explainx.ainewsletter3.5k
TrendingNewsPathwaysSkills
Pricing
explainx.ai

Upskill in AI — 16 free pathways, live workshops & bootcamps, and 50+ courses from practitioners. Plus the skills, tools, and MCP servers to practice on.

follow us

corporate training

support@explainx.ai

get started

Find your pathTake Free Evaluation

learn

pathways — start freeworkshopsbootcampscoursescertificationsmock testsexplainx universitycorporate traininglearn skills & mcp

discover

skillsmcp serversexplainx mcptoolsagentsllmsdesignsdictionaryagi trackerranks

company

aboutvisionmissionteaminstructorscommunityhackathonscareers

content

daily AI newsstate of AI — live resultsblogreleasespromptsgeneratorsresource libraryfor LLMsexplainx.ai kids

solutions

all solutionsdeveloper upskillingmarketing upskillingproduct manager upskillingleadership upskilling

newsletter · weekly

Get AI news, tools, and insights in your inbox.

supportprivacytermsdata rightshow we create contentsubmission guidelines

© 2026 AISOLO Technologies Pvt Ltd

On this page

  • TL;DR
  • Included vs Not Included (the table The Register flattened)
  • What people are asking
  • What this means if you build with AI
  • Honest limitations
  • Related on explainx.ai
← Back to blog

explainx / blog

Google’s $10M Spirit Data Auction: What the Docket Actually Includes

Google won Spirit’s deidentified data auction for $10M. The Register listed 30M call recordings; the docket marks those Not Included. Here’s the schedule.

Aug 18, 2026·9 min read·Yash Thakker
Training DataGooglePrivacyBankruptcyAviation AI
go deep
Google’s $10M Spirit Data Auction: What the Docket Actually Includes

The Register’s headline is true. Several of its counts are not. Google LLC won a $10 million bankruptcy auction on August 14, 2026 for Spirit Aviation’s Deidentified Data. The underbidder, Mercor.io Corporation, was $7.5 million, not a rounding error. A sale hearing is set for August 19 in the Southern District of New York (Case No. 25-11897, notice ECF 1463). Until Judge Sean H. Lane signs an order, Google has a winning bid, not a hard drive.

What the docket actually does is more interesting than the puns: it is a two-column asset schedule. Customer call recordings, chat sessions, and 13.7 million marketing email addresses sit in Not Included. One hundred million work emails, five hundred million Teams items, yield-management tables, and ~30 million lines of Spirit-owned code sit in Included. If you only read the viral summary, you were trained on the wrong mix.

The primary source is the auction-results notice and bill of sale (also on Epiq’s Spirit docket). Google told reporters the lot “can be helpful in improving our products and AI models,” and that it is not buying customer or credit-card information.

Weekly digest3.5k readers

Catch up on AI

Curated AI updates on agents, skills, and MCP — delivered to your inbox. Unsubscribe anytime.

TL;DR

table · 2 cols
QuestionAnswer
What happened?Virtual auction August 14, 2026 for Spirit’s Deidentified Data
Winner?Google LLC — $10,000,000
Backup?Mercor.io Corporation — $7,500,000
Closed?No. Hearing 11:00 a.m. ET, August 19, 2026
Call recordings in the Google lot?No — ~30.9M recordings and ~15.8M chats are Not Included
Passenger CRM in the Google lot?No — 97.5M profiles, loyalty, 740k cardholders Not Included
What is in?Emails/Teams/M365, ops, pricing, PNRs/transactions (deidentified), HR extracts, 516 repos / ~30M LOC
Scrub?Third-party Deidentification Agent Google designates and pays; CCPA (+ HIPAA if needed); keep join keys
Re-ID promise?Google publicly commits not to intentionally associate the set with a person or household

Included vs Not Included (the table The Register flattened)

Figures below are from the Assets Schedule attached to Google’s form bill of sale. “Not Included” is not a rumor; it is a column.

table · 3 cols
AssetVolume (docket)Google lot
Work emails100,000,000 (80,000 accounts)Included
Teams500,000,000Included
OneDrive17,082,644Included
SharePoint20,577,677Included
ServiceNow tickets667,563Included
Source repos516 repos, ~30M LOC; 43,170 PRs; 372,585 commitsIncluded
Flights (Movement Manager)763,391Included
Crew pairings5,014,676Included
Fuel slips1,239,196Included
Parts received787,452Included
Competitor fare observations (Infare)7,250,630,887Included
Navitaire transactions7,510,221,520Included
PNRs190,312,864Included
Wi-Fi sales (Thales codes)~11.2M codesIncluded
Employee records175,658 (UKG from 1986)Included
Payroll records3,426,618Included
Employee tax forms148,018Included
Passenger profiles97,500,000Not Included
Free Spirit members50,200,000Not Included
Co-branded cardholders740,000Not Included
Oracle Responsys active emails13,700,000Not Included
Handled call recordings30,865,471Not Included
Handled chat sessions15,784,473Not Included
Phone numbers7,341,857Not Included

That last block is exactly what most passengers worry about — and it is not the Google lot. The first block is what employees, contractors, and anyone who mailed spirit.com should worry about.

Wi-Fi codes and deidentified PNRs are still in. Those are operational traces, not a CRM dump, but they are also the kind of high-cardinality leftovers re-identification papers love.

What people are asking

How is this even legal?

US bankruptcy sales under § 363 transfer estate assets free and clear of most claims, after notice and a hearing. Spirit’s bill of sale is explicit that Personal Data is carved out of the Assets, then everything else is run through a Deidentification Agent before Google takes delivery. Google pays that agent; the $10 million purchase price is not reduced by scrubbing costs.

This is not GDPR-by-default. It is New York bankruptcy plus a contractual CCPA deidentification standard “regardless of whether such statute does, or does not, apply.” If you wanted a European-style purpose limitation (data collected for flying cannot later train Gemini), this docket is the counterexample.

Does “deidentified” mean I cannot be found?

The contract’s own language is the tell. Deidentification must remove or transform elements so the data cannot reasonably be associated with a particular consumer, while preserving referential integrity across the data set. Join keys stay. That is how you still train on “this ticket, this delay, this refund, this aircraft” as one trajectory.

Google also publicly commits (i) to keep the data deidentified and (ii) not to intentionally associate it with a person or household. It may on-sell Deidentified Data to others who sign the same clause. “Not intentionally” is not “cannot.” Hacker News’s first long comment was stylometry: writing style as a fingerprint even after names are stripped. That concern is technically real. It is also the class of risk the contract pushes onto Google’s promise, not onto a zero-knowledge proof.

Yahoo’s old “anonymized” search logs and medical de-identification failures are the historical prior. This filing cites CCPA and, for health-like fields, 45 C.F.R. § 164.514. Certification is to Google’s reasonable satisfaction, not to an independent public auditor named in the notice.

Why buy a famously messy airline?

Because training data for how work actually happens is scarce. Public web text is blogs and docs. Spirit’s Included set is IRROPs recommendations (3.0 billion rows in ITS), ServiceNow, sap-ish finance, CrewTrac pairings, git review threads, and a decade of internal email. That is closer to an enterprise simulator than to a FAQ scrape.

Same week, Razorpay Vulcan is the living-company version of this idea: pretrain on proprietary operational events. Google is buying the corpse version — one org’s full trace, once, AS IS, no warranty of accuracy (Section 7 of the bill of sale is a classic “WHERE IS / WITH ALL FAULTS” dump).

The ~30 million lines of code, 372,585 commits, and PR comment threads are a private cousin of The Stack: not 5 trillion open tokens, but airline-specific software with the human review text still attached.

Will Gemini start denying my boarding the way Spirit did?

Not from this lot in any simple “upload PNR → refuse you” way: customer profiles are Not Included, and Google promised not to buy credit-card files. The more plausible product use is domain post-training — customer-ops agents, disruption recovery, pricing, and code models that have seen real MRO and Navitaire-shaped schemas.

If a future Gemini answer ever regurgitates a unique complaint thread from a Spirit inbox, that is the SNAFU The Register is waiting for. The legal theory will be “deidentified + no intentional re-ID,” not “we never ingested workplace mail.”

What this means if you build with AI

1. Bankruptcy is now a data market. When a company dies, collaboration suites and git become inventory. If you are designing a data-retention policy, add “363 sale to a frontier lab” next to “e-discovery” and “subpoena.”

2. Read the Included column. Viral coverage will keep flattening schedules. For this story the correction is the product: calls out, emails in.

3. Referential integrity is the training feature. A pile of isolated, shuffled sentences is weak. Linked tickets → aircraft → crew pairing → refund is how you get long-horizon trajectories for agents. The contract requires keeping those links.

4. Deidentification is a vendor you do not pick. Spirit delivers to an agent acceptable to or designated by Google. Builders evaluating “safe to train” third-party dumps should ask who paid the scrubber.

5. License tables still matter elsewhere. Getty’s OpenAI deal is the consent-and-check path. This is the estate-sale path. Both feed models. Only one looks like a marketplace with a price list.

6. Keep personal compute off the corporate tenant. The bill of sale’s M365 clause is to transfer email, OneDrive, SharePoint, and Teams in the native Microsoft 365 environment. That is the whole corpus, then scrubbed — not a curated export of “non-personal” folders.

Honest limitations

  • Hearing not held as of this writing. Terms can still move.
  • Schedule OCR from the PDF is messy (typos like “Reponsys,” “Sharpoint”). Volumes are Spirit’s own counts.
  • Privileged material is carved out with a clawback; plenty of “legal memos” are still listed as Included — expect fights at the edges.
  • Spirit may still sell a customer list (traveler spend by year) to hospitality/travel buyers. That is a different lot than Google’s.
  • We have not seen the Deidentification Agent’s name, methods, or error rates.
  • Google can transfer Deidentified Data onward under contract. Downstream use will not all be public.

Related on explainx.ai

  • Hanover Institute — retrieval-time GEO farms vs buying a training dump
  • Razorpay Vulcan — live payments foundation model — operational pretraining while the company still exists
  • The Stack v3 — 5T open code tokens — the public-code analogue to Spirit’s 30M private LOC
  • Getty × OpenAI — licensing instead of a fire sale
  • Claude shared chats indexed by Google — another Google-shaped privacy surprise, different mechanism
  • Cloudflare: AI crawlers vs human traffic
  • EU AI Act after enforcement — GPAI training-data duties this US sale does not have to meet
  • How AI text watermarking works — stylometry is the other direction of “who wrote this”
  • Deidentification · Training data · PII

Sources

  • S.D.N.Y. notice of auction results + Google/Mercor bills of sale (ECF 1463, Aug 14, 2026)
  • Spirit Airlines bankruptcy claims / docket portal (Epiq)
  • The Register — viral write-up this post fact-checks
  • Contemporaneous reporting: Bloomberg Law, Skift, Business Insider, Axios (Google spokesperson: internal data and custom software, not customer or credit-card files)

Auction results and asset volumes are from the August 14, 2026 court notice and attached form agreements. The sale remains subject to Bankruptcy Court approval at the August 19, 2026 hearing. Deidentification quality cannot be verified from the docket alone. This is not legal advice and is not affiliated with Google, Spirit, or Mercor.

Spotted something out of date? Let us know.
Yash Thakker

Written by

Yash Thakker

Yash is an AI expert with over 300K learners. Join his workshops →

Related posts

Aug 15, 2026

Google HEIR: A Compiler for Running AI Inference on Encrypted Data

Google open-sourced HEIR, a compiler that lowers the barrier to fully homomorphic encryption from "needs a cryptography team" to "point it at your model." It lets a server run AI inference on encrypted data without ever seeing the plaintext — but Hacker News's practitioner reaction is a useful reality check on how far this is from production-speed.

Aug 18, 2026

How to Disable AI Features in Windows, Chrome, Zoom, and More

Librarian Jessamyn West published a practical, Creative Commons-licensed reference for turning off AI features you did not ask for — Windows Copilot, Chrome's Gemini flags, Zoom AI Companion, Gmail smart features, and more. It hit #1 on Hacker News. Here is the accurate, condensed version for developers and practitioners who want intentional control over which AI features run on their machines.

Aug 17, 2026

Google's Generative UI: Gemini Now Builds a Custom App for Every Prompt

Google Research's new generative UI implementation has Gemini 3 write and render a fully custom, interactive web interface for any prompt — not a templated app, code generated fresh every time. It's live in the Gemini app's "dynamic view" and in Google Search's AI Mode. Here's the actual system architecture, and what it means for anyone building AI products.