Spirit Airlines’ Proposed Data Sale to Google Turns Privacy Review Into a Customer Check That Leaves Out Employees

MIT Technology Review reported on October 13 that “tech undertakers” are selling what failed startups leave behind to AI developers. Founders who shut a company down can now sell its code and its workspace data (the contents of tools such as Slack, Notion, Google Drive and Figma) to buyers that include AI labs. In the story, Keelan Miskell netted roughly $25,000 for a crypto startup’s code base and expects around $15,000 more for its workspace data. The largest deal it names is far bigger: Google’s winning $10 million bid for Spirit Airlines’ archive in that airline’s bankruptcy, which had not yet been approved by the court in the latest filings I could read.

Spirit’s sale is the deal in that story that has a public court file, so I read it. That meant the sale notice and the agreement attached to it, the objections filed by unions and by a business partner, and the October 5 report of the court-appointed Consumer Privacy Ombudsman, all in the bankruptcy docket (Bankr. S.D.N.Y., No. 25-11897). I read them next to SimpleClosure’s own announcements and the text of the Bankruptcy Code. The file shows what the MIT story only touches on. The privacy review that bankruptcy law provides looks at customers’ data. The ombudsman wrote that she did not investigate the risks to employees or third parties, and the only exclusions her report lists are passenger systems; corporate email that includes employee data stays in the sale. In a private startup wind-down like the ones SimpleClosure runs, there is no judge and no ombudsman, so the safeguard is whatever the seller and the marketplace decide it is.

What the MIT Technology Review story reports

SimpleClosure, a company that handles startup shutdowns, added data sales to its services in April through a platform called Asset Hub. MIT reports that Asset Hub has handled over 150 deals and expects to return over $10 million to startups by the end of 2026, and that 47% of SimpleClosure deals now include code or data sales, transactions that CEO Dori Yona has watched go from rare and “somewhat taboo” to increasingly routine. A rival, Sunset, values a dying client’s assets by head count, years in operation, volume of data and how the team used its tools. Its CEO, Brendan Mahony, uses a particle-board analogy: “This operational exhaust data is like the wood shavings that everyone thought had no value, but it actually has a ton.”

The buyers want more than documents. MIT says Surge AI’s Nick Heiner looks for archives that reflect a range of situations, “from disagreements to tricky judgment calls,” to build practice environments where AI agents rehearse tasks. Real archives avoid errors common in synthetic workplaces, such as dates that do not exist, but MIT adds that “the extent of performance gains remains unclear. No company has yet released tests comparing results of real-world data training with the alternatives.” MIT also notes a September Bloomberg report that SpaceX’s AI unit is weighing purchases of data from struggling startups.

On the law, MIT says this “could get iffy.” Attorney Greg Ewing calls such data “legally radioactive.” Emory law professor Ifeoma Ajunwa argues that “Employers should have express consent before selling workers’ data,” while noting there is no federal law banning such sales. And founders, MIT writes, have “little to no visibility into who is buying the data and what they plan to do with it.” One caution on the numbers: SimpleClosure’s own September 3 release (a syndicated copy of its Business Wire announcement) counts something different from MIT’s 150 deals, saying that “AssetHub has supported more than 350 companies,” so the two figures are not directly comparable.

What Spirit’s court file adds

The sale and its calendar

On Friday, August 14, Spirit’s debtors filed a notice of the results of an auction for what they call “deidentified data.” The winner was Google LLC at $10,000,000. The alternate bidder was Mercor.io Corporation at $7,500,000. Objections were due at 4:00 p.m. the following Monday. The hearing was first set for August 19 and has been adjourned repeatedly. A September 28 notice moved it to 11:00 a.m. on October 14 before Judge Sean H. Lane in White Plains, with replies due by noon on October 9. The docket also lists a notice of adjournment of a hearing on October 9, and other filings through October 12, that I could not read, so the date and the terms may have moved since the documents quoted below were filed.

The unions that objected say the timeline was tight. The International Association of Machinists and Aerospace Workers (IAM) and Transport Workers Union of America Local 570 (TWU), in a joinder to the flight attendants’ objection, say the sale “raises novel technological and legal concerns on very short notice.”

What is on offer

The Association of Flight Attendants-CWA (AFA) tallied the assets schedule in its limited objection, and I confirmed that each of its counts appears in the schedule attached to the August 14 sale notice (the version before the revisions described below). It lists about 100 million emails across 80,000 email accounts, 17,082,644 OneDrive items, 20,577,677 SharePoint items and 500,000,000 Teams items. It also lists 175,658 employee records going back to August 1986, 3,426,618 payroll records from June 2016, 148,018 employee tax forms, 1,092,000 time-card records from December 2012, and 5,014,676 crew pairings for a crew base of 4,600.

The consumer side looks different. The schedule marks customer profiles, loyalty data, call recordings and similar categories “Not Included.” The ombudsman’s report adds that the original plan also covered several systems that might include customer data, such as passenger transactions and name records, refunds, inflight purchases and WiFi sales, and that those have since been pulled out. She counts 97 million consumers who gave Spirit their data and about 190 million flights taken by Spirit passengers.

As a scale check, and my own arithmetic: $10,000,000 spread over the four Microsoft 365 counts above (637,660,321 items) comes to about 1.6 cents an item. That overstates the per-item price, because the same sum also buys code, finance and operations data.

What the ombudsman reviewed

Because the sale touches passengers’ information, the court has a Consumer Privacy Ombudsman, Lucy L. Thomson, who filed a report on September 8 and a supplemental report on October 5. The October report describes concrete changes. The parties revised the scope of the personal data in the sale, “excluded passenger databases from the sale to eliminate much of the Spirit passenger PII,” and hired Tonic.ai as a “Deidentification Agent” to scrub unstructured data such as corporate email. The excluded systems include passenger transactions and name records, refunds, inflight purchases and WiFi sales. Certified completion of the de-identification is a closing condition of the sale, and the ombudsman recommends that the certificate also give “the number of individual Spirit passengers in the unstructured data whose personal information was detected and deidentified.” Her conclusion is that “from a privacy perspective, approval of the sale of the deidentified data would be appropriate.”

The report also draws its own boundary. Two footnotes carry the same sentence: “The Ombudsman did not investigate or reach any conclusion about whether the data offered for sale or the deidentification process to be followed presents any risks to affected employees or third parties.” The body of the report says “Corporate emails that include employee data and possibly consumer personal data will be included in the sale.” Her footnote on the wider problem is blunt: “technology has far outpaced the law.” The report treats the buyer’s use as AI training. One section is headed “Reasonable and Appropriate Security in the AI Training Environment,” and it points to Google’s paper “Our Approach to Protecting AI Training Data,” whose abstract describes Google’s general data-protection practices rather than this dataset. She also asks for time to research Mercor’s early-2026 security incident, and micro1’s proposal to do the de-identification itself rather than use an independent party, if either becomes the buyer. Mercor’s own June 25 post says it “was affected by a supply chain attack involving LiteLLM” in late March.

Why the check stops at customers

The statute

The limit comes from the text of the Bankruptcy Code. Under section 363(b)(1), if a debtor told individuals it had “a policy prohibiting the transfer of personally identifiable information about individuals to persons that are not affiliated with the debtor,” the trustee may not sell such information unless the sale is consistent with that policy or a court approves it “after appointment of a consumer privacy ombudsman in accordance with section 332.” Section 332 then lists what the ombudsman may present: the debtor’s privacy policy, “the potential losses or gains of privacy to consumers,” “the potential costs or benefits to consumers,” and alternatives that would limit harm to consumers. And section 101(41A) defines personally identifiable information, in its first branch, as items such as a name, a home address, an email address, a home telephone number, a Social Security number or a credit card number, but only “if provided by an individual to the debtor in connection with obtaining a product or a service from the debtor primarily for personal, family, or household purposes.”

On a plain reading, a flight attendant’s payroll file does not fit that definition. I am not a lawyer, and this is a reading of the text rather than legal advice. The AFA reads the scheme the same way. It says it “does not contend that section 363(b)(1) is triggered here” and that “the statutory scheme the parties borrowed from was built for customers.” The ombudsman was appointed anyway, and her report treats sections 332 and 363 as the framework.

The agreement

The sale agreement follows the same consumer logic. It excludes “Personal Data,” defined as information that “relates to, describes, or is reasonably capable of being associated with a consumer.” It measures de-identification by whether data can be linked to “a particular consumer.” The parties volunteered the California Consumer Privacy Act’s standard “regardless of whether such statute does, or does not, apply to the Assets.” The agreement does not define “consumer” in the pages I read, so how far these protections reach employees is a question the AFA raised and the debtors’ reply, which I could not read, may address.

Four other terms matter to anyone whose messages are in the archive. De-identification must happen “while preserving referential integrity across the data set,” which the AFA reads as keeping the joins among timekeeping, pairing, training, payroll and communications data. The de-identification agent is a third party “acceptable to or designated by Buyer,” and it certifies “to Buyer’s reasonable satisfaction.” The buyer “shall have the right to transfer Deidentified Data to third parties,” provided the buyer contractually obligates them to the same commitments. And the agreement gives no other person “any legal or equitable right, remedy or claim.” The AFA says this leaves flight attendants with “no practical means of monitoring or enforcing compliance at any point in that chain.”

Identity is not confidentiality

The AFA’s objection is narrow by its own account: it “does not seek to disrupt the Debtors’ sale process, to unwind the Auction, or to prevent the estates from monetizing data assets.” Its argument is about what de-identification does. “Deidentification addresses whether a record can be traced to a named individual. It does not address whether the contents of the record are confidential.” It lists what stays sensitive after a name is removed: “A flight attendant’s disciplinary correspondence, a crew training deficiency, a leave or accommodation request, an internal Teams exchange about staffing or scheduling grievances, and a payroll adjustment history.” Its summary of the deal is “The privacy architecture of this transaction is consumer-facing; its payload is disproportionately employee-facing.”

The same gap shows up for a business partner. International Aero Engines (IAE) filed a limited objection and cross-motion saying the proprietary technical and commercial data it gave Spirit is “almost certainly stored on the Debtors’ computer and data storage systems.” IAE “does not oppose the Data Sale in its entirety,” but argues that “deidentification is insufficient to ensure that the IAE Proprietary Information has been removed.” Removing names does not remove a supplier’s technical data from an email, which is IAE’s point, any more than it removes the content of a worker’s grievance, which is the AFA’s.

One observation here is mine and not a claim from the filings. The material that AI buyers prize and the material the AFA calls confidential look like the same material. MIT says Surge AI seeks archives with “disagreements” and “tricky judgment calls,” and the AFA’s list of sensitive items includes grievances, discipline and accommodation requests. Both are candid records of friction at work, which is why it matters who reviews them.

Linkage is the feature and the risk

Buyers want the links between records, and the vendor says so. Tonic’s page on company data de-identification promises to “turn a wind-down workspace into a licensable asset,” and describes replacing every name, email and ID with a synthetic value that holds its references across the corpus, “so the person who opened a ticket is still the person who closed it.” That is the same property the AFA worries about. With a crew base of 4,600 described across linked datasets spanning more than a decade, the union says the risk that information about identifiable individuals or small groups can be inferred “is not speculative.” It adds that it “does not make a technical claim that any particular record can be re-identified, and it does not need to.” The IAM and TWU joinder argues that AI changes the economics, because the benefit of de-identification was “to make the juice not worth the squeeze.” The ombudsman’s footnote makes a similar point about AI making inferences by combining de-identified data with other sources.

On accuracy, the vendor’s number is a benchmark score. Tonic says that on its open PrivacyBench, which uses synthetic email and Slack datasets, it “scores 92.0% on the combined detection-and-replacement metric.” That is a score on a test set, not an error rate on Spirit’s mail, and I found no figure for Spirit’s data in the filings I could read. The ombudsman’s request for counts of detected passengers would put a real number on the table, though it covers passengers only. For a hands-on look at why a stable token that lets you find a record also lets someone link records, see our tutorial on encrypting PII with blind indexes, which measures what each design leaks.

The private route

A company that dissolves through SimpleClosure goes through state dissolution filings (the firm says it handles “all dissolution filings and state withdrawals”) rather than a bankruptcy case, so no bankruptcy judge approves the data sale. SimpleClosure describes the sale process in its AssetHub materials and its September 3 release. “With the seller’s approval, qualified assets are surfaced to SimpleClosure’s buyer network.” Data goes through “a multi-pass de-identification protocol designed to detect and replace personal and sensitive information, including names, emails, credentials, keys and customer identifiers.” Because SimpleClosure works directly with sellers, it “can also confirm chain of title and use purpose-built lists of affiliated individuals as an additional detection layer.” And: “Sensitive employee and health records, as well as any information sellers designate as out of scope, are not sold.”

That last sentence is a stronger employee rule than the one visible in Spirit’s schedule. But the product itself is workspace data, described as “Operational data including communications, workflows and documents that capture how teams worked, made decisions and built products over time,” and the seller is the company. I did not find employee notice or consent mentioned in the release or the blog post. SimpleClosure’s own October 1 guidance for founders states the problem plainly: “Company data is rarely just the founder’s information. Employees, customers, vendors, and investors may be represented throughout the material.” It tells sellers to “work out what you have the right to transfer” and warns that “If the buyer receives the raw files first and promises to clean them up later, you have already given that buyer access to the material you meant to remove.”

So the private route has a policy but no outside reviewer. Nothing I read about it involves a court, an ombudsman or a public record of the deal terms, and MIT’s point about founders’ limited visibility into buyers applies. That is not evidence that anything has gone wrong. It means the check is the marketplace’s own.

Two doorways for the same kind of data

Questions to ask before a dead company’s data is sold

Ask for the schedule, by system. In Spirit’s August 14 schedule customer profiles were out and “Team Member” records were in, and anyone could see the difference only because the schedule was attached to a public notice.

Ask whether the privacy review covers workers and partners, in writing. In Spirit’s case the ombudsman said she did not investigate employee or third-party risk.

Ask what the certificate will contain. The ombudsman wants counts of passengers detected. The same count for employees, and accuracy results on your own data rather than a vendor benchmark, would tell you more than a statement that data was de-identified.

Ask whether the buyer can resell and who can enforce the pass-through. Spirit’s agreement allows onward transfer and gives no one else a claim.

Ask whether links between records are kept and whether small groups can be picked out. A crew base of 4,600 is the kind of small population the AFA says is exposed when linked datasets stay joined.

If you are a founder, define the scope first. SimpleClosure’s own checklist says to look for employee details, health information, payment data and live credentials, and to avoid giving a buyer raw files before cleaning.

What is still open

The sale has not been approved in anything I could read. The October 14 hearing may have moved again, the debtors’ reply was due October 9, and the carve-outs the AFA and IAE asked for may or may not appear in a final order. No figure exists in the public filings I read for how well Tonic’s process works on Spirit’s mail. And MIT’s caveat stands: nobody has published tests showing that real company archives train better models than synthetic ones. MIT says data prices will likely stabilize in the coming year, “unless an AI bubble pops first.”

The bottom line

The MIT story describes a market. The Spirit file shows what a market’s safeguards look like when a judge is involved: a careful review of the customers’ data, written by an official who says she did not examine employee or third-party risk, plus objections from the workers and the business partner that the review did not cover. In the private channel the safeguard is a seller’s policy. Either way, the question for anyone whose emails sit in a dying company’s archive is the one the AFA asked in court: who is the check for?

Sources and method

I read MIT Technology Review’s story in full, and from the Spirit bankruptcy docket on CourtListener the notice of auction results with the sale agreement and assets schedule, the AFA’s limited objection, the IAE objection, the IAM and TWU joinder, the September 28 adjournment notice and the ombudsman’s October 5 supplemental report. I also read SimpleClosure’s AssetHub page and its September 3 and October 1 publications, Tonic’s company data de-identification page, Mercor’s incident post, and sections 101, 332 and 363 of title 11 on Cornell’s Legal Information Institute site. Every quotation was checked against the fetched text.

Limits: the docket lists filings I could not read, including the AFA’s September 24 supplemental objection, an objection by the Air Line Pilots Association on September 25, and a reply, an objection, an adjournment notice and other entries docketed on October 9 and later. I did not see the ombudsman’s September 8 report. Forbes, Business Wire and the UpStarts newsletter were not readable by my tools, and Bloomberg’s report on SpaceX’s AI unit comes only through MIT. I did not test any de-identification system. The per-item price is my own arithmetic on the AFA’s counts and the $10,000,000 price. Statutory readings are mine and are not legal advice.