Skip to main content

Google Agrees to Privacy Limits on Spirit Airlines AI Training Data

Google has agreed to extra privacy protections for the Spirit Airlines data it bought for $10 million to train AI. Employee health and medical records and other sensitive files are excluded, an independent agent chosen with the unions will check de-identification, and Google may not try to re-identify anyone. The flight attendants' union dropped its objection, and a bankruptcy judge still has to approve the sale.

On this page

What Google is buying from Spirit

Spirit Airlines halted operations in May and has been selling off its assets in bankruptcy. In August, Google won an auction for the airline's internal business data with a $10 million bid, beating the AI recruiting company Mercor, which offered $7.5 million. The archive includes roughly 100 million emails and 500 million Microsoft Teams messages, along with spreadsheets, calendars, software code, and other internal records. Google says it wants the material as an enterprise dataset to improve its products and AI models.

Customer data was handled separately from the start. A privacy ombudsman recommended excluding personal information on 97 million Spirit passengers. The fight since August has been about the employees whose working lives fill those inboxes and chat channels.

Why the flight attendants objected

The Association of Flight Attendants-CWA filed an objection in the US Bankruptcy Court for the Southern District of New York, arguing that the sale did not adequately protect employee data, which it saw as more sensitive than the customer records that had received stronger treatment. The union said it was not trying to block the sale, only to strip out identifying information. It also warned that the planned de-identification might preserve links across the dataset that could allow confidential records to be traced back to people.

Union president Sara Nelson noted that Spirit flight attendants had not been paid accrued vacation and sick leave, and called the data sale "adding insult to injury". Political pressure followed: about 120 lawmakers, led by Senator Elizabeth Warren and Representative Steven Horsford, raised concerns about the volume of employee communications involved.

The privacy terms Google accepted

Reuters reported on 10 November that Google had agreed to additional protections, clearing the way for a bankruptcy judge to consider the deal the following week. The union says the de-identification process is now more rigorous than first planned and that significant employee data, including health and medical information, is excluded. Reporting on the agreement lists drug testing information, disciplinary records, identification documents, and workplace safety records among the other exclusions.

An independent verification agent, chosen in consultation with the unions, will review the de-identified data before Google receives it. Google is barred from trying to re-identify the data and must process it in a secure environment and treat it as confidential. According to The420, the de-identification will follow ISO 27559, a standard for de-identifying personal data, and the court hearing is set for 17 November. Google had already said it was "not looking to buy any personal information from Spirit".

What this means for your own workplace messages

The new terms are real improvements, but they do not change the basic fact this deal established: years of workplace email and chat can be sold as an asset when a company fails, and AI training is now a buyer's use case. Employees wrote those messages to do their jobs, not to train a model, and de-identification of free text is hard because names, roles, and incident details are woven through ordinary conversation.

For individuals, the practical lesson is to treat anything typed into an employer's email or chat system as company property that could outlive the company. For businesses, it is a reason to check what retention policies and bankruptcy terms say about communications data, and to keep genuinely sensitive drafting and analysis in tools that do not store it on someone else's servers. Local, on-device AI is one way to keep that work out of a future dataset entirely.

CuriousLM runs supported AI models locally on your device. Try CuriousLM.