How Long AI-Training Roles Stay Open

Most write-ups of the AI-training gig market - including our own Pay Index and Market Data pages - describe a single moment: what a role pays, right now. Neither answers a different, equally practical question: once a listing appears, how long does it actually stay open?

We can answer that because every role this site tracks carries two real timestamps - when it was first seen, and when it was last seen - going back to the corpus’s earliest scrape. This piece reads those two columns directly, with nothing estimated and nothing typed by hand.

The finding

The typical listing that has already closed stayed open for about a week. That is read live from the data below, not fixed at publish time - it will move as more listings close, which is the point of building this as a component instead of a sentence.

What this does and doesn’t show

It is a lower bound, not a survival estimate. The median above only counts roles that have already closed. A newly-tracked corpus like this one always has more young open roles than young closed ones, so the true median lifespan - once every currently-open role eventually closes too - is at least this high, likely higher. A proper survival-curve (Kaplan–Meier) estimate needs roughly 90 days of history to be meaningful; this corpus started being tracked in July 2026, so that estimate isn’t ready yet.

It assumes continuous listing. A role is treated as live for every day between its first and last sighting. If a listing disappears from a source and reappears later under the same identifier, our scrapers treat that as one continuous listing rather than two - so a real gap in posting would be invisible to this measurement.

It carries no pay figure, on purpose. Pay data is re-parsed in place every time the extraction logic improves, so a historical pay figure computed today would describe today’s parser applied to old listings - not what was actually advertised at the time. Lifecycle timing has no equivalent problem: once a role is marked closed, that date doesn’t change retroactively.

“New” and “closed” counts for the current week are provisional. The table above only shows weeks that have fully elapsed. A week still in progress will show fewer closures than it eventually will, simply because the week isn’t over yet.

Why this matters for anyone comparing platforms

A platform whose median advertised rate looks attractive is a different proposition if its listings also churn fastest - less certainty that a given posting will still be open by the time someone applies. Once enough history accumulates to break this down per platform rather than site-wide, that comparison is the natural next piece in this series.