Best Public Data Providers for AI Training and Grounding

Most comparisons of AI data providers put labelling shops, dataset catalogues and licensed marketplaces in one table and rank them against each other. They solve different problems, so the ranking is meaningless before you know which problem you have.

The first split is the one in the title. Training data is a corpus you buy, ingest and train on, usually once or on a periodic refresh. Grounding data is retrieved live at inference time, so a model can answer using information it was never trained on. A provider who is excellent at one is frequently irrelevant to the other.

The second split is legal, and in 2026 it decides more purchases than quality does. This guide covers both: the provider categories, what a licence needs to say, what it costs, and how the 2 halves differ. It is not legal advice, and this area of law is moving fast enough that anything here should be checked against counsel before it drives a contract.

Why does licensing matter more than the dataset?

Because the courts have drawn a line, and it is not where most people expect it. A pattern has emerged across the major US cases: the act of training has repeatedly survived fair use scrutiny, while the act of acquiring the material has produced the largest payouts.

The numbers behind that pattern, verified against reporting as of mid 2026:

• Anthropic agreed a settlement of roughly $1.5 billion covering about 482,000 works. It received preliminary approval, with a final fairness hearing set for 23 April 2026.
• In Thomson Reuters v Ross Intelligence, decided in Delaware in February 2025, a US court rejected a fair use defence for AI training data for the first time. The ruling was explicitly limited to a non-generative system that competed directly with the source product, so it does not settle the question for generative models.
• In Kadrey v Meta, decided in Northern California in June 2025, a court reached a broadly similar fair use conclusion on book training claims.
• The US Copyright Office published a report in May 2025 concluding that certain uses of copyrighted material to train generative AI cannot be defended as fair use.
• Article 53(1)(d) of the EU AI Act requires developers of general purpose AI models to publish summaries of the content used for training.
• By early 2026, more than 90 cases had been filed against AI companies relating to training data.

Read together, the practical rule is simple. Training on licensed material is defensible. Training on material you cannot show you acquired legitimately is where the money goes. That makes provenance the primary product attribute, ahead of volume, freshness or price.

One distinction worth keeping straight, because vendors blur it. Lineage is the technical history of data inside your systems: transformations, versions, pipelines. Provenance is the rights history: who owned it, how it was acquired, and what permission came with it. Lineage tooling does not give you provenance.

What are the 5 categories of AI data provider?

Match your gap to the category before you compare vendors, or you will end up comparing a labelling shop with a marketplace and drawing conclusions from the difference.

These are different businesses with different risk profiles. A labelling shop transfers no acquisition risk to you, because you supplied the data. A licensed marketplace transfers acquisition risk away from you, because the licence names the rights holder. A dataset catalogue may transfer nothing at all, depending on how the catalogue was assembled, which is the question to ask first.

What should a training data licence actually say?

6 things. A vendor who hesitates on any of them is transferring their risk to you, and the whole point of buying licensed data is that the risk moves the other way.

1. Per-asset rights documentation. Not a blanket assurance covering a corpus, but a record you can trace to individual items.
2. A licence that names AI training explicitly. A general content licence written before 2023 very often does not grant training rights, and silence is not permission.
3. An authorised chain of acquisition. Who obtained it, from whom, and under what agreement at every hop.
4. Per-layer clearance for complex content. Video carries music, images, performances and trademarks. Clearing the video does not clear what is inside it.
5. Evidence that rights holders are paid. A licence nobody is being paid under is a dispute waiting to surface.
6. Clear withdrawal terms. What happens if a rights holder pulls out, and what that means for models already trained.

Add 1 more that is specific to contract structure rather than rights. Reported deal terms run multi-year: News Corp and OpenAI at 5 years, Reddit contracts reported at 2 to 3 years. Ask what happens at expiry, including whether models already trained on the data are affected. That clause decides whether you are renting or buying.

What does licensed training data cost?

Published bands vary by licence type more than by dataset size. Treat these as ranges reported across the market rather than quotes, and confirm current figures with any vendor before they reach a budget.

The trap in that table is the first row. A research licence is the cheapest way to get data and the most common source of an expensive mistake, because teams evaluate on research terms and then ship a production model without re-licensing. Read the permitted use clause before the price.

Payment structures vary too: flat fees, recurring annual payments, usage-based pricing and revenue shares all appear in reported deals. Usage-based looks attractive at evaluation scale and behaves very differently at production scale, which is the same trap we documented across the social data API pricing market.

If your requirement is live creator data rather than a bulk corpus, the field list for each platform is public. See the coverage list

How is grounding data different from training data?

Grounding data is retrieved at inference and inserted into the context, so the model answers from it rather than from its weights. That changes almost every requirement.

The last row is the one that reorders vendor shortlists. For training, a dataset from last quarter is normal. For grounding, a follower count from last quarter presented as current is the exact failure the retrieval layer exists to prevent.

This also changes what "public data" means for you. Grounding needs a live source with a documented refresh behaviour, not an archive. We covered what the public surfaces actually return, and how often, in social media public data.

What should you look for in a grounding source?

• A documented refresh interval per field. Some fields change hourly and some change annually. A single freshness claim across a whole dataset tells you nothing.
• Rate limits that survive your query volume. Grounding traffic scales with user queries, not with your corpus size, so the shape of your load is completely different from a training pull.
• Structured output rather than raw pages. Anything your pipeline has to parse at inference adds latency and a new failure mode.
• Provenance per field, not per dataset. If a user challenges an answer, you need to say where that specific number came from and when it was fetched.
• A defined behaviour when data is unavailable. A source that returns a stale value rather than an empty one will quietly poison your outputs.

Where does Phyllo fit?

On the grounding side, and specifically for creator and social data. We are not a training data provider and we do not sell bulk corpora for model training.

Phyllo's social data API returns live creator data across 25+ platforms through one schema. Where it differs from a public web source is the permission model: creators connect their own accounts and authorise access, so the data arrives from the platform with a documented consent record attached. For an AI product that has to explain where a number came from, that record is the provenance answer rather than a claim about it.

It also reaches fields no public source holds, because platforms release them to the account holder alone: true impressions, audience demographics, performance on ephemeral formats and earnings where the platform exposes them. We set out that boundary in authenticated versus public social data. Identity resolution ties a creator's accounts to 1 record so a grounded answer refers to 1 person rather than 5 guesses. The field list per platform is at getphyllo.com/coverage and the API reference is public.

For public conversation and mention data, where there is no account holder to ask, social listening covers the other half.

When is Phyllo the wrong choice?

Whenever you need a bulk corpus. If you are pre-training or fine-tuning a model and need volume, buy from a licensed marketplace or a web data platform. That is a different product and we are not a substitute for it.

The same applies if you need data on creators who have never authorised your product. Our model requires the creator to connect, so cold breadth across people with no relationship to you needs a public index instead. The wider stack, and which layer each vendor occupies, is in what is web data infrastructure.

How do you choose, in 4 questions?

1. Are you training or grounding? Training buys a corpus with a rights history. Grounding buys live access with a refresh guarantee. Very few vendors are genuinely good at both.
2. Can the vendor name the rights holder? If the answer is a category rather than a counterparty, the acquisition risk has not moved off your balance sheet.
3. Does the licence say the words "AI training"? General content licences frequently predate the question and do not grant it.
4. What happens at expiry or withdrawal? For training, whether trained models are affected. For grounding, whether access simply stops mid-quarter.

The short version

Decide whether you are training or grounding first, because that decides the vendor category, the contract shape and what you should be negotiating.

If you are training, buy provenance. The case law has been consistent enough that a documented licence naming AI training, with an identifiable rights holder behind it, is worth more than a larger corpus with a vaguer history.

If you are grounding, buy freshness and a clear answer to the question of where each field came from and when. A stale number returned as current is the failure your users will find first.

Working out whether your AI product needs licensed training data or a live grounding source? We can walk through the field-level difference. Get a demo