First-party data, collected through our own network.

In AI training, first-party data is data the provider collected itself. The provider recruited the contributors, briefed them on the task, paid them, holds their signed releases, and ran the review that decided what ships. The opposite is third-party data: footage scraped from the web, bought from an aggregator, or relicensed from a collection someone else built for a different purpose.

The distinction matters because a training dataset inherits every gap in its history. If nobody can say who recorded a clip, what they agreed to, or why it was accepted, those questions do not disappear when the clip enters a training run. They resurface later, during a licensing review or a customer’s procurement process, when they are much harder to answer.

Firsthand runs a distributed contributor network with a web portal for assignment, upload, and review. A buyer’s written spec is turned into assignments; contributors are briefed and capture against it; uploads come back through the same portal and are reviewed before anything is accepted.

Contributors are briefed, paid hourly, and collected against the spec, with a specialized rig such as our synchronized egocentric hardware where the project needs it. People who want to capture for the network can apply on the for-collectors page, which publishes the release they would be signing.

Every contributor signs a participant release before capture. The release is published verbatim on the provenance page, together with the de-identification steps applied to every clip, so a buyer can read exactly what each person agreed to.

Contributors are paid the hourly rate stated in their schedule for all time spent in session, including setup, calibration, and breaks. Payment is not contingent on recordings being accepted, delivered, or licensed, and there is no per-task bounty and no rejection clawback. The current rates are listed on the provenance page.

Every recording is human-reviewed before delivery for task match, visible hands, and the absence of staged or idle footage. Uploads missing camera telemetry are rejected automatically, before a reviewer sees them.

If an item cannot meet the spec, or a bystander cannot be de-identified, it is dropped rather than shipped at a discount to pad the count. You see the reject reasons in a reject log, not just the accepted items.

Faces, licence plates, and screens are detected and irreversibly blurred at ingest, then confirmed by a reviewer before anything ships. Speech is scrubbed of names, addresses, and card numbers while ambient audio is retained.

If a person in frame cannot be de-identified, the clip is discarded rather than shipped. Files are organised in a human-readable, PII-free layout of project, category, activity, slot, and version.

You receive a perpetual, irrevocable, worldwide, buyer-owned licence that is non-exclusive by default, with a data card recording the jurisdictions of capture and the consent basis. Exclusivity, where a program needs it, is a separate term to negotiate.

Because the data was collected first-party, the licence is backed by releases Firsthand itself holds rather than by a chain of onward licences. These are the general terms stated on this site’s provenance and terms pages, not legal advice for a specific contract.

Where coverage for the requested tasks and regions is already deep, the median time from signed spec to first validated batch is a median of 48 hours. Where it is thin, turnaround runs three to four weeks while recruiting expands in parallel across more regions.

Most orders begin with a Spec Pilot: a small, fast first batch that proves the spec and accept criteria hold up on real footage before you commit to volume.