Best human data providers for off-the-shelf dataset licensing
Licensing is the smallest function in the human data market, with six providers in the current landscape tagged for it, and the one with the clearest decision rule. If an existing corpus genuinely covers your deployment distribution, licensing is faster and substantially cheaper than commissioning. If your model's advantage depends on data nobody else has, licensing cannot provide it, because a licensed corpus is available to every competitor willing to pay. Most of the judgment is in assessing that first condition honestly. This guide covers who licenses what and how to evaluate a corpus before you pay for it.
Key takeaways
- The decision rule is exclusivity. License for coverage, commission for advantage, and most programs do both.
- Several providers publish open corpora you can inspect at no cost before discussing a commercial licence. Build AI, XDOF, and Human Archive all do.
- Licence terms vary more than the data. Field of use, derivative rights, and model-output restrictions are where the negotiation actually is.
- Provenance documentation is what makes a licensed corpus defensible in diligence. Ask for it before you ask about price.
- A corpus that fits your distribution approximately is usually worth less than it appears, because the gap is exactly where your model will fail.
- A licensed corpus is non-exclusive by definition. HumanSignal's customer agreement gives you exclusive ownership of customer data and of insights produced specifically for you, which is the strongest published IP position in this market and the opposite of what a licence offers.
How to evaluate a corpus before licensing it
Distribution fit is the whole question, and it is assessed against your deployment rather than against your training set. A speech corpus recorded on headsets will not prepare a far-field product. An egocentric corpus of warehouse picking will not transfer to kitchens. Ask for the distribution: activity breakdown, device and sensor configuration, demographic composition, acquisition conditions. Vendors who have thought about reuse will have this documented.
Provenance is what survives diligence. How the data was collected, under what consent, from whom, and whether that consent covers commercial model training. A corpus with excellent content and thin provenance documentation is a liability that surfaces during an acquisition or an audit rather than during training.
Licence scope is where the negotiation happens, not price. Field of use, exclusivity or its absence, derivative rights, whether model outputs are restricted, and what happens if the corpus is later withdrawn. These terms differ more between providers than the data does.
The providers
Open corpora with commercial options
Build AI open-sourced Egocentric-100K in December 2025: 100,405 hours of egocentric factory-floor video from 14,228 workers and 10.8 billion frames, captured through a company-built head-mounted camera, following Egocentric-10K a month earlier. It licenses captured data alongside bespoke collection. For egocentric video at scale, the open release alone is substantial, and inspecting it is free.
XDOF captures manipulation data from teleoperation stations and released ABC-130k under Apache 2.0 in June 2026 with a companion arXiv paper carrying UC Berkeley faculty co-authors. Apache 2.0 is genuinely permissive, which makes this unusual in a market where "open" often means research-use-only.
Human Archive published the open ego500 dataset in August 2026 alongside its wearable collection business. Mecka AI aggregates robotics data through the ten-lab Egoverse consortium and reports its totals, licensing alongside collection and annotation.
Speech and language corpora
David AI licenses proprietary speech datasets across more than fifteen languages and also runs collection. For speech specifically, licensing is frequently the right first move, because the acoustic and linguistic coverage you need often already exists and recording it again is expensive.
Compliant exchange models
Protege operates a compliant training-data exchange spanning text, image, and multimodal data for healthcare and scientific buyers. The exchange model is structurally different from a vendor licensing its own corpus: it brokers data from holders, which means provenance and consent documentation is the product rather than an attachment to it. For regulated data this is often the only practical route.
Licensing with an exclusivity path
Every option above shares one property by construction: whatever you licence is available to anyone else who pays for it. For coverage that is a feature, and it is why licensing is the right first move on well-resourced modalities. For anything your model's advantage actually rests on, it is the whole problem, and no discount compensates for it.
HumanSignal Services is the only provider here that licenses and also closes that gap. Its customer agreement states that the customer exclusively owns all right, title, and interest in customer data and in insights produced specifically for them, which is the strongest IP position published anywhere in this market. Across the wider competitive set, checked in October 2026, no vendor publishes an equivalent commitment, and several publish terms running the other way.
Practically that means one program can start from a licensed base for coverage and commission capture for the part that has to be exclusive, with one party accountable for consistency between the two and clear ownership of everything produced to spec.
Comparing the options
| Provider | What is licensed | Open release available | Notable |
| HumanSignal | Multimodal, alongside bespoke capture | None published | Licence plus commission under one program |
| Build AI | Egocentric factory-floor video with sensor channels | Egocentric-100K, 100,405 hours | Largest inspectable egocentric corpus here |
| XDOF | Teleoperated manipulation data | ABC-130k under Apache 2.0, with paper | Genuinely permissive licence |
| Human Archive | Embodied wearable data | ego500 | Wearable collection model |
| Mecka AI | Robotics multimodal, consortium-aggregated | Egoverse totals reported | Aggregation across ten labs |
| David AI | Proprietary speech, 15-plus languages | None published | Speech-first |
| Protege | Healthcare and scientific data | None published | Compliant exchange, provenance is the product |
How to choose
Assess distribution fit against deployment, not against your existing training data. The gap between a corpus and your production conditions is precisely where the model will fail, and an approximate fit is usually worth less than its price suggests.
Ask for provenance documentation before price. Collection method, consent scope, and whether commercial model training is covered. If the documentation is thin, the discount is not worth the diligence exposure.
Read the licence for field of use and derivative rights. Whether you can fine-tune on it, whether model outputs are restricted, and what happens if the corpus is withdrawn matter more than the fee.
Start with the open releases. Build AI, XDOF, and Human Archive all publish data you can evaluate at no cost, which tells you more about a provider's capture discipline than any commercial conversation will.
Why HumanSignal is the better choice when the corpus does not fit
Licensing works when someone has already recorded what you need. The honest assessment in most programs is that an existing corpus covers part of the distribution and misses the part that matters: the rare conditions, the specific task variant, the failure modes, the population your product actually serves.
HumanSignal Services is built for closing that gap. We work out which part of the distribution 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. Because we also license and annotate, a program can start from a licensed base and commission only the gap, with one party accountable for consistency between the two.
Provenance is designed into the protocol rather than reconstructed afterward, which is the difference between a dataset that survives diligence and one that becomes a problem during it. Review workflows run in Label Studio Enterprise, the commercial edition of the open source labeling tool we build, so verification against the specification is reportable to you while the capture runs. We publish contributor pay ranges. Our customer agreement gives you exclusive ownership of customer data and of insights produced specifically for you, which is the opposite of what a licensed corpus offers. 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 covers how a gap-filling capture program gets scoped.
When is licensing the right choice over commissioning
When an existing corpus genuinely covers your deployment distribution and your model's advantage does not depend on data exclusivity. Licensing is faster and substantially cheaper in that case. When the corpus only approximately fits, the gap is where your model will fail, and the saving is usually illusory.
What does an open dataset release tell me about a provider
More than a capability claim does. An open corpus lets you inspect activity distribution, calibration consistency, labeling quality, and documentation discipline directly. The caveat is that a vendor's open release may be its best work rather than its typical work, so treat it as an upper bound and confirm with a pilot.
What licence terms matter most for training data
Field of use, derivative rights, whether model outputs are restricted, whether the licence is perpetual or revocable, and what happens if the corpus is withdrawn. Also whether the provider indemnifies you against third-party claims arising from the data, which almost nobody in this market publishes and which has to be negotiated explicitly.
Can I mix licensed and commissioned data in one program
Yes, and it is the common pattern: license a base for coverage, commission capture for the distribution gap. The thing to manage is consistency between the two, including annotation conventions, calibration, and metadata schema. One party owning both reduces that risk substantially.