Can a Bankrupt Company Sell Your Work Emails to Train AI? The Spirit Airlines Case Explained

Yes—potentially. A bankrupt company can generally sell corporate data and other intangible assets that it legally owns or has transferable rights in. But bankruptcy does not give it unlimited ownership over every fact, message or file sitting on its servers, nor does it erase privacy law, confidentiality, privilege, third-party intellectual-property rights or other legal restrictions.

Spirit Airlines has turned that abstract question into a remarkably concrete one.

Why We Investigated This: The headline is that Google bid $10 million for Spirit Airlines’ data. The more consequential question is what happens to decades of employee emails, chats, HR records, code and other workplace history when a company dies—and AI companies decide that history has value.

On August 14, Google won a bankruptcy auction with a $10 million bid for a vast portion of Spirit’s internal digital archive. Mercor.io was designated the alternate bidder at $7.5 million. According to the sale schedule, the proposed package includes roughly 100 million emails, 500 million Microsoft Teams items, 17 million OneDrive items, 20.6 million SharePoint items, 175,658 employee records, 3.4 million payroll records, 148,018 employee tax forms, approximately 30 million lines of software code and billions of operational and transactional records. Major customer datasets, including 97.5 million passenger profiles and 50.2 million Free Spirit member records, are specifically marked “Not Included.”

But Google has not yet received that archive. The sale agreement requires bankruptcy-court approval and de-identification before transfer. After the Association of Flight Attendants-CWA objected, the court postponed the approval hearing until September 9, 2026. Reuters reports that Google intends to use the data for product development and AI-model training.

The union’s objection exposes the part of this story that matters far beyond Spirit Airlines:

Removing a worker’s name is not necessarily the same thing as removing what the company knows about that worker.

And the proposed sale contains one unusually important technical requirement: the data must be de-identified while “preserving referential integrity across the data set.”

That phrase may explain both why this archive is attractive to AI companies and why former employees are worried about it.

What Exactly Is Spirit Airlines Trying to Sell?

Calling this a sale of “old emails” seriously understates it.

The assets schedule attached to the Google sale agreement reads more like an inventory of the airline’s institutional memory: what employees communicated, how work was scheduled, what systems recorded, how software changed, how flights operated, how prices were set and what happened afterward.

The court filing identifies the following among the proposed assets. All included data remains subject to the sale agreement’s de-identification requirements.

The court filing verifies the categories, counts and whether each category is included. The explanations of possible AI value in the table are sherafy.com’s technical analysis, not a claim that Google has announced a particular use for every category.

That distinction matters because Google has said broadly that Spirit’s enterprise data could help improve its products and AI models, but it has not publicly disclosed which models will receive which data, the precise training methodology, or whether every category will be used for model training at all. (Business Insider)

There is another detail easy to miss in summaries of the transaction. Although the large customer datasets listed above are excluded from the proposed Google package, the agreement separately reserves Spirit’s right to sell a customer data list, including annual traveler spending information, to parties in the hospitality or travel industries. That is separate from what Google is buying and should not be conflated with the Google transaction.

Can a Bankrupt Company Really Sell Employee Emails?

The basic legal answer is yes, if the company possesses a property interest that can legally be transferred.

Section 541 of the Bankruptcy Code creates a bankruptcy estate containing broadly the debtor’s legal and equitable interests in property. That can include intangible assets—not merely airplanes, buildings, cash or physical equipment. (Legal Information Institute)

Section 363 then provides a mechanism for selling estate property outside the ordinary course of business after notice and a hearing. Spirit’s proposed Google transaction expressly invokes Section 363. (Legal Information Institute)

But there is an equally important qualification.

A bankruptcy estate does not automatically receive more rights than the debtor had.

The legislative notes accompanying Section 541 put the principle plainly: only the debtor’s interest in property becomes property of the estate. If the debtor’s interest was limited, the estate generally inherits that limitation. (Legal Information Institute)

So the correct rule is not:

“If it is on the company’s server, the bankruptcy court can sell it.”

It is closer to:

“The bankruptcy estate can monetize the legal interests the company actually possesses, subject to bankruptcy law and other applicable rights and restrictions.”

That difference becomes important when one archive contains employee information, vendor material, licensed software, privileged legal communications, trade secrets and intellectual property belonging to multiple parties.

Does Your Employer Actually “Own” Your Work Email?

The common shorthand—“your work emails belong to your employer”—contains some truth but is legally too crude to answer this case.

An employer generally controls its corporate email environment and normally has extensive rights to access, retain and use records generated through its business systems. Employees should therefore not treat a company mailbox or Teams account like a private diary.

But control of the mailbox does not automatically establish unlimited ownership over every legal interest contained inside every message.

An email might contain the employer’s ordinary business record. It might also contain an employee’s personal information, another company’s trade secret, licensed material, attorney-client privileged advice, protected health-related information, a document subject to a confidentiality agreement, or intellectual property whose ownership depends on separate law or contract.

Spirit’s own sale agreement recognizes some of those distinctions. It expressly excludes privileged material and provides procedures for returning or destroying privileged material discovered after transfer. It also says no Spirit contracts, agreements or licenses are being assigned to the buyer through the transaction.

The better lesson for workers is therefore not simply “your emails aren’t yours.”

It is:

What you create on a work system can become part of a durable corporate record with a life—and potentially a market value—far beyond the job for which you created it.

Why Are Customer Records Treated Differently From Employee Records?

This is one of the most revealing aspects of the Spirit case.

The Bankruptcy Code contains a specific privacy mechanism for certain consumer information.

Under Section 363(b)(1), when a company has a privacy policy restricting transfers of personally identifiable information obtained while offering products or services, the bankruptcy estate can face additional conditions before selling that information inconsistently with the policy. In qualifying circumstances, the court can appoint a consumer privacy ombudsman to evaluate the privacy consequences of the transaction. (Legal Information Institute)

The statutory definition of “personally identifiable information” in this particular bankruptcy provision is itself consumer-oriented. Section 101(41A) begins with information supplied by an individual in connection with obtaining a product or service primarily for personal, family or household purposes. (Legal Information Institute)

That history reflects a familiar bankruptcy problem.

In 2000, failed online retailer Toysmart.com attempted to sell customer information after having promised users that their information would not be shared with third parties. The Federal Trade Commission sued, and a settlement imposed strict conditions on any transfer of the database. (Federal Trade Commission)

Spirit presents a different version of the problem.

Its most controversial proposed asset is not a conventional customer mailing list. It is a sprawling record of work itself.

AFA explicitly acknowledges this distinction in its objection. The union does not argue that the Bankruptcy Code’s special Section 363(b)(1) consumer-data provision necessarily governs this sale. Its argument is narrower: the privacy architecture used in the Spirit agreement borrows heavily from consumer privacy concepts while the material being transferred is disproportionately employee-facing.

That does not mean employees have no privacy rights.

California is an obvious counterexample. The California Privacy Protection Agency states that CCPA rights extend to California residents who are employees and job applicants, and the old exemptions for employment-related personal information expired at the end of 2022. (California Privacy Protection Agency)

Other state laws, employment rules, contracts and data-specific statutes can also matter.

So the accurate conclusion is:

There is a potential gap in the bankruptcy-specific consumer privacy framework—not a blanket rule that employee data is legally unprotected.

Exactly which protections apply to a particular employee or record can depend on the jurisdiction, the type of information, the company involved, contractual terms and the purpose of the transfer.

“De-identified” Does Not Mean “Nothing Sensitive Is Left”

Spirit’s proposed agreement does contain meaningful privacy requirements.

Before Google receives the assets, the data must first pass through one or more third-party “Deidentification Agents” acceptable to or designated by Google. The agreement requires reasonable measures intended to prevent the resulting data from being associated with, linked to, or reasonably used to infer information about a particular consumer. It calls for the California de-identification standard to be used for U.S. consumer information even where the CCPA itself might not apply, and for a HIPAA de-identification standard for protected health information or other consumer health-related data where applicable.

Google also commits to keep the transferred dataset in de-identified form and not intentionally associate it with a person or household. The agreement permits Google to transfer the de-identified data to another party only if that party is contractually bound to the same restriction.

Those are important facts.

But they do not completely answer AFA’s objection, because identity and confidentiality are different problems.

Consider a hypothetical record:

Employee X at a particular crew base failed a recurrent training event, later filed a scheduling grievance and received a specific payroll adjustment.

Deleting the employee’s name may substantially reduce the risk of direct identification.

It does not make the underlying facts suddenly non-sensitive.

AFA gives similar examples involving disciplinary communications, training deficiencies, leave or accommodation requests, grievances and compensation adjustments. Its argument is that the confidentiality of such information may reside in its substance and context, not merely in a name field that can be removed.

That is arguably the strongest part of the union’s objection.

Why “Referential Integrity” Is the Most Important Phrase in the Deal

The sale agreement requires de-identification to occur while preserving referential integrity across the dataset.

That sounds like database jargon. It is central to understanding the controversy.

Imagine that an employee’s real name disappears and the person instead becomes:

Employee 004731

The archive could still retain valid relationships showing that Employee 004731 appears in:

a time card → a crew pairing → a training record → a Teams discussion → a payroll adjustment → an operational event

The identity can be obscured while the relationships among records remain intact.

That is not necessarily a defect in de-identification. A useful de-identified dataset can preserve relationships through pseudonymous or otherwise non-identifying keys without retaining a person’s real identity.

But keeping those relationships dramatically changes what remains in the archive.

A pile of disconnected, anonymized text fragments tells you relatively little about how a company works.

A linked archive can potentially preserve process.

AFA says the Spirit agreement means joins among timekeeping, crew pairing, training, payroll and communications information will remain. It points to the relatively small size of the crew-base population—4,600 in the dataset—and argues that combining numerous linked attributes across a structured group can allow sensitive information concerning individuals or small groups to be inferred.

There is an important limit to that claim.

AFA explicitly says it is not making a technical claim that any particular Spirit record can actually be re-identified.

That distinction should not be lost.

The union has not demonstrated that Google can recover the identity behind Employee 004731. Nor has Google been caught attempting to do so; the proposed dataset has not yet been delivered.

The dispute is prospective:

Are the protections sufficient before a highly interconnected corporate archive leaves Spirit’s control?

Google Has More Influence Over the De-identification Process Than “An Independent Third Party Will Scrub It” Suggests

Some accounts of the transaction describe the data as simply being anonymized by an independent third party.

The contract is more specific.

The “Deidentification Agent” must be acceptable to or designated by the buyer, Google pays the cost of the process, and the completed de-identification must be certified to Google’s reasonable satisfaction. Spirit must also give Google a reasonable opportunity to review and comment on the de-identification process and consider those comments in good faith.

None of that, by itself, shows that the de-identification will be inadequate.

But “an independent third party handles everything” leaves out a material part of the contractual structure.

The buyer has a legitimate interest in receiving a dataset that remains technically useful after privacy-sensitive information has been removed. The union’s concern is that the qualities that preserve usefulness can overlap with the qualities that preserve sensitive context.

That brings us to the other half of this story.

Why Would AI Companies Pay Millions for a Dead Airline’s Digital Memory?

Google has said Spirit’s enterprise dataset “can be helpful in improving our products and AI models.” Mercor, the $7.5 million alternate bidder, described corporate archives as records showing “how real work gets done.” (Business Insider)

The exact training plans are not public.

But the structure of the archive suggests why multiple AI companies see value here.

Reasonable inference: the prize is not merely 100 million sentences

For years, frontier AI models have consumed enormous quantities of publicly available books, websites, code, forums, documentation and other internet material.

That material contains tremendous knowledge.

It has a weakness: much of it shows the finished product of human work.

A corporate archive can contain something different:

problem → discussion → document → revision → decision → action → system change → outcome

Consider an IT ticket.

A public webpage might contain a polished tutorial explaining how a technical problem should be solved.

A real corporate archive can contain the original messy complaint, the employee’s questions, the wrong diagnosis, the escalation, the eventual fix and the evidence that the fix worked.

Or consider software development.

Spirit’s proposed package contains not merely approximately 30 million lines of source code but complete Git histories, hundreds of thousands of commits, pull requests, review discussions, bug reports and CI/CD artifacts.

That potentially records not only what the finished code was, but how engineers got there.

The same principle can apply to pricing, scheduling, customer operations, finance and internal communication.

This creates several plausible uses for an AI developer: model training or post-training, evaluation of AI agents performing multi-step business tasks, enterprise-search research, code-model development, workflow simulation, or the construction of realistic task environments.

Those are technical possibilities, not confirmed descriptions of Google’s plan.

Google has not publicly identified which models it intends to use, whether Spirit data would enter base-model training, post-training, evaluation systems or other products, or how much of the archive would be useful after de-identification.

The File Format Makes the Dataset More Interesting

The court filing shows that Spirit is not proposing to hand Google a giant folder of PDFs.

Microsoft 365 material—including email, OneDrive, SharePoint and Teams data—is supposed to be retained and transferred within its native Microsoft 365 environment.

Other systems can be exported in machine-readable formats such as CSV, JSON or native SQL database dumps. Source-code repositories can be transferred as Git repositories or bundles, with metadata such as commits, pull requests, issues and comments represented in structured form.

That matters because structure is information.

A machine-readable database can retain fields, timestamps and relationships that a flattened document may lose. A Git repository retains software history. A collaborative environment can preserve threads and chronology.

Again, this does not establish Google’s exact intended training technique.

It helps explain why referential integrity and native structure are potentially valuable in the first place.

Corporate Email Has Become AI Training Material Before

There is a historical comparison, although the circumstances were very different.

After Enron collapsed, a large collection of the company’s internal email became public through the Federal Energy Regulatory Commission’s investigation. Carnegie Mellon hosts a cleaned research version containing approximately 500,000 messages from about 150 users. It became an important real-world dataset for email and machine-learning research. (CMU School of Computer Science)

Spirit’s proposed archive contains roughly 100 million emails alone—about 200 times the raw email count—plus hundreds of millions of Teams items, operational databases, documents, software histories and other structured systems.

The comparison should not be pushed too far. Enron’s email entered the public research ecosystem through a regulatory investigation; Spirit is proposing a commercial bankruptcy sale subject to de-identification.

But Enron demonstrated something decades ago that is now becoming economically important:

The internal communications of a dead company can acquire an entirely new life as machine-learning data.

Spirit may represent that concept at industrial scale.

Does Spirit Actually Own Everything It Is Trying to Sell?

That question is no longer hypothetical.

On August 21, software company Springshot Inc. filed a limited objection asserting that portions of the proposed dataset contain Springshot proprietary intellectual property covered by a 2022 software agreement with Spirit.

Bloomberg Law reported that Springshot claims it owns certain data included in the proposed transaction. That is an allegation in a pending bankruptcy dispute, not a judicial finding that Spirit is attempting to sell property it does not own. (Bloomberg Law)

But the objection illustrates an important principle.

A database can look technically like one asset while containing many separate legal interests.

Spirit’s Google agreement represents that Spirit is the sole owner of the assets being sold and has the right to transfer them. But the same agreement says no contracts, licenses or other arrangements are being assumed or assigned to Google.

Springshot is effectively asking the court to examine the boundary between those propositions.

Bankruptcy can sell Spirit’s rights.

It cannot manufacture an ownership interest Spirit never possessed.

Can Employees Object to a Sale Like This?

Yes.

That is exactly what AFA has done.

But there is no general rule giving every individual employee a personal veto over every transfer of corporate business records generated during employment.

Instead, the available arguments depend on the rights involved.

AFA’s objection is relatively targeted. It says it does not seek to unwind the auction or prevent Spirit from monetizing data assets generally. Its primary request is to exclude flight-attendant information; alternatively, it asks for additional safeguards around confidential employee material.

Among the issues raised by the union are free-form employment communications, the possibility of inference from linked records, downstream transfers and the absence of a separate confidentiality review focused specifically on employee material.

That is a stronger argument than pretending that every work email is inherently private.

Some workplace material may be routine and commercially useful without presenting a meaningful privacy problem.

Other records may describe investigations, medical or accommodation issues, grievances, disciplinary events, performance problems or highly specific combinations of events that remain sensitive even after direct identifiers are stripped.

The hard question is where to draw that line across hundreds of millions of records.

The Buyer May Not Even End Up Being Google

The phrase “Google bought Spirit Airlines’ emails” is premature.

Google is currently the successful bidder, not the recipient of a completed transfer.

The August 14 auction notice names Google as the successful bidder at $10 million and Mercor as the $7.5 million alternate bidder. The agreement itself makes bankruptcy-court approval a condition to consummating the transaction.

Then the situation changed.

After missing the original auction deadline, AI startup Micro1 sent Spirit’s legal team a $12.5 million offer, according to Business Insider. Micro1 CEO Ali Ansari said the company was working to present details of the offer to the court. Bankruptcy experts interviewed by the outlet said reopening a properly conducted auction for a late bidder would be unusual but is not categorically impossible. (Business Insider)

As of August 28, the more precise description is therefore:

Google won the authorized auction. Micro1 subsequently made a higher late offer. The Google transaction still requires court approval, and it is not yet clear whether Micro1’s proposal will alter the result.

That uncertainty is one reason this article should not be titled “Google Bought Spirit Airlines’ Emails.”

What Happens at the September 9 Hearing?

The bankruptcy court originally scheduled the sale hearing for August 19. After AFA filed its objection, the hearing was postponed until September 9. (Investing.com)

Several questions are now hanging over the transaction: whether the Google sale should be approved under its current terms; whether additional employee-data protections or exclusions are warranted; how the Springshot ownership objection should be addressed; and whether Micro1’s late higher offer changes the economic calculus.

The court could approve the transaction substantially as written, approve a sale with additional protections, require further work, continue one or more disputes, or otherwise alter the process.

What the court has not done yet is approve Google taking possession of the data.

That distinction is especially important because some coverage already describes the acquisition in completed terms.

What Should Employees at Ordinary Companies Learn From the Spirit Case?

Spirit is unusual because of its size, bankruptcy and the sheer quantity of data involved. The underlying problem is not unusual at all.

Five lessons are worth carrying beyond this case:

1. Workplace data can outlive both your job and the company itself. A message created for an ordinary Tuesday meeting can remain inside a corporate archive years later.

2. Corporate data can become an asset in its own right. AI has increased the potential value of records that companies historically treated as operational exhaust.

3. De-identification and confidentiality solve different problems. Removing who someone is does not necessarily remove what a record says about them.

4. “The company owns the email” is not the end of the legal analysis. Privacy law, contracts, privilege, intellectual property, employment rules and other rights can still matter.

5. The relationships among records may be more valuable than the records themselves. For AI systems attempting to learn real workflows, a linked trail from problem to discussion to action to outcome can carry information that isolated documents cannot.

Spirit’s bankruptcy therefore raises a question companies and employees rarely had reason to contemplate until now:

What happens to an organization’s digital memory when the organization no longer exists?

For decades, a dead company’s old inbox looked mostly like a storage expense, a discovery risk or something to delete under a retention policy.

In the AI economy, it can also look like a training corpus.

And apparently, it can be worth millions of dollars.

The Bottom Line

Yes, a bankrupt company can potentially sell employee emails and other workplace data for AI use if those assets represent transferable property interests belonging to the company. But bankruptcy does not create unlimited ownership over everything stored in a corporate system, and the information remains subject to applicable privacy, confidentiality, privilege, intellectual-property and other legal restrictions.

In Spirit’s case, the proposed buyer would not receive the raw archive as it exists today. The agreement requires de-identification first, excludes major customer datasets and privileged material, and obligates the buyer to keep the resulting data de-identified.

The unresolved issue is more difficult.

Spirit’s flight attendants argue that a system designed to remove identity does not necessarily remove confidential meaning, particularly when the sale deliberately preserves relationships among payroll, scheduling, training, operational and communication records. Their objection does not prove that individual employees can be re-identified, and the union expressly says it is not claiming that. It asks whether de-identification alone is the right test for a dataset containing decades of workplace history.

That is the part of the Spirit case likely to outlive the bankruptcy itself.

AI has created a market for something businesses have accumulated almost accidentally for decades:

the record of how human beings actually did their jobs.

The law is now being asked what happens when that record becomes valuable enough to sell.

References and Further Reading

Primary Bankruptcy Records

Spirit Aviation Holdings — Notice of Auction Results and Google/Mercor Sale Agreements, ECF No. 1463 The central primary source. Identifies Google as the $10 million successful bidder, Mercor as the $7.5 million alternate bidder, defines the de-identification requirements and contains the detailed asset schedule.

Association of Flight Attendants-CWA — Limited Objection to Proposed Spirit Data Sale, ECF No. 1489 The union’s actual court filing. Particularly important for the distinction between de-identification and confidentiality and for its argument concerning preserved referential integrity.

AFA-CWA — Spirit Bankruptcy: Objection to Sale of Your Data Union update explaining the status of the objection and confirming that the proposed transfer had not yet occurred.

Bankruptcy and Privacy Law

11 U.S.C. § 541 — Property of the Estate Defines the property interests entering a bankruptcy estate and explains the principle that the estate generally receives only the debtor’s own interest.

11 U.S.C. § 363 — Use, Sale, or Lease of Property Governs sales of bankruptcy-estate property and contains the special provisions addressing certain personally identifiable consumer information.

11 U.S.C. § 101 — Bankruptcy Code Definitions, Including “Personally Identifiable Information” Shows why the Code’s specialized bankruptcy privacy definition is oriented toward information supplied while obtaining consumer products or services.

11 U.S.C. § 332 — Consumer Privacy Ombudsman Explains the bankruptcy mechanism for appointing a consumer privacy ombudsman in qualifying data-sale proceedings.

California Privacy Protection Agency — CCPA Frequently Asked Questions Authoritative California guidance confirming that CCPA privacy rights extend to California residents who are employees and job applicants.

Historical Context

Federal Trade Commission — Toysmart.com Bankruptcy Privacy Settlement A landmark example of a failed company attempting to monetize a customer database despite prior privacy promises.

Carnegie Mellon University — Enron Email Dataset Documents the roughly 500,000-message Enron corpus that became an important real-world research dataset after the company’s collapse.

Current Reporting and Pending Disputes

Reuters — Court Delays Google-Spirit Data Sale Hearing After Union Objection Confirms the September 9 hearing date, Google’s AI-related purpose and the flight attendants’ concern about reconstructing information from linked data.

Business Insider — Micro1 Challenges Google With a Late $12.5 Million Offer Primary reporting on Micro1’s post-auction proposal and the uncertainty surrounding whether a bankruptcy court would entertain a late offer.

Bloomberg Law — Spirit Data Sale to Google Prompts Springshot Ownership Objection Reports Springshot’s allegation that the proposed dataset includes intellectual property it owns under a 2022 software agreement.

Fast Company — Why Spirit Airlines’ Internal Data Has Become Valuable to AI Companies Useful context on the emerging commercial market for real-world enterprise datasets.

Editorial currency note: This is an active bankruptcy proceeding. Auction status, bidder terms, objections and the September 9 hearing can change. The article should be reviewed immediately after any new substantive docket filing and updated after the September 9, 2026 hearing.