straight from each organization's own hiring system rather than from a curated job board, which avoids the selection built into those boards. Every organization included publishes a machine-readable feed through Greenhouse, Lever or Ashby, so the whole collection runs unattended each week. The 30 September 2026 snapshot holds 141 concrete vacancies.

Organizations with no feed are left out rather than read by hand. Hand capture was tried for four of them (GovAI, Apart Research, Redwood Research, Palisade Research) and dropped: the rows went stale between captures, and leaving those organizations out moves the headline share by less than one percentage point. Every collection is also compared with the one before for organizations that fall to zero, because a hiring feed that has been switched off looks exactly like an organization that stopped hiring. That check caught the Center for AI Safety's move from Lever to Greenhouse in July 2026.

Standing "expression of interest" and "talent community" listings are kept in the data but left out of every vacancy count.

Which organizations count

An organization is in if AI safety, security or governance is its main purpose and it has a hiring system we can read. That includes nonprofits, one government body (the UK AI Security Institute) and four safety-focused companies (Apollo Research, Goodfire, Haize Labs, Irregular). The companies sell products and hire sales and marketing staff accordingly, so the findings page shows every figure with and without them. Anthropic is tracked as a frontier-lab comparison and never mixed in. The coverage page lists every organization considered, including the ones left out and why.

How roles are classified

Three published rules, all in classify.py, with a regression test for each fix in test_classify.py:

• Title markers. Whole-word markers sort titles into entry (intern, internship, fellowship, resident, residency, junior, entry level, entry-level, new grad, graduate, trainee, apprentice, early career, early-career, scholar, all levels, any level, all experience levels, "associate", and fellowship programs), senior (senior, staff, principal, lead, head of, director, vice president, chief, founding, engineering manager, expert, distinguished, partner), or unspecified. A title that names two rungs ("Researcher / Senior Researcher") is not treated as senior-only. "Member of Technical Staff" is exempt from "staff", and a bare "Manager" in a function title ("Product Manager", "Office Manager") is not read as a rung.
• Open-door statements. A short list of plain phrases in the posting body that invite below-senior applicants: "open to hires at junior…", "at all experience levels", "we don't require a formal background", "an offer at the Associate level first". Phrases that describe juniors the new hire would manage are deliberately excluded.
• Stated experience floor. The largest "N years" requirement in the posting, since listed requirements are normally all required at once. The parser skips non-experience uses of "years" (visa residency, post-employment bans, contract lengths). 43 of the 141 vacancies state a number.

A role reads as open if its title carries an entry marker, or (with no senior marker) its text states an open door or a floor of two years or less. It reads as needs more experience if the title is senior or the floor is above two years, and as doesn't say otherwise. Kind of work (research, engineering, operations and so on) comes from title keywords, with the hiring system's department as a fallback.

Checking the rules by hand

Rules miss things, so every one of the 141 vacancies in the 30 September 2026 snapshot was also read in full and labelled by hand, with the line that decided each label recorded next to it (download the labels). The labels mean:

• Open to early-career: the posting's own text makes it reachable with two years' experience or less.
• Needs more experience: a floor above two years, a senior or leadership scope, or a required PhD or publication record. A conditional exception ("earlier-career applicants with a standout record may apply") does not make a role open.
• Doesn't say: the posting gives no usable signal.

The findings page reports these hand labels. The rules are what keep the weekly figures and the role list up to date between audits, so it matters how well they match:

The rules agree with the hand labels on 106 of 141 vacancies. Of the 24 they call open, 22 are open by hand; of the 25 open by hand, they find 22. Two caveats. The open-door phrases were written after reading this snapshot, so this agreement is partly by construction, and the real test is the next audit. And the audit caught mistakes in both directions: one hand label was corrected after the rules found a requirement the first reading missed.

The rules' main weakness is on the other side. 25 roles they cannot place are, on reading, clearly closed: a required PhD, "several years" written in words, a leadership scope described rather than titled. So the rules overstate "doesn't say" and understate "needs more experience", while their count of open roles stays close. For comparison, titles alone flag 4% of vacancies as open, about four times fewer than the hand-checked 18%.

What to keep in mind

• Postings are not need. Senior and network hiring often goes unadvertised, so this undercounts senior demand and should not be read as where the field most needs people.
• Fellowships are the main way in, and mostly invisible here. MATS, SPAR, GovAI's seasonal fellowships and similar programs recruit on their own cycles. A field-wide picture of entry routes would need to count them.
• Small numbers. 141 vacancies is the whole population for these organizations, not a sample, but per-organization and per-function figures rest on a handful of roles and move a lot week to week.
• One reader. The hand labels are one careful reading. Where you disagree, the evidence column shows exactly what the label rests on.

Corrections

This project publishes its mistakes. Every figure on the site is re-computed from the raw snapshots with the current rules, so older figures quoted elsewhere may differ.

• 30 September 2026, method overhaul.
  • Hand audit added. The old headline put the open share at about 5%, counting title keywords. Reading every posting gives 18%, because most open roles say so in the body rather than the title.
  • Markers now match whole words. Substring matching had read "International" as "intern", so "Director, US International Tax" counted as an early-career role in the Anthropic comparison.
  • The experience floor now takes the largest stated requirement, not the smallest. "10+ years… 2+ years of people management" had read as a two-year floor.
  • Lever postings are now stored in full. Before, only the opening paragraph was kept, leaving requirements unreadable for METR, Epoch AI, Apollo Research and FAR.AI.
  • Hand capture dropped. Rows for the four organizations without a hiring feed were carried forward between captures, and closed GovAI roles, including the Research Scholar position the old headline counted as a way in, had stayed in the count for two months. Those organizations are now left out; with them, the hand-checked open share was 18.3%, without them 18%.
  • Seven organizations added (FAR.AI, LawZero, Kairos, AVERI, CAIS Action Fund, Haize Labs, Irregular), and each organization is now typed as nonprofit, government or company.
  • Two-rung titles, bare "Manager" titles and "Associate" titles are handled as described above.
• 26 July 2026. The Center for AI Safety's move from Lever to Greenhouse was caught and fixed; "Fellow" was split into program fellowships (entry) and think-tank Research Fellow roles (not entry).