# Is selling company data to AI safe? The real risks

URL: https://sellmydatatoai.net/guides/is-selling-company-data-to-ai-safe
Reviewed: 2026-11-08

Selling company data to AI can be done safely, but it is never risk-free, and the decision cannot be undone once a model is trained. The real risks are breaching privacy law or client contracts, licensing content you do not own, people being re-identified, a buyer misusing the data and a public backlash. Exclusions, de-identification, buyer vetting and tight license terms reduce most of them.

## Key takeaways

- Training is a new purpose. FTC staff warn against surreptitious, retroactive privacy policy changes, and the CCPA requires notice before incompatible new uses.
- Client contracts and NDAs are the risk sellers most often miss. Email and chat hold other companies' confidential information.
- You give warranties about rights you license, and the $1.5 billion Anthropic authors' settlement shows why buyers ask for them.
- De-identified records can be re-identified, and researchers have extracted names and email addresses verbatim from a trained model's training data.
- Some data should never be sold. Daniel Goldberg of Frankfurt Kurnit Klein & Selz advises excluding high-risk data entirely.

AI buyers are spending heavily on company records. The Next Web reported in October 2026 that micro1 had committed $1bn over 12 months to buying and licensing operational data. The money is real, and so are the risks, because a data license is hard to reverse.

Daniel Goldberg of Frankfurt Kurnit Klein & Selz put it plainly in October 2026: "Once data is used to train a model, it cannot practically be pulled back out." This guide sets out each risk honestly, what the law and regulators say about it, and the safeguards that reduce it. It is not legal advice for your company, so take independent advice before you sign anything.

## Is it safe to sell company data to AI companies?

Selling company data to AI companies is safe enough for many sellers once the high-risk records are excluded and the rest is prepared properly, but no sale is risk-free. The table summarizes the six risks that come up most and who usually carries each one.

| Risk | What can go wrong | Main safeguard | Who usually carries it |
| --- | --- | --- | --- |
| Privacy law | Personal data disclosed for a purpose people were never told about | Exclude or de-identify, and give notice first | The seller |
| Client contracts and NDAs | Breach of confidentiality owed to clients, suppliers or partners | Contract review and exclusion of client material | The seller |
| Copyright | Licensing material the company does not own | Chain of title review and removal of third-party content | The seller, through warranties |
| Re-identification | People identified from records or from model outputs | Documented method, re-identification test, contract ban | Both sides |
| Buyer misuse | Onward transfer, wider use or poor security | Buyer vetting and tight use limits | The seller in practice |
| Reputation | Staff, clients or users object in public | Plan disclosure before signing | The seller |

## What are the privacy law risks of selling data to AI?

The privacy law risk of selling data to AI is that training is a new purpose, and most records were collected without any mention of it. Goldberg notes that "most privacy policies, employee notices, and handbooks do not contemplate selling workplace communications for AI training".

Section 5 of the FTC Act makes unfair or deceptive acts or practices in commerce unlawful. FTC staff wrote on February 13, 2024 that it may be unfair or deceptive for a company to adopt more permissive data practices, including AI training, through "a surreptitious, retroactive amendment to its terms of service or privacy policy". The post cites the FTC's 2004 case against Gateway Learning, the maker of Hooked on Phonics.

State law adds specific duties. Frankfurt Kurnit Klein & Selz notes that a transfer of personal information for value is a sale under the CCPA unless the data meets its de-identification standard, and a sale gives consumers the right to opt out. The California Privacy Protection Agency's adjusted fines, in force from January 1, 2025, are up to $2,663 per violation and $7,988 per intentional violation.

Sector laws raise the stakes for specific records. HIPAA requires an authorization for any sale of protected health information, and Illinois' Biometric Information Privacy Act sets liquidated damages of $1,000 per negligent violation and $5,000 per intentional or reckless one.

The safeguard is to keep personal data out of the package where you can. Our guides to [CCPA, state privacy laws and GDPR](https://sellmydatatoai.net/guides/gdpr-and-selling-data-for-ai-training) and [de-identification](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data) cover the detail.

## Can client contracts stop you selling company data?

Client contracts can stop you selling company data, and they are the risk sellers most often miss. Goldberg warns that "email and chat contain other companies' confidential information, often covered by NDAs and commercial agreements", so an archive you hold is not automatically an archive you can license.

Agencies and service firms face a sharper version. If you hold client personal information as a CCPA service provider, California Civil Code section 1798.140(ag) requires your contract to bar you from selling it or using it outside the business purposes the contract specifies. Licensing a client's data for training is a purpose the client never set. Buyers know this: Polyshares says its license does not cover data a company holds on behalf of its own clients, and Endless Data says it never licenses client books or privileged client matter content.

- Read the confidentiality, data use and IP clauses in your largest client contracts first
- Check supplier and partner NDAs for restrictions on disclosing their information
- Check software vendor terms, since Hunton Andrews Kurth reported in August 2025 that Slack's revised API terms ban bulk export of data accessed through its APIs for use in large language models
- Exclude client deliverables and client data unless a contract clearly allows the use

## What copyright risks come with selling data for AI?

The copyright risk in selling data for AI is licensing material your company does not own, and then standing behind a warranty that says it does. Under section 201(b) of the US Copyright Act, the employer is the author of work employees prepare within the scope of their employment, but contractors keep their copyright unless they signed a written assignment, which section 204(a) requires. Statutory damages under section 504(c) run from $750 to $30,000 per work infringed, and up to $150,000 where the infringement is willful.

Courts have split on whether training itself is fair use. In Bartz v. Anthropic, Judge William Alsup held on June 23, 2025 that training on books was fair use but that building a library from pirated copies was not, according to Akin Gump. Two days later, in Kadrey v. Meta, Judge Vince Chhabria granted Meta partial summary judgment on fair use because the authors had not shown market harm. In Thomson Reuters v. Ross, the Third Circuit held on September 29, 2026 that copying Westlaw headnotes to train a non-generative legal research tool was not fair use, according to Haynes Boone.

Buyers take copyright seriously because their own exposure is large. A federal court in California gave final approval in July 2026 to a $1.5 billion settlement in the authors' suit over Anthropic's use of pirated books, about $3,000 for each of more than 400,000 books, according to Jurist. Those cases explain why buyers want a clean license and strong warranties rather than a fair use argument, and they pass that risk back to sellers through contract terms.

Troy Lieberman of Nixon Peabody wrote in October 2026 that remedies deserve as much attention as representations: "indemnities, liability caps, holdbacks, insurance, and termination rights", covering privacy claims, confidentiality breaches, IP disputes and the cost of removing data or retraining a model. He also recommends settling whether representations are qualified by knowledge, materiality or time period. Negotiate a cap and a knowledge qualifier before you sign.

## Can de-identified data be traced back to people?

De-identified data can sometimes be traced back to people, either from the records themselves or from the model trained on them. A 2019 study in Nature Communications by Luc Rocher and colleagues estimated that 99.98% of Americans would be correctly re-identified in any dataset using 15 demographic attributes. That is why HIPAA's Safe Harbor method removes 18 types of identifier, not just names.

Models add a second route. Nicholas Carlini and colleagues showed at USENIX Security 2021 that they could extract verbatim training sequences from GPT-2, including names, phone numbers and email addresses.

The safeguards are a documented method, a re-identification test run by someone who did not do the redaction, and a license term banning re-identification, which the CCPA's deidentified standard requires recipients to accept by contract. Our glossary entry on [re-identification risk](https://sellmydatatoai.net/glossary/re-identification-risk) explains the test.

## What happens if an AI buyer misuses your data?

If an AI buyer misuses your data, your remedy is usually a contract claim after the damage is done, so the protection has to come before transfer. Misuse covers passing data to affiliates or model partners, using it beyond the agreed purpose, holding raw copies longer than agreed and weak security.

Derived assets matter as much as the raw data. Nixon Peabody lists "cleaned datasets, annotations, embeddings, evaluation sets, fine-tuned models, model weights, and outputs" as layers a license should address one by one, and warns that deleting source files may not remove their influence from a trained model. It suggests a staged approach that starts with "a limited sample in a controlled environment".

Regulators can reach the models too. The FTC's January 2021 proposed settlement with the photo app developer Everalbum required it to delete "any facial recognition models or algorithms developed with Ever users' photos or videos". A buyer that trains on data it should not hold risks losing the model, which is one reason careful buyers check your rights so closely.

Buyers handle access differently. Scale AI says it anonymizes data inside the seller's environment before extraction, while other buyers connect extraction tools to company accounts. Goldberg advises companies to "avoid giving buyer tools direct access to company systems". Ask each buyer which model it uses before you share anything.

Our own buyer vetting checks company identity, intended use, security and retention commitments, onward transfer limits and, for larger deals, ability to pay, before a buyer sees a sample. The full standard is on [how we vet](https://sellmydatatoai.net/how-we-vet).

## What are the reputational risks of selling data to AI?

The reputational risk of selling data to AI is that staff, clients or users learn about the deal from someone else and object in public. Three published cases show the pattern.

- Stack Overflow: after its partnership with OpenAI, some users overwrote their answers in protest and had their accounts suspended, The Register reported on May 9, 2024.
- Wiley: The Bookseller reported in August 2024 that Wiley expected $44 million from AI rights deals and had confirmed there was no opt-out for authors.
- Spirit Airlines: flight attendants' unions and consumer privacy advocates objected to the sale of the bankrupt airline's data, Fortune reported in September 2026.

Staff messages carry a specific US risk. Goldberg notes that employee communications "may be protected concerted activity under Section 7 of the National Labor Relations Act", and that a collective bargaining agreement may restrict the sale. Our guide to [telling staff and clients about a data license](https://sellmydatatoai.net/guides/telling-staff-and-clients-about-a-data-sale) covers what to say and when.

## If you hold UK or EU data

UK and EU rules raise the same risks with stricter notice duties and larger fines. If your records include people in the UK or EU, check these points as well.

- Notice: Articles 13(3) and 14(4) of the UK GDPR require a controller that intends to process personal data for a new purpose to give people information about it before that processing starts. The ICO says: "You must bring any new uses of an individual's personal data to their attention before you start the processing."
- Fines: the ICO's fining guidance puts the higher maximum fine under UK GDPR at £17.5 million or 4% of total worldwide annual turnover, whichever is higher.
- Processors: Article 28(10) of the UK GDPR treats a processor that determines the purposes and means of processing in breach of the regulation as a controller for that processing.
- Copyright: under section 11 of the Copyright, Designs and Patents Act 1988, the employer owns work made by employees in the course of employment, and section 90(3) requires contractors' assignments to be in writing.
- Anonymization: the ICO says "simply removing direct identifiers from a dataset is insufficient", and the EDPB's Opinion 28/2024 says AI models trained with personal data "cannot, in all cases, be considered anonymous".
- Re-identification: section 171 of the Data Protection Act 2018 makes knowingly or recklessly re-identifying de-identified personal data an offense.

## How do you reduce the risks before you sign?

You reduce the risks before you sign by deciding what never leaves the company, then preparing, vetting and contracting in that order.

1. **Decide what never leaves**: Exclude HR and payroll files, health and benefits records, privileged material, payment card data and client data your contracts protect.
2. **Check notices and contracts**: Read the privacy notices in force when the data was collected, your client and supplier contracts, software vendor terms and any collective agreements.
3. **De-identify and test**: Apply a documented method and run a re-identification test on a random sample before any buyer sees it.
4. **Vet the buyer**: Confirm who the buyer is, what models the data will train, where it will be stored, who else may receive it and when raw copies are deleted.
5. **Share samples, not systems**: Show redacted samples under a mutual NDA, and transfer an exported copy rather than giving buyer tools access to live systems.
6. **Negotiate the license**: Set use limits, rules for derived assets, deletion dates, a ban on re-identification and capped, knowledge-qualified warranties. Our [licensing agreement checklist](https://sellmydatatoai.net/guides/ai-data-licensing-agreement-checklist) lists the clauses.
7. **Take independent legal advice**: Have your own lawyer review the license and the privacy position before you sign.

## When should you not sell company data?

You should not sell company data when the value sits in material you cannot clear. That includes records whose worth depends on personal data, archives your client contracts protect, code or content you cannot show you own, and deals where the buyer will not accept limits on use or onward transfer.

> **Rule of thumb** If you would not be comfortable explaining the sale to the people in the records, fix that first or leave those records out.

We are a broker, not a law firm. You approve every buyer, every sample and every license term, and you can withdraw at any point before signing. For a view of the rights issues in your own records, start with a [free data valuation](https://sellmydatatoai.net/data-valuation) or [book a consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). Our guide to [whether you can legally sell company data](https://sellmydatatoai.net/guides/can-you-legally-sell-company-data-for-ai) covers the legal tests in more depth.

## Frequently asked questions

### Will my data leak if I sell it to an AI company?

Leaks are possible at two points: in transfer and storage, and through the trained model. Researchers have extracted verbatim training text, including names and email addresses, from GPT-2. Reduce the risk by removing personal data before transfer, sending an exported copy rather than system access, and agreeing security, retention and deletion terms with the buyer in writing.

### Can I get my data back after a sale?

You can require the buyer to delete raw copies by a set date, but you cannot practically remove data from a model already trained on it, as Daniel Goldberg of Frankfurt Kurnit Klein & Selz notes. Nixon Peabody recommends that licenses say whether deletion, retraining or model retirement applies if a problem emerges. Agree those terms before transfer, not after.

### Am I liable if the buyer misuses my data?

You may be, depending on the law and the contract. Under the CCPA, a business that sells personal information must tell consumers and honor opt-outs, and data only counts as deidentified if recipients are bound by contract not to re-identify it. Contractually, your exposure depends on the warranties and indemnities you give. Limit onward transfer and use in the license, vet the buyer first and take independent legal advice on liability.

### Is de-identified data safe to sell?

Data that meets a legal de-identification standard, such as the CCPA's three-part test or HIPAA's Safe Harbor method, falls outside the main privacy rules, but no method is perfect. A 2019 Nature Communications study estimated 99.98% of Americans could be re-identified from 15 demographic attributes. Treat de-identification as a tested result, documented with a re-identification check, rather than a label.

### Do I need a lawyer to sell company data to AI?

You should take independent legal advice before signing any data license. The questions involve privacy law, client contracts, copyright and employment law, and the license will contain warranties and indemnities that set your exposure. A broker can tell you what buyers check and help you prepare, but a broker is not a substitute for your own lawyer.

## Sources

- [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein & Selz, accessed November 2026
- [micro1 commits $1bn to enterprise data for AI training](https://thenextweb.com/news/micro1-1bn-enterprise-data-ai-training-citi-hercules), The Next Web, accessed November 2026
- [15 U.S. Code section 45: unfair methods of competition unlawful](https://www.law.cornell.edu/uscode/text/15/45), Legal Information Institute, Cornell Law School, accessed November 2026
- [California Civil Code section 1798.140 (definitions)](https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=CIV&sectionNum=1798.140), California Legislative Information, accessed November 2026
- [California Privacy Protection Agency announces 2025 increases for CCPA fines and penalties](https://cppa.ca.gov/announcements/2024/20241217.html), California Privacy Protection Agency, accessed November 2026
- [45 CFR 164.508: uses and disclosures for which an authorization is required](https://www.law.cornell.edu/cfr/text/45/164.508), Legal Information Institute, Cornell Law School, accessed November 2026
- [Biometric Information Privacy Act, 740 ILCS 14](https://www.ilga.gov/legislation/ilcs/ilcs3.asp?ActID=3004&ChapterID=57), Illinois General Assembly, accessed November 2026
- [17 U.S. Code section 201: ownership of copyright](https://www.law.cornell.edu/uscode/text/17/201), Legal Information Institute, Cornell Law School, accessed November 2026
- [17 U.S. Code section 204: execution of transfers of copyright ownership](https://www.law.cornell.edu/uscode/text/17/204), Legal Information Institute, Cornell Law School, accessed November 2026
- [17 U.S. Code section 504: remedies for infringement, damages and profits](https://www.law.cornell.edu/uscode/text/17/504), Legal Information Institute, Cornell Law School, accessed November 2026
- [District court rules AI training can be fair use in Bartz v. Anthropic](https://www.akingump.com/en/insights/ai-law-and-regulation-tracker/district-court-rules-ai-training-can-be-fair-use-in-bartz-v-anthropic), Akin Gump, accessed November 2026
- [Second district court rules AI training can be fair use](https://www.akingump.com/en/insights/ai-law-and-regulation-tracker/second-district-court-rules-ai-training-can-be-fair-use), Akin Gump, accessed November 2026
- [Third Circuit rejects fair use for ROSS's copying of Westlaw headnotes](https://www.haynesboone.com/news/alerts/third-circuit-rejects-fair-use-for-rosss-copying-of-westlaw-headnotes), Haynes Boone, accessed November 2026
- [3rd Circuit rules for Thomson Reuters in its copyright fight against legal research startup Ross](https://www.lawnext.com/2026/09/3rd-circuit-rules-for-thomson-reuters-in-its-copyright-fight-against-legal-research-startup-ross.html), LawSites (LawNext), accessed November 2026
- [FTC press release on the proposed Everalbum settlement (January 11, 2021)](https://www.ftc.gov/node/48633), Federal Trade Commission, accessed November 2026
- [UK GDPR Article 13](https://www.legislation.gov.uk/eur/2016/679/article/13), legislation.gov.uk, accessed November 2026
- [UK GDPR Article 14](https://www.legislation.gov.uk/eur/2016/679/article/14), legislation.gov.uk, accessed November 2026
- [UK GDPR Article 28](https://www.legislation.gov.uk/eur/2016/679/article/28), legislation.gov.uk, accessed November 2026
- [The right to be informed](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/individual-rights/the-right-to-be-informed/), Information Commissioner's Office, accessed November 2026
- [The maximum amount of a fine under UK GDPR and DPA 2018](https://ico.org.uk/about-the-ico/our-information/policies-and-procedures/data-protection-fining-guidance/statutory-background/the-maximum-amount-of-a-fine-under-uk-gdpr-and-dpa-2018), Information Commissioner's Office, accessed November 2026
- [AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive](https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/02/ai-other-companies-quietly-changing-your-terms-service-could-be-unfair-or-deceptive), Federal Trade Commission, accessed November 2026
- [Polyshares homepage](https://polyshares.com/), Polyshares, accessed November 2026
- [Endless Data homepage and FAQ](https://endless-data.com/), Endless Data, accessed November 2026
- [Salesforce locks down Slack data: time to review your Slack API terms](https://natlawreview.com/article/salesforce-locks-down-slack-data-time-review-your-slack-api-terms), The National Law Review (Hunton Andrews Kurth), accessed November 2026
- [Copyright, Designs and Patents Act 1988, section 11](https://www.legislation.gov.uk/ukpga/1988/48/section/11), legislation.gov.uk, accessed November 2026
- [Copyright, Designs and Patents Act 1988, section 90](https://www.legislation.gov.uk/ukpga/1988/48/section/90), legislation.gov.uk, accessed November 2026
- [Judge approves record $1.5 billion AI copyright settlement involving Anthropic](https://www.jurist.org/news/2026/07/judge-approves-record-1-5-billion-settlement-involving-anthropic/), Jurist, accessed November 2026
- [Shopping for AI training data? Five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody, accessed November 2026
- [How do we ensure anonymisation is effective?](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/how-do-we-ensure-anonymisation-is-effective), Information Commissioner's Office, accessed November 2026
- [Estimating the success of re-identifications in incomplete datasets using generative models](https://pmc.ncbi.nlm.nih.gov/articles/PMC6650473), Nature Communications (Rocher, Hendrickx and de Montjoye, 2019), accessed November 2026
- [Extracting training data from large language models](https://www.usenix.org/conference/usenixsecurity21/presentation/carlini-extracting), USENIX Security Symposium 2021, accessed November 2026
- [Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models](https://edpb.europa.eu/system/files/2024-12/edpb_opinion_202428_ai-models_en.pdf), European Data Protection Board, accessed November 2026
- [Data Protection Act 2018, section 171](https://www.legislation.gov.uk/ukpga/2018/12/section/171), legislation.gov.uk, accessed November 2026
- [Data partnership](https://scale.com/data-partnership), Scale AI, accessed November 2026
- [Stack Overflow simply bans folks who don't want their advice used to train AI](https://www.theregister.com/2024/05/09/stack_overflow_banning_users_who/), The Register, accessed November 2026
- [Wiley set to earn $44m from AI rights deals, confirms 'no opt-out' for authors](https://thebookseller.com/news/wiley-set-to-earn-44m-from-ai-rights-deals-confirms-no-opt-out-for-authors), The Bookseller, accessed November 2026
- [Google wants Spirit Airlines' data. Micro1 wants to pay more](https://fortune.com/2026/09/14/micro1-tries-to-beat-googles-bid-to-buy-bankrupt-spirit-airlines-data-for-12-5-million-highlighting-value-of-data-to-ai-companies/), Fortune, accessed November 2026

---

Book a free consultation: https://sellmydatatoai.net/book-a-consultation