Google Buys Bankrupt Spirit Airlines' Data for AI Training

Google's purchase moves through a court-supervised bankruptcy sale rather than a scrape of the open web - a distinction that matters after Anthropic agreed to pay $1.5 billion to settle a lawsuit over training Claude on pirated books ^[1]. Forbes contributor John Werner framed the Spirit deal in exactly those terms, arguing real-world enterprise records have 'no substitute' and that buying them with clean legal title sidesteps the litigation risk that hit AI labs relying on scraped or pirated content ^[2]. That clean title still has to clear a judge: the sale runs through the US Bankruptcy Court for the Southern District of New York, where Judge Sean Lane must approve it before Google gets access to anything ^[3].

The headline number is small next to what it buys: roughly 100 million employee emails, about 500 million Microsoft Teams messages, and around 30 million lines of internal code and software ^[4], plus 7.2 billion records tracking competitors' flight pricing and 7.5 billion passenger transaction records dating back to 2008 ^[3]. Google's own opening bid was around $5 million; Mercor, an AI data-labeling company, pushed it to $7.5 million before Google closed the auction at $10 million ^[5]. The Register framed the value differently: this kind of domain-specific operational data - aviation pricing, crew scheduling, maintenance records, customer-service logs - is prized less for training another general-purpose chatbot than for building narrower, more specialized AI models ^[6].

Google says a third party will scrub the dataset of personally identifiable information before it ever reaches Google, and that passenger and loyalty records - about 97.5 million passenger profiles and up to 52 million Free Spirit accounts - are excluded entirely ^[4]. The Association of Flight Attendants-CWA is not convinced that scrubbing is enough: its bankruptcy-court motion argues a pseudonymized dataset can still reveal which crew bases generated grievances, who underwent disciplinary investigation, or what employees said about management and their union ^[7][12]. That skepticism turned out to be a cross-platform reaction rather than a niche complaint. Reddit discussion leans openly mocking, with a recurring view that de-identification at this scale is largely symbolic, countered by a smaller camp welcoming a licensed, court-approved dataset over the scraped and pirated training data that has landed other AI labs in legal trouble. Social video coverage of the deal ranged from straightforward explainers of the mechanics to comedic dramatizations of the Google-Mercor bidding war that poked fun at the de-identification promise, alongside more adversarial framing of the union's objection. Even the news accounts that reported the purchase matter-of-factly leaned hard into the sheer scale of the data involved, and at least one outlet's own social account pushed a sharper privacy angle by citing the union directly. The dispute is not just rhetorical - it already pushed the bankruptcy court's sale-approval hearing from August 19 to September 9 ^[8].

Selling internal company data out of bankruptcy has a paper trail. The FTC intervened when Toysmart tried to sell customer data in 2000 despite promising never to share it, and again in RadioShack's 2015 bankruptcy, in both cases narrowing what could actually be sold ^[9][10]. Online discussion of the deal has also drawn an informal parallel to the Enron email corpus - roughly $10,000 for 600,000 emails - though Spirit's haul is a far larger and more varied mix of emails, chats, and operational records than that single archive.

Coverage of the deal frames it as more than a one-off curiosity: a clean, structured internal dataset drew real competitive bidding, which signals to other distressed companies that their internal records - not just their planes, gates, and leases - can be monetized in bankruptcy ^[11]. A Mercor spokesperson made the underlying logic explicit, arguing that 'companies are sitting on decades of records that show how real work gets done,' and that this operational archive material is now some of the most valuable input for training and evaluating AI models ^[4]. Whether more bankrupt companies follow Spirit's path may depend on how the AFA-CWA's privacy objection is resolved when the court reconvenes on September 9 ^[8].