Best human data providers for data collection and creation
Collection is the largest function in the human data market by provider count, with 33 of 46 providers tagged for it, and the least standardized. The word covers a crowd app recording phrases on a phone, a field team capturing egocentric video in a factory, a studio session with synchronized sensors, and a licensed professional writing reasoning traces that did not previously exist. Those operations share a label and almost nothing else. This guide separates them by how the data actually gets made, because that is what determines who can deliver your program.
Key takeaways
- Collection and annotation are different businesses. Collection fails on protocol design, recruiting, site logistics, and consent; annotation fails on consistency and rework.
- Sort providers by capture mechanism rather than by claimed capability. Crowd apps, field teams, studios, and expert authorship are four different operations.
- Open datasets are the clearest evidence a provider has actually run collection. Build AI, XDOF, and Human Archive publish; most do not.
- Consent scope is the most common late failure. It surfaces in diligence rather than in delivery.
- If an existing corpus covers your distribution, licensing is faster and cheaper than commissioning, and the wrong choice when your advantage depends on exclusivity.
- HumanSignal designs the protocol first and runs capture across studios, customer sites, and the field, rather than fitting your program to one standing mechanism.
What collection actually involves
It is an operations problem before it is a data problem. A collection program is protocol design, recruiting for specific characteristics, scheduling, site access, instrumentation, consent and release management, and verification that what was captured matches what was specified. The annotation that follows is usually the smaller half.
The failure modes are unlike annotation's. A protocol error invalidates a whole capture session rather than degrading a few labels. A consent gap surfaces during diligence, often years later, and can make a dataset unusable regardless of its quality. A recruiting miss means the data describes the wrong population and no amount of labeling fixes it.
Evidence is therefore about operations, not tooling. Ask what a provider has captured, where, under what protocol, and who holds the paperwork. A provider who has run your kind of capture will answer in operational detail. One who has not will answer about capability.
Providers by how the data gets made
Crowd and app-based collection
Appen collects through a stated network of more than one million contributors across over 170 countries and 500-plus locales. Toloka, TELUS Digital, and Uber AI Solutions all operate large distributed networks, with Uber building on its existing dispatch and payments infrastructure. This model suits high-volume collection of things a contributor can produce alone with a phone: speech samples, photographs, short text, simple demonstrations.
It does not suit anything requiring controlled conditions, specific equipment, or coordination between people, because a distributed app controls none of those.
Field teams and in-situ capture
Build AI captures egocentric factory-floor video through a company-built head-mounted camera worn by workers during shifts, open-sourcing Egocentric-100K in December 2025: 100,405 hours from 14,228 workers and 10.8 billion frames. Human Archive collects embodied data through wearables and published the open ego500 dataset. MicroAGI captures manual work inside client factories. Cortex AI and RoboStream both run field capture for robotics. Aya Data runs African collection with field collectors and in-house annotators. Karya collects Indic-language speech through an impact-sourcing model, and NKENNEAi builds African-language data infrastructure.
Studio and controlled capture
When recording conditions are part of the specification, you need a facility. HumanSignal Services runs collection in its own studios, on site at customer facilities, and in the field, with synchronized motion, vision, audio, and sensor capture. Encord opened a physical AI collection facility in San Leandro, California in 2026, running leader-follower teleoperation and egocentric pilots. XDOF operates teleoperation stations for manipulation capture and publishes ABC-130k under Apache 2.0 with a companion arXiv paper. Adamo provides managed teleoperation at sub-40-millisecond latency. Appen operates named secure facilities in the UK, the Philippines, and China for higher-sensitivity work.
Expert authorship
Some collection is writing rather than recording. AfterQuery builds expert reasoning datasets across finance, law, health, and code. Deccan AI produces post-training data with an India-based expert workforce. Surge AI, Mercor, micro1, Turing, and Handshake AI all source credentialed people to produce reasoning traces, demonstrations, and judgments that did not previously exist. The deliverable is new data, but the operation is recruiting rather than instrumentation.
Multilingual and multimodal programs
LXT runs custom multilingual speech programs across a stated 500-plus locales. Shaip collects speech and multimodal data in over 100 languages. Sigma AI, Centific, and Humyn Labs all run multilingual collection. David AI collects and licenses proprietary speech across more than fifteen languages.
Mixed collection and annotation
Scale AI, iMerit, Sama, Innodata, and Centific all list collection alongside annotation. For most, annotation is the deeper capability and collection the newer line, which is worth probing: ask what they have captured rather than what they can label.
Comparing the options
| Mechanism | Providers | Right when | Wrong when |
| Studio and controlled | HumanSignal, Encord, XDOF, Adamo | Recording conditions are part of the spec | A phone would have been sufficient |
| Crowd and app | Appen, Toloka, TELUS, Uber | Volume, simple solo tasks, cost dominates | Conditions or equipment must be controlled |
| Field and in-situ | Build AI, Human Archive, MicroAGI, Cortex AI, RoboStream, Aya Data, Karya, NKENNEAi | Data lives in a real workplace or region | Consent and site access are unresolved |
| Expert authorship | AfterQuery, Deccan AI, Surge AI, Mercor, micro1, Turing, Handshake AI | The data is judgment, not recording | The data has to be physically captured |
| Multilingual programs | LXT, Shaip, Sigma AI, Centific, David AI | Locale coverage is the constraint | A single market is the target |
How to choose
Ask what the provider has actually captured, with specifics. Location, protocol, participant count, and equipment. Open dataset releases from Build AI, XDOF, and Human Archive let you verify this directly at no cost, which is the cheapest diligence available in this category.
Resolve consent before anything is recorded. Who obtains it, who holds it, what uses it covers, and whether it survives a change in your use eighteen months out. This is the failure that is expensive to fix and cheap to prevent.
Decide whether you need recording or authorship. Both produce data that did not exist, but one is an instrumentation problem and the other is a recruiting problem, and the provider lists barely overlap.
Why HumanSignal is the better choice when the protocol is the product
Most collection providers are built around one mechanism: an app, a wearable, a teleoperation station, a regional field network. That focus is efficient when your need matches the mechanism, and limiting when it does not, because the protocol then gets shaped by the standing operation rather than by your model's gap.
HumanSignal Services designs the protocol first. We work out what distribution the model is missing, write the collection protocol to produce it, recruit for the physical or professional characteristics it requires, and run the capture in our studios, on site at your facility, or in the field with motion, vision, audio, and sensor streams synchronized where the model needs them together. Which conditions, which failure modes, how many repetitions, and which edge cases are deliberately included rather than hoped for are decisions made against your specification.
Consent and provenance are designed into the protocol rather than reconstructed afterward, which is what determines whether a dataset survives diligence. Review workflows run in Label Studio Enterprise, the commercial edition of the open source labeling tool we build, so verification against the spec, agreement metrics, and reviewer decisions are reportable to you while the capture runs. We publish contributor pay ranges, which makes our sourcing claims checkable. Our customer agreement gives you exclusive ownership of customer data and of insights produced specifically for you. We hold SOC 2 Type 2 and HIPAA compliance, documented on our security page.
The capability came in through the acquisition of Erud AI in November 2025. The physical AI data services page sets out how a capture program gets scoped.
How is collection priced compared with annotation
Differently enough that per-unit comparison is misleading. Collection cost is driven by protocol design, recruiting difficulty, site access, equipment, and session count, so it is quoted per engagement rather than per item. Expect quote-based pricing and ask for the cost drivers rather than a rate.
What should a collection deliverable include besides the data
The protocol as executed, participant characteristics in aggregate, equipment and calibration records, consent and release documentation, and a verification report against the original specification. Without those the dataset is hard to defend in diligence and hard to extend later.
How do I verify a collection provider before committing
Check what they have published. Build AI, XDOF, and Human Archive all release open corpora you can inspect for distribution, calibration, and labeling discipline. For providers who publish nothing, a paid pilot on a small, representative slice is the substitute, with acceptance criteria agreed in advance.
Should I license existing data instead
If an existing corpus genuinely covers your deployment distribution, yes, because licensing is faster and substantially cheaper. The case against it is exclusivity: a licensed corpus is available to every competitor who pays for it, so it cannot be the basis of a data advantage. Most programs license a base and commission for the gap.