Google paid $10M for a dead airline's paperwork, and the flight attendants object

Google won a bankruptcy auction on Friday for a vast trove of Spirit Airlines company records to train AI models, bidding $10 million and beating AI training startup Mercor's $7.5 million ^[1][2]. On Tuesday the Association of Flight Attendants filed an objection in the U.S. Bankruptcy Court for the Southern District of New York, and the Wall Street Journal reported Wednesday that the court postponed approval of the deal ^[4][14].

Two things are now established. First, when a company dies, its filing cabinets have a market price: at least two bidders competed, and the winning bid sat $2.5 million, or about 33 percent, above the underbid ^[2][16]. Spirit shut down all operations earlier this year after a second bankruptcy filing ^[8]. What is being sold is not a customer list. According to Ars Technica, the dataset excludes personal and customer information and instead nearly covers the airline's entire employment and workplace record ^[17]. Friday's filing lists employee business travel records, crew training records, employee tax forms, and 100 million emails ^[6].

Second, the privacy machinery attached to the deal points at the wrong population. Google agreed to work with a third-party deidentification agent, following the standard in the California Consumer Privacy Act, so that nothing traceable to Spirit customers reaches the training data ^[3]. Per Ars Technica, a court-appointed ombudsman would oversee the stripping of personally identifying information, Google would keep the data de-identified and never intentionally re-identify it, and any third party buying access would be bound by the same terms ^[5]. That is a real set of commitments. It is also, as AFA lawyer Charles Rubio put it, built for customers, with "no comparable screen" applied to the employment record the transaction actually conveys ^[7]. The union's summary is blunter: "The privacy architecture of this transaction is consumer-facing; its payload is disproportionately employee-facing," and so the employee data is more confidential yet less protected than the customer data ^[9].

Rubio's argument is that removing names does not solve it. A pseudonymized dataset, he writes, can still reveal which crew bases generated grievances, how a small subset of flight attendants performed on recurrent training, which employees were investigated, what compensation adjustments followed which events, and what employees said to each other about management, staffing, or their union ^[12]. Workers fear confidential material could still be used to link them back to the scrubbed records ^[11]. The AFA wants the deal rejected unless it expressly and categorically excludes all flight attendant information ^[13]; the union says no comparable protections currently exist for the thousands of flight attendants who worked there ^[10].

The commercial logic explains why anyone paid. The internet's usable text has largely been harvested, pushing buyers toward creative sourcing, and synthetic data brings its own problems ^[18]. A full corporate record is a detailed picture of day-to-day life inside a modern business, which is the kind of material that makes AI tools useful to enterprise customers ^[15]. Google has also been losing AI talent to competitors ^[19].

What to watch: whether the court lets the CCPA standard stand as sufficient for a dataset that is mostly employment records, or requires a separate worker-confidentiality screen. That ruling sets the template for every distressed-asset sale that follows, because the next seller will read this docket before pricing its own email archive.