The compliance layer of the AI hiring stack
Snapshot 2026-08-19 · 22,102 listings from 36 sources · CC BY 4.0
Of 6,349 live remote listings, 77.2% state no eligibility rule at all. Among the 1,445 that do, 59.8% (Wilson 95% CI 57.2–62.3) exclude candidates outside a named region. Listings tagged "worldwide" are barely different: 58.1% of those that state a rule still restrict it.
Why the number is a range, not one figure
A restriction can be written by the employer or imposed by the feed. Adzuna boards, for example, are single-country datasets, so a listing there is closed to outsiders whether or not the ad says so. We report all three counts rather than picking the most quotable one.
| Definition | Restricted | n | 95% CI |
|---|---|---|---|
| Listing text only (pre-registered primary) | 59.8% | 1,445 | 57.2–62.3 |
| Listing text, excluding single-country feeds | 69% | 1,036 | 66.1–71.8 |
| Text plus feed-level country scope (upper bound) | 78.8% | 2,246 | 77.1–80.4 |
Where the restrictions point
Half of all restricted listings require one country. Meanwhile only 2.9% of remote listings name a time-zone requirement at all, so the constraint is jurisdictional rather than operational.
Advertised pay
46.2% of remote listings advertise pay. The median among them is $170,000 a year, skewed toward US technology roles. Split by eligibility: restricted listings $173,502 (n=465) against open listings $163,000 (n=229) — too close to claim as a real difference. The pay hypothesis from the July study did not reproduce.
Openness looks like a company property (exploratory)
The pre-registered test required 20 or more stated-rule listings per employer; only six clear that bar, so it is underpowered and we report it as a failure. At a threshold of 10 listings — exploratory, not pre-registered — 16 employers qualify, and 68.8% of them are open on one in ten or fewer of their listings.
| Employer | Open rate | n |
|---|---|---|
| Multiverse | 100% | 14 |
| Supabase | 87.5% | 16 |
| Bjak | 65.2% | 23 |
| ElevenLabs | 27.8% | 18 |
| EverAI | 26.7% | 30 |
| 1Password | 0% | 29 |
| OpenAI | 0% | 61 |
| ButterflyMX | 0% | 13 |
| UiPath | 0% | 15 |
| Tenstorrent | 0% | 14 |
Method
- Single cross-sectional read from production D1. No time-trend claims.
- Labels re-derived from current listing text, never read from the stored column.
- Classifier validated first on 120 hand-labelled rows: 90.9% precision and 87% recall on restricted listings, 100% on open ones.
- Re-derived labels agree with the production labels on 86.4% of rows; the gap is reported rather than hidden.
- Listings stating no eligibility rule are excluded from denominators rather than counted as open. Onsite roles are excluded.
Limitations we state up front: silence is not openness; text is not law (an employer can write “open to anyone” and still be unable to run payroll in your country); and one index is not the global labour market — our sources skew toward English-language technology employers.
Aggregate dataset (CC BY 4.0)
Every figure on this page, with its sample size and confidence interval.
Download CSVThe write-up: The AI hiring stack breaks at the border