AI data market
The future of deidentified work data: four scenarios, eight dated signposts
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
The future of deidentified work data turns on two questions: whether the law judges identifiability from the recipient's position, and whether AI buyers keep paying for real work context as synthetic data improves. Crossed, they give four scenarios, from a licensable archive to a locked one. Since September 2025 the EU legal signal has moved toward the recipient test; in California worker data stays inside privacy law, and demand is unsettled.
Redaction
Deidentified work data
Key takeaways
- Work records are valuable for roles, sequence and outcomes, not names, so the commercial skill is preserving context while removing identity.
- De-identification settles only the privacy question; ownership of the work product and contract terms are separate permissions.
- The EU Court of Justice held in September 2025 that pseudonymised data are not personal data in every case for every person, and the Commission has proposed codifying it.
- California has covered employee data under the CCPA since its exemption expired at the end of 2022, and the FTC says hashing is not anonymization.
- NVIDIA's report that over 98 percent of Nemotron-4 340B's alignment data was synthetic is the strongest public signal against the demand side of the thesis.
- Companies that can show utility after de-identification on a held-out task hold value in all four scenarios.
What is the investor question behind deidentified work data's future?
The question is whether records of past work, such as tickets, case notes, dispatch logs and approval trails, can be licensed at scale once the workers who created them are no longer named, and who captures the value if they can. The answer depends on two things an investor can watch: the legal test for when de-identified records stop being personal data, and whether AI buyers keep paying for real work context as synthetic data improves.
Work records matter to AI developers because they show how work actually got done: who acted, in what role, in what order, under which policy, and with what result. That is the material an agent needs to learn a workflow rather than a description of one. Most of that value survives if names come out, and much of it dies if roles, sequence and outcomes come out with them.
Useful evidence for this decision is dated and checkable: court rulings, regulator guidance, legislative proposals, enforcement statements and model developers' own disclosures about their data. This article turns that evidence into a scenario matrix with eight signposts, then names the risks and the diligence questions that follow.
Why stripping names can strip the value out of work records
Because the context that makes work records useful is also what makes workers identifiable. A support ticket whose agent is replaced with a stable placeholder still teaches escalation and resolution; the same ticket stripped of every role, timestamp and handoff teaches almost nothing.
The difficulty is that rich context re-identifies people. A 60-person distributor may have one credit controller, so the role alone points to a person, and free-text notes carry signatures, nicknames and writing habits. Rocher, Hendrickx and de Montjoye's 2019 Nature Communications paper built a model that estimates how likely a specific person is to be correctly re-identified even in a heavily incomplete dataset, and concluded that even heavily sampled anonymized datasets are unlikely to meet the GDPR's standard for anonymization.
Regulators read shortcuts the same way. In a July 2024 technology post, the U.S. Federal Trade Commission argued that hashing does not make data anonymous and that data is anonymous only when it can never be associated back to a person. A hashed employee ID is a pseudonym, not an exit from privacy law.
So the commercial skill is context preservation under control: replacing names with consistent role-level placeholders, keeping sequence and outcomes, removing or generalizing rare roles and free-text identifiers, and binding recipients by contract not to re-identify. That skill is a prerequisite in every scenario below; which scenario prevails decides how much it is worth.
Ownership, privacy and contracts: three permissions a pseudonym does not merge
De-identification answers only one of three separate questions, and a licensing business needs all three answered. Who owns the work product, whether worker information is still regulated personal data, and what customer and vendor contracts allow are decided by different rules and often by different reviewers.
Ownership is usually the company's. Under U.S. copyright law, the Copyright Office's Circular 30 explains, work an employee creates within the scope of employment is a work made for hire, and the employer is treated as its author. That gives the company authority over the copyright in its records, but it says nothing about the privacy of the people named in them.
Privacy coverage has widened for workers. The California Privacy Protection Agency states that the CCPA's exemptions for employee and business-to-business data expired on December 31, 2022, and that the law's protections now include employees, job applicants and contacts at business customers, vendors and contractors.
Contracts are the third layer: customer agreements, vendor terms and data processing agreements can restrict a use that ownership and privacy law would allow. Each of the three needs expert review on the actual records, because a clean answer to one never implies the others. This is general information, not legal advice.
Two uncertainties, four futures for licensed work records
Cross the two uncertainties and four distinct futures appear. The legal axis runs from a strict reading, in which pseudonymised work records stay personal data in everyone's hands, to a recipient-relative reading, in which identifiability depends on what the holder can realistically do. The demand axis runs from substitutes being good enough, through synthetic and public data, to buyers needing recorded decisions, roles and outcomes that only real work leaves behind.
Each scenario below states what must be true, who would capture the value and what an investor should look for in a company exposed to it. None is a forecast of any company's results. The matrix:
| Scenario | What must be true | Who captures value | What to look for in a company |
|---|---|---|---|
| Licensable archive (relative test, real context needed) | Recipients that cannot re-identify may treat pseudonymised records as non-personal, and buyers still pay for real decisions and outcomes | Licensors who can preserve roles and sequence while proving recipients cannot re-identify | A repeatable de-identification method with measured utility loss, and recipient contracts banning re-identification |
| Controlled-access premium (strict test, real context needed) | Pseudonymised records stay personal data for every holder, yet buyers still need real context | Operators of controlled environments where models come to the data instead of copies leaving | Audit trails, access controls and the ability to run buyer workloads without releasing records |
| Evaluation niche (relative test, substitutes good enough) | Licensing is legally easier, but synthetic data covers most training needs | Holders of small, trusted, held-out test sets and seed data for generators | Evaluation-grade outcome labels and sealed holdouts, priced per test rather than per record |
| Locked archive (strict test, substitutes good enough) | Records stay personal data and buyers rarely need them | Companies using their own records for their own models | Low licensing volume; value shifts to internal tooling rather than outside sales |
Eight dated signposts and the scenario each one favors
Six of the eight signposts bear on the legal axis and two on demand, and they do not all point the same way. Each row gives the date, what the source says, the scenario it favors and the development that would reverse its effect, so a reader can check the call against later events without relying on anyone's upkeep.
Three signposts are EU legal texts on pseudonymised data, two are U.S. agency positions, one is peer-reviewed re-identification research, and the two demand signposts are a model developer's disclosure and a published forecast. The scorecard:
| Date and signpost | What the source says | Favors | What would reverse it |
|---|---|---|---|
| Jan 2025: EDPB Guidelines 01/2025 | Pseudonymised data attributable with additional information is personal data; deleting the key alone does not make it anonymous | Strict test | Revised guidance after the 2025 ruling |
| Sep 4, 2025: CJEU, EDPS v SRB (C-413/23 P) | Pseudonymised data are not personal data in every case for every person; the controller's duty to inform is judged from its own position at collection | Relative test for recipients | A narrow reading on remand or in later cases |
| Nov 19, 2025: Commission digital package | Proposes codifying the ruling so datasets can be shared when the recipient cannot re-identify; the pseudonymising controller keeps its GDPR duties | Relative test, if adopted | Parliament and Council dropping or narrowing the change |
| Jan 1, 2023: California employee data in scope | CCPA exemptions for employee and B2B data expired December 31, 2022 | Strict handling in the U.S. | A restored exemption or a federal preemptive law |
| Jul 24, 2024: FTC technology post | Hashing does not make data anonymous; data is anonymous only if it can never be tied back to a person | Strict test | Enforcement accepting pseudonyms as anonymous |
| 2019: Rocher et al., Nature Communications | Even heavily sampled anonymized datasets are unlikely to meet the GDPR anonymization standard | Strict test | Methods that measurably defeat such re-identification models |
| Jun 2024: NVIDIA Nemotron-4 340B report | Over 98 percent of the data used in the model's alignment was synthetically generated | Substitutes good enough | Evidence that synthetic alignment data fails on real workflows |
| 2024: Epoch AI public-data forecast | On current trends, training sets reach the size of the public human text stock between 2026 and 2032 | Real context needed | Synthetic or multimodal data closing the gap without private records |
Where the evidence leans in October 2026, and what would flip it
On the legal axis the EU evidence has moved toward the recipient-relative test since September 2025, while the U.S. evidence keeps pointing to strict handling of worker data. On the demand axis the evidence is split: one model developer, NVIDIA, reports mostly synthetic alignment data, and public text is forecast to run short. Neither axis is settled, so the honest position is conditional.
Each call below is dated 2026-10-26 and names the signal that would confirm or overturn it:
- If the Commission's change to the definition of personal data is adopted in substance, EU-origin work records move toward the licensable-archive or evaluation-niche scenarios. If it is dropped and the EDPB's January 2025 reading prevails, they move toward controlled access or a locked archive. Check against the final legislative text and any revised EDPB guidance.
- In the U.S., California worker records stay within the CCPA, so a licensor's de-identification evidence matters in every scenario. Check against new state laws and FTC enforcement that treat pseudonymised records differently.
- If developers' disclosures keep reporting synthetic shares near the Nemotron-4 level for post-training, the demand axis tilts toward substitutes and value concentrates in evaluation and seed data. If public disclosures start naming licensed real work records as inputs for agent training, it tilts back toward real context.
- Whatever the legal outcome, the companies best placed are those that can show utility after de-identification on a held-out task, because that evidence is worth something in all four futures.
Risks that could break the work-data thesis
Five developments could make the work-data thesis wrong or much smaller. Name them before the upside; each comes with a source where one exists.
At a glance
Risks that could break the work-data thesis
- Synthetic substitution: NVIDIA reported in 2024 that over 98 percent of Nemotron-4 340B's alignment data was synthetic. If synthetic workflows match real records on buyers' acceptance tests, the premium for licensed work records shrinks to evaluation use.
- A re-identification event: research such as Rocher and colleagues' shows how readily rich records re-identify people, and the FTC has said pseudonyms are not anonymity. One public re-identification of workers in a licensed dataset could tighten practice across the market.
- Legal reversal: the Commission's proposal is not law, and the EDPB's January 2025 guidelines take a stricter line. A narrow final text, or a narrow reading on remand in the SRB case, would push EU-origin records toward controlled access.
- Context destroyed by preparation: aggressive redaction can remove the roles, sequence and outcomes buyers pay for. A company that cannot measure utility after de-identification may be selling records that no longer teach anything.
- Buyer concentration: if only a few model developers buy work records, prices and terms follow their needs. No current public figure on buyer concentration for work records was found for this article, so treat it as an open risk to test in diligence.
Questions for any company licensing work records, SourceX included
Ask for evidence at the step where de-identification can destroy value or leak identity, and set the definitions yourself. SourceX should face the same test.
SourceX is one application of this thesis. Its business is turning business records a company selects into datasets licensed for AI, and it is SourceX that sources and qualifies each supplier, reviews rights, negotiates the contract, prepares the records, delivers them and pays the company. The step this article turns on, de-identification, is performed by SourceX itself or through its processing partner. The contract comes first: no work record reaches SourceX until the supplying company has approved what the deal covers and on what terms and has signed, and the signed agreement then bounds what SourceX may do with the de-identified dataset, from using it to sublicensing and distributing it.
SourceX has not published de-identification, utility or deal metrics; it shares company and commercial metrics with investors directly. It is raising from accredited investors under Rule 506(c), and investor materials are available on request through the investors page. Questions worth bringing:
- How are names replaced while roles, sequence and outcomes are kept? Show a before-and-after on a sample record.
- Which re-identification test is run, against which attacker and which auxiliary data, and who signs off on the result?
- Who holds the pseudonym key, and does any recipient ever receive it?
- Which contract terms bar recipients from re-identifying workers, and how are they monitored?
- For EU-origin records, which reading of pseudonymised data does the pipeline assume, and what changes if the strict reading prevails?
- What share of candidate work records fail de-identification or rights review, defined over which denominator and period?
Methods note: how the matrix and signposts were chosen
The scenario matrix and the eight signposts are editorial frameworks built for this article on 2026-10-26. Signposts were included only if they came from an official or primary source (a court, regulator, government agency, peer-reviewed journal or a model developer's own technical report), carried a date, and bore directly on either the identifiability of de-identified records or demand for real work context.
Sources: the CJEU press release on EDPS v SRB (September 2025), the EDPB's pseudonymisation guidelines (January 2025), the European Commission's digital package materials (November 2025), the California Privacy Protection Agency's FAQ, the U.S. Copyright Office's Circular 30, the FTC's July 2024 technology post, Rocher and colleagues (2019), NVIDIA's Nemotron-4 340B report (2024) and Epoch AI's data forecast.
Not measured: the volume or price of licensed work records, the share of records that survive de-identification with their value intact, the number of buyers, the current legislative status of the Commission's proposal beyond its publication, and any SourceX figure. None of these can be measured from public sources today.
Frequently asked questions
Can a company license employee work records without naming the employees?
Often, but removing names settles only the privacy question, and only partly. The company also needs to own the work product and to check that customer and vendor contracts allow the use. In the EU and California, pseudonymised worker data can still be regulated personal data, so counsel should review the specific records.
Is pseudonymised data still personal data under the GDPR?
For the company that holds the key, yes. The EU Court of Justice held in September 2025 that pseudonymised data are not personal data in every case for every person, and the Commission has proposed writing that into the GDPR, while the EDPB's January 2025 guidelines take a stricter view. Which reading applies to a given recipient depends on the facts and on whether the proposal is adopted.
Does the CCPA cover employee data?
Yes. The California Privacy Protection Agency says the exemptions for employee and business-to-business data expired on December 31, 2022. Employees, job applicants and business contacts who are California residents now have CCPA rights.
Will synthetic data make licensed work records unnecessary?
Possibly for some training uses: NVIDIA reported that over 98 percent of Nemotron-4 340B's alignment data was synthetic. Real records keep a role as held-out tests and as seeds for generators, which is the evaluation-niche scenario. Watch developers' disclosures for which way demand moves.
What should an investor ask a company that licenses work data?
Ask how it keeps roles, sequence and outcomes while removing names, which re-identification test it runs, who holds the pseudonym key, and what share of candidate records fail review. Set the definitions and denominators yourself so answers can be compared across companies.
Sources
- CJEU press release on EDPS v SRB (C-413/23 P), 4 September 2025: pseudonymised data are not personal data in every case for every person, and the controller's duty to inform is assessed from its own point of view at the time of collection. Source
- EDPB Guidelines 01/2025 on pseudonymisation (January 2025): pseudonymised data that could be attributed to a person with additional information is personal data, and deleting the additional information does not by itself make it anonymous. Source
- The European Commission's digital package FAQ says the proposal codifies a recent Court of Justice judgment so datasets can be shared when the recipient cannot re-identify individuals, while the pseudonymising controller keeps its GDPR obligations. Source
- Commission staff working document SWD(2025) 836 of 19 November 2025 records stakeholder concern about when a person is indirectly identifiable, including when pseudonymous data go to a recipient. Source
- The California Privacy Protection Agency says the CCPA exemptions for employee and business-to-business data expired on December 31, 2022, and the law now protects employees, job applicants and business contacts. Source
- U.S. Copyright Office Circular 30: a work created by an employee within the scope of employment is a work made for hire, and the employer is considered the author. Source
- Rocher, Hendrickx and de Montjoye (Nature Communications, 2019) estimate the likelihood of correct re-identification even in heavily incomplete datasets and conclude that heavily sampled anonymized datasets are unlikely to meet GDPR anonymization standards. Source
- FTC Office of Technology post (July 24, 2024): hashing does not make data anonymous; data is anonymous only when it can never be associated back to a person. Source
- NVIDIA's Nemotron-4 340B technical report (June 2024) states that over 98 percent of the data used in the model's alignment process was synthetically generated. Source
- Epoch AI's data forecast projects that, on current trends, models will be trained on datasets about the size of the public human text stock between 2026 and 2032, or earlier if overtrained. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.