Synthetic Data for AI Training: What It Is, Who Buys It and Its Limits

Synthetic data is data generated by software or by AI models rather than recorded from real people, systems or events, and AI developers now train on it next to crawled, licensed and human-made data. Anthropic's Form S-1 of 2026-10-26 lists synthetic data generated by its own models among its training data sources, and xAI's Grok 4.7 training-content summary of 2026-09-21 answers "Yes" for synthetic data from its own models. No filing or buyer page in this reference records a lab buying a synthetic dataset.

Within AI training data licensing, synthetic data is mostly generated rather than bought, but it still shapes deals for real records: micro1's license terms give the buyer ownership of synthetic data derived from licensed data, and Mercor's listings turn synthetic records away. This page covers the definition, four generation methods, who generates and refuses synthetic data, its effect on licenses, and its quality, privacy and disclosure limits. DataPacts sells no data and no generation software.

What synthetic data means in AI training

Synthetically generated means produced by a program, a simulation or a model to resemble real data, without each record describing a real person or event. In general machine learning the term covers any artificial dataset, including software test data, and that wider sense is not developed here. As data acquired for AI training, synthetic data reaches a developer by three routes: the developer generates it in-house with its own models, licenses software that generates it, or derives it from real records it has licensed. The table shows four types that appear in public statements, as of 2026-11-09, with the party behind each example.

Two of the four types start from real data. Synthetic tabular records come from a model fitted to a real table, so their legal status depends on what the generator retained from the source records. Synthetic identities differ in kind: micro1 describes, in a research post of 2026-09-24, replacing names and identifiers with consistent invented ones while the records themselves stay real, which works much like pseudonymization rather than generation from scratch. Model-generated text and code and synthetic agent tasks start from a trained model or a task template instead, and the labs and builders that produce them describe them as their own output, not as data bought from another party. The distinction matters for a data holder, because only the first two types use the holder's records as raw material.

How synthetic data is generated: four methods in outline

Synthetic data comes from four families of methods: rules and simulation, statistical models fitted to real records, outputs of generative AI models, and generated tasks for AI agents. Each method starts from something different, and the starting point decides both what the output looks like and where it fails. The EU AI Act treats the work as part of training itself: Recital 111 of Regulation (EU) 2024/1689 lists "pre-training, synthetic data generation and fine-tuning" among the activities counted in cumulative training compute. The table compares the four methods as of 2026-11-09, as an outline rather than a tool guide.

Software vendors sell the second and third methods as products. NVIDIA's synthetic-data page, checked 2026-11-09, names NeMo Data Designer and NeMo Safe Synthesizer, and MOSTLY AI publishes its software development kit under the Apache 2.0 license. Generated tasks are the newest family: CAMEL-AI stated on 2026-01-09 that its SETA project for terminal agents started from 400 initial synthetic tasks. Wrapped in tools and a scoring check, such tasks become the material of RL environments, which builders sell to labs as a product of their own. Generation tools and platforms are compared on the vendor page listed in the hub table further down, and step-by-step generation tutorials are outside this reference.

Who uses and buys synthetic data: what filings and buyers' pages show

The public record shows labs generating synthetic data with their own models, vendors selling generation software, and buyers of company data refusing synthetic records, but no lab stating that it bought a synthetic dataset. The table lists each statement in date order, as of 2026-11-09.

Each row is the named party's own statement, and no row records a purchase of a synthetic dataset.

AI labs that generate their own synthetic data

Anthropic, xAI and Apple each state in a filing, a regulatory summary or a technical report that synthetic data is part of their training data. Anthropic, PBC lists public web data, datasets licensed or commissioned from third parties, and synthetic data generated by its own models in its Form S-1 of 2026-10-26. SpaceXAI LLC, formerly xAI LLC, answers "Yes" for synthetic data generated with its own models, such as Grok 4.6, in the public summary of training content for Grok 4.7. Apple's 2025 foundation-model technical report names responsible web crawling, licensed corpora and synthetic data.

Labs that acquire data this way disclose categories, not volumes. The EU summary template, C(2025) 8311 final, discloses commercially licensed datasets only as Yes or No with their modality and is never a full dataset list, and the Grok 4.7 summary answers Yes for synthetic data without a volume, so it shows that synthetic data is used, not how much. Epoch AI's estimate, published 2024-06-06, frames the pressure behind generation: it puts the effective stock of public human-generated text at about 300 trillion tokens and expects it to be fully used between 2026 and 2032 (an 80 % interval). Generated data sits beside the paid routes set out in how frontier AI labs acquire data, from platform licenses to insolvency bids.

Vendors of synthetic data software and their 2026 ownership changes

Synthetic data vendors mostly sell software that generates data, not finished training sets, and two well-known names changed hands or addresses in the record by 2026. Their products are generators, such as the NeMo tools on NVIDIA's page and the MOSTLY AI software development kit named above. The record for the two names, as of 2026-11-09:

• MOSTLY AI: Syntho announced in Amsterdam on 2026-06-09 that it acquired the MOSTLY AI brand, trademark and related assets, with the value not disclosed, and the brand now reads "MOSTLY AI, powered by Syntho".
• Gretel: www.gretel.ai returned a permanent redirect (HTTP 301) to NVIDIA's synthetic-data page on 2026-11-09, and that page does not mention Gretel; the redirect alone shows no transaction terms or date, and this page states none.

Neither change is a sale of training data: the Syntho deal moved a brand and assets, and a redirect moves visitors. Vendors and their generation tools are compared on the vendor page in the hub table below, and the Syntho deal also appears among the dated deals linked at the end of this page.

Data buyers that turn synthetic records away

Buyers of real company data say synthetic records are what they do not want: Mercor's own listings name synthetic charts and synthetic traffic as not a fit. Mercor states the following, checked 2026-11-09:

• Medical practice or health system listing: "synthetic or sample charts" are not a fit, next to records without physician notes and data already licensed exclusively through a broker.
• AI product company listing: "synthetic or scripted traffic" is not a fit, next to logs that cannot be stripped of personal information and truncated or partial traces.

Both listings ask for records of real work instead: longitudinal patient records with free-text physician notes, and complete session logs from a product in production for 6 or more months with real users. The two listings, their requirements and the ranges Mercor states are set out in Mercor's data buyer profile, and generated look-alikes of the same records do not qualify under either.

What synthetic data changes in licenses for real company data

When a company licenses real records, the buyer may keep the synthetic data it derives from them, and some buyers' terms say so outright. micro1's Data License Master Terms, last updated 2026-11-08 and checked 2026-11-09, give the licensee ownership of derivatives, models, weights, outputs and synthetic data (s.2.2). The same terms require deletion of the delivered data at the earlier of 120 days after delivery completes or termination, but impose no duty to delete or retrain Licensee Materials (s.4.4). The table sets out the three clauses on record that touch synthetic or model-generated data, as of 2026-11-09.

Read together, the two micro1 clauses let synthetic data derived from a holder's records outlast the deletion of the records themselves. micro1 also shows a second use of the technique on the seller's side: its research post of 2026-09-24 describes consistent synthetic identities that keep records linked while names and identifiers are replaced. Anthropic's Section D.4 adds a limit from the provider's side, because a customer may not access Anthropic's services to train a competing AI model under terms effective 2025-06-17.

A holder negotiating a license can ask three questions about synthetic data: whether the buyer may generate it from the records at all, who owns what is generated, and whether deletion reaches derived data as well as the raw delivery. The clause-by-clause guide to AI data licensing agreements explains derived-data, deletion and ownership terms in full, including buyer terms that run the other way.

Where synthetic data reaches its limits

Synthetic data runs into three limits: model quality when training loops on generated data, privacy law when it comes from personal records, and EU rules on marking and disclosure.

Quality limits when models train on generated data

Training models on recursively generated data can degrade them, a process Shumailov and colleagues described in Nature in 2024. Their paper, published online on 2024-07-24 in Nature volume 631, pages 755 to 759, after a preprint of 2023-05-27 (arXiv 2305.17493), named the effect model collapse. The lab statements above all list real sources, such as web data and licensed corpora, next to synthetic data, and none describes training on generated data alone.

Two later papers disagree on how much generated data it takes, and the disagreement is not settled:

• Dohmatob and colleagues reported on 2024-10-07 that as little as 1 % synthetic data in the training set can trigger collapse, in a regression setting.
• Gerstgrasser and colleagues reported on 2024-04-01 that when real and synthetic data accumulate rather than replace each other, test error stays under a finite upper bound however many iterations run.

The two results rest on different setups, so neither shows that synthetic data always or never degrades a model; the outcome depends on how generated data is mixed with real data. The research is treated in depth on the model collapse page in the hub table.

When generated records still count as personal data

Synthetic data generated from personal records leaves the GDPR only if it is anonymous under Recital 26, judged by all means reasonably likely to be used, such as singling out. Three texts set the test, as of 2026-11-09:

• GDPR Recital 26, Regulation (EU) 2016/679, applying from 2018-05-25 (Art. 99): data protection principles "should therefore not apply to anonymous information", and identifiability is judged by all means reasonably likely to be used, such as singling out.
• EDPB Guidelines 02/2026 on Anonymisation, version 1.0, adopted on 2026-07-07 for public consultation and not final as of 2026-11-09: three criteria (No Record Isolation, No Linkage, No Inference), and section 3.4 (paragraphs 79 to 82) says inference from aggregated correlations "would, in particular, be the case for AI models or synthetic data", so a meaningful inference breaches No Inference.
• Utah Code 13-61-101, part of the Utah Consumer Privacy Act in force since 2023-12-31: "deidentified data" includes synthetic data, provided the holder makes a public commitment and passes the conditions on to recipients by contract; only the Attorney General enforces the act.

The EU draft and the Utah statute start from opposite ends: Utah writes synthetic data into its deidentified category by definition, while the EDPB draft names synthetic data as a typical place where inference survives. A holder that generates synthetic data from customer or employee records and offers it in the EU cannot treat it as outside data protection law by default, and the legal basis, notices and transfers for real records are covered under GDPR and AI training data.

Disclosure and marking rules under the EU AI Act

The EU AI Act, Regulation (EU) 2024/1689, treats synthetic data generation as part of model training and requires generated content to be marked. The act entered into force on 2024-08-01, and its obligations for general-purpose AI models apply from 2025-08-02.

Art. 50(2) requires providers of systems that generate synthetic audio, image, video or text to mark the outputs in machine-readable form. Systems placed on the market before 2026-08-02 must comply by 2026-12-02 under Art. 111(4), which Regulation (EU) 2026/1744 inserted. Marking concerns what a system outputs, not what a model was trained on.

On the training side, the provider's training-content summary is never a full dataset list, and the Grok 4.7 summary answers only Yes for synthetic data from its own models. The content and deadlines of that document are set out under the EU AI Act Article 53 training data summary. This section is information, not legal advice.

Choosing synthetic data or licensed real records

Synthetic data fits volume and privacy needs, while licensed real records fit work that depends on how real people and systems behave. The table compares the two on five needs, as of 2026-11-09.

The table is the hub's selection logic only. The full comparison of where each wins, including synthetic data made from real data, is the versus page in the hub table below, and neither column carries a public price in this reference.

Benefits, risks, downsides and alternatives for data holders, including not licensing at all

A company holding real records can license them, generate synthetic data from them, or do neither, and each choice changes what leaves the company.

Benefits:

• Less personal data in what is shared: synthetic data generated from the company's own records leaves the GDPR if it is anonymous under Recital 26 and meets the inference test in the EDPB draft.
• Demand for the real thing: Mercor's listings ask for real clinical notes and real session logs and turn synthetic versions away (checked 2026-11-09).
• Linked records without names: consistent synthetic identities, as micro1 describes them, keep records linkable while names and identifiers are replaced.

Risks:

• The buyer keeps what it derives: under micro1's terms (s.2.2 and s.4.4), synthetic data and models built from licensed records stay with micro1 after the delivered data is deleted.
• Generated data may still be personal data: the EDPB draft names synthetic data as a typical source of inference.
• Quality loss downstream: a buyer that trains on generated data in recursive loops risks the collapse effect described in Nature in 2024.

Downsides:

• No price on record: no filing, deal or vendor page covered in this reference states a price for a synthetic dataset or a generation license.
• Cost and work stay with the holder: generating synthetic data in-house means running generation software and testing the output for inference, with no buyer paying for either.

Alternatives:

• License real records under a narrow derived-data clause that limits what the buyer may generate and keep.
• License de-identified records instead, after names and identifiers are removed or replaced.
• Keep the data in-house and generate synthetic data for internal use only.
• Do not license at all: nothing leaves the company, and the guide on whether to sell or license your data treats declining as a valid outcome.

How this hub is organized and checked

Four deeper pages sit under this hub, one per question: vendors, synthetic versus real data, model collapse, and privacy law. The table lists them as of 2026-11-10.

Sources on this page are filings, regulators, statutes, peer-reviewed papers and buyers' own pages, each with the date it was stated or checked, and vendors appear whether or not they pay referral fees. The DataPacts Editorial Team checked every figure against its primary source, following how we source, date and check every figure.

Where synthetic data meets human data, RL environments and other acquisition routes

Synthetic data is one of several routes to training data, next to crawling, licensing, human data and RL environments. Related pages cover the neighboring routes and the steps a holder takes with real records.

• How AI companies get training data: crawling, licensing, users, humans and synthetic.
• Human data for AI: the counterpart to generated data, made by paid contributors.
• RLHF data: what reinforcement learning from human feedback needs and who supplies it.
• Data types AI companies buy: the real data types AI companies buy.
• AI data licensing deals: the Syntho and MOSTLY AI asset deal among the dated deals.
• Anonymization vs pseudonymization: what each achieves legally.
• How DataPacts is funded: no fees and no cut of any sale.

Frequently asked questions

What does synthetic data mean?

Synthetic data means data generated by a program, a simulation or an AI model instead of recorded from real people or events. In AI training it includes model-generated text and code: Anthropic's S-1 of 2026-10-26 lists synthetic data generated by its own models among its training data sources.

Is synthetic data reliable?

Synthetic data is as reliable as the way it is used. Shumailov and colleagues showed in Nature in 2024 that recursive training on generated data degrades models. Later studies disagree on the threshold: one reports collapse from a small synthetic share, another bounded error when real and synthetic data accumulate.

Do AI companies pay for synthetic data?

No filing or buyer page in this reference records an AI company paying for a synthetic dataset. Labs state that they generate synthetic data with their own models, and vendors sell generation software. Company data programs state ranges for real records instead, covered under what company data is worth to AI companies.

What happened to MOSTLY AI and Gretel?

Syntho acquired the MOSTLY AI brand, trademark and related assets, announced on 2026-06-09 with the value not disclosed. The brand now reads "MOSTLY AI, powered by Syntho". The Gretel website redirected to an NVIDIA synthetic-data page on 2026-11-09, and that page does not mention Gretel.

Can synthetic data replace a company's real records in a license?

Synthetic data does not replace real records for buyers that ask for real work. Mercor's listings, checked 2026-11-09, name synthetic charts and synthetic or scripted traffic as not a fit. Where a buyer generates synthetic data from licensed records, micro1's terms (s.2.2) give the buyer ownership of it.

Does synthetic data count as data acquired for AI training?

Synthetic data counts as data acquired for AI training, but it is mostly generated rather than bought. Anthropic's S-1 and xAI's Grok 4.7 summary both name synthetic data from their own models, and no record in this reference shows a lab buying a synthetic dataset.