Google Is Buying 500 Million Employee Messages as AI Training Data for $10 Million

Google is paying $10 million for Spirit Airlines' internal data, and 500 million Microsoft Teams messages written by Spirit employees are inside the package. So are nearly 100 million passenger names, roughly 13 million active email addresses, and about 176,000 employee records. All of it is being bought as AI training data. Google says the records will be deidentified before any model reads them, and the Association of Flight Attendants-CWA has filed a court objection making a narrower point: the deidentification preserves referential integrity across the data set, and with 5,500 Spirit flight attendants inside it, individuals and small groups could be reconstructed from records that no longer carry anyone's name.

Sara Nelson, international president of the Association of Flight Attendants-CWA, said it directly.

This is outrageous! We are filing a court objection to Google's attempt to buy data that has no business being sold.

The filing itself is narrower than the quote. AFA-CWA, which represents more than 5,500 Spirit flight attendants, is asking the court for restrictions on the sale of flight-attendant employee data and for added protections for workers if the sale proceeds.

The court moved the hearing so the objection could be heard

The sale was scheduled for approval on Wednesday, August 19, and after AFA-CWA filed, the court adjourned to September 9 so the objection can be argued on the record. A federal bankruptcy judge decided a workforce privacy question was worth three additional weeks, in a proceeding where the pressure to convert assets into cash never lets up. Credit where it belongs: the adjournment is the process functioning as designed.

The bankruptcy code anticipated part of this. Section 332, added in 2005, requires a court-appointed consumer privacy ombudsman whenever personally identifiable information is sold in a bankruptcy, and one is overseeing the deidentification here. A rule exists, an independent overseer is assigned to it, and the transaction is being scrutinized rather than approved by default.

Here is what the $10 million buys.

A 2005 privacy rule is carrying a 2026 question

Section 332 was written when the asset in a data sale was a customer list: names and purchase histories, sold to a buyer who wanted to mail catalogs. Spirit Airlines is selling something different in kind. A corpus of 500 million work messages carries the recorded behavior of an organization, how people argued, escalated, apologized, and solved problems in their own words. That is what makes it useful to Google, and what makes reconstruction a live concern for AFA-CWA.

The AI data privacy question underneath this sale is easy to state and hard to answer. 500 million messages were written by people who had no idea they were producing AI training data, because at the time nobody was buying data for that purpose.

We talk about the half-life of skills, the shrinking interval between learning something and having to learn it again. Rules have a half-life too. A provision built for one kind of transaction is now the closest instrument available for a very different one, and a single judge in a single case is being asked to settle something that reaches well past Spirit Airlines.

I watched a version of this at Napster and Liquid Audio. Digital distribution arrived years ahead of the rules written for it, and everyone in that fight was following the rules that existed. The rules that existed were about physical copies.

Every party in this case is following the rules that exist

Google is buying data offered in an open court process, and it outbid Mercor's $7.5 million to win it. A Google spokesperson says the company will not receive any personal information from the data set, and that any data it receives will be scrubbed of personally identifiable information by a third party first. Spirit Airlines, operating in Chapter 11, owes its creditors an effort to get real money for its assets, and its data has real money in it. AFA-CWA is using the one venue open to it, which is the sale hearing. All of that is legitimate under current rules.

Nobody has answered the underlying question anywhere: whether the internal communications of a company in bankruptcy should be available as AI training material at all, and under what conditions. So it is being answered here, one docket at a time, by a judge handling it carefully.

The question worth asking before someone asks it for you

If your company were sold tomorrow, what would be in the package? Most executives can name the customer database, and considerably fewer can describe what 10 years of internal chat and recorded calls would be worth to a buyer training a model, or what their own employment agreements permit. That inventory is a governance exercise, and it costs one meeting.

If you sit nowhere near that decision, the question still belongs to you. Ask what your employer's policy says about internal communications in a sale or a merger. Most of us have never read that provision, and the answer usually sits in the handbook we signed on day one.

Now multiply this case out. Every organization running Teams or Slack holds a data set with the same properties, and bankruptcy is one of several doors through which it can change hands. Mergers and asset sales move data under rules written when data meant records rather than training material. Most of the people whose words sit in those data sets have no idea they are in there.

This is the kind of question where practitioners can be useful early, while a rule is being drafted rather than tested in a courtroom. I have written before about governance structures moving slower than the commercial pressure around them, both when OpenAI dissolved its own catastrophic-risk oversight team ahead of its IPO and when that same company paused a model release under its own framework. The through-line across the AI governance work is steady: the institutions holding the pen are working with instruments built for a slower world, and the people closest to the technology can help draft the next version rather than critique it later.

September 9 will produce a ruling about Spirit Airlines, and the broader policy question stays open after it. So here is one to carry into your next leadership meeting: if a buyer wanted our company's data tomorrow, whose words would be in it, and did we ever tell them? I have no settled answer for where that line belongs. I would rather we draw it deliberately than discover it in a docket. I am easy to find.

Here is what makes Alex a credible voice on this topic: Alex advises the California State University system on AI governance, where the work is writing usable rules for technology that keeps moving after the policy is approved.

For more, book a conversation →