Lawmakers Urge Google and Spirit Airlines to Halt $10M AI Data Deal

A proposed $10 million data transaction between Google and defunct carrier Spirit Airlines has drawn intense scrutiny from federal lawmakers over the privacy and security risks of exposing sensitive corporate and employee records during artificial intelligence model training. The proposed Google Spirit Airlines data AI training deal would transfer hundreds of millions of internal enterprise communications and personnel records from Spirit Airlines to Google. The transaction highlights growing regulatory and security concerns surrounding the commercial monetization of enterprise datasets during corporate liquidations, particularly when those datasets are fed into large-scale machine learning systems.

More than 100 members of Congress, led by Rep. Steven Horsford (D-NV) and Sen. Elizabeth Warren (D-MA), sent a letter to the CEOs of both companies demanding an immediate halt to the transfer until privacy concerns and independent reviews are addressed.

The Scope of the Asset Sale

Spirit Airlines, which announced plans to shut down in May—resulting in nearly 1,000 job losses in Las Vegas alone—is seeking to sell a massive repository of historical enterprise data as part of its liquidation process. Google intends to acquire the dataset to train its proprietary foundation AI models.

According to details disclosed by Rep. Horsford, the proposed sale encompasses an immense dataset of sensitive internal communications and operational records:

• Approximately 100 million corporate emails
• Roughly 500 million Microsoft Teams messages
• Binding employment contracts and internal timecard records
• Detailed employee payroll and tax information

While Google has stated that the data will undergo third-party scrubbing and de-identification prior to ingestion, security and privacy experts—alongside the 114 signing lawmakers—warn that modern machine learning techniques render traditional de-identification methods largely ineffective.

Why Anonymization Fails in Modern AI Training

The primary risk in training foundation models on unstructured corporate data lies in the mechanics of data re-identification and model memorization. Google’s plan relies on conventional data scrubbing—removing direct personal identifiers such as names, employee IDs, phone numbers, and email addresses. However, in the context of modern generative AI and high-dimensional vector embeddings, removing surface-level personally identifiable information (PII) does not make a complex dataset truly anonymous.

Unstructured communications datasets containing 600 million combined emails and chat logs possess rich contextual relationships. Internal Microsoft Teams messages and email threads frequently contain implicit identifiers: unique project codenames, specific dates and times of operational incidents, organizational hierarchies, geographical references, and distinct writing styles.

When deep learning models are trained on unstructured chat logs, they risk experiencing model memorization. During inference, large language models can inadvertently regurgitate sensitive context when prompted with specific, related inputs—a vulnerability known as training data extraction or membership inference. A attacker or user interacting with a model trained on raw corporate archives could potentially reconstruct confidential business strategies, internal legal disputes, or private personal disclosures even if direct names were omitted during pre-processing.

Furthermore, cross-referencing scrubbed internal communications with publicly available data—such as LinkedIn profiles, public news reports, or regulatory filings—allows threat actors or researchers to execute re-identification attacks. By linking known public events to temporal sequences in the scrubbed Teams messages, third parties can re-identify specific employees and map their private internal conversations with high confidence.

Regulatory Guidance and Demands

In their letter, lawmakers stressed that conventional privacy safeguards are inadequate for modern artificial intelligence training pipelines. The coalition requested that if Google and Spirit Airlines proceed with the transaction, they must implement several mandatory security and governance controls:

• Establish a standardized de-identification protocol that explicitly incorporates direct feedback from former Spirit Airlines employees.
• Exclude sensitive employee records, including payroll data, tax documents, employment contracts, and granular workplace communications, from the transfer scope.
• Implement legally binding usage restrictions limiting how and where the acquired data can be processed or retained.
• Commission an independent employee confidentiality review by an external security firm prior to any data transfer.

Neither Google nor representatives from the defunct carrier immediately responded to requests for comment regarding whether they will pause the deal or adopt the recommended privacy frameworks.

Enterprise Governance and Data Lifecycle Risks

This incident highlights a broader security challenge for enterprise risk management: the lingering data trail left behind during corporate insolvency and asset sales. Corporate data governance policies must explicitly account for the entire data lifecycle, including business shutdown scenarios.

When organizations store years of unencrypted or minimally aggregated communications in cloud platforms like Microsoft Teams or Exchange, those archives become high-value liquidatable assets during bankruptcy proceedings. Organizations should enforce strict data retention and disposition policies, ensuring that sensitive internal communications are regularly sanitized, archived, or destroyed according to defined legal schedules rather than left sitting indefinitely in cloud tenants where they can be monetized during distress sales.

For tech vendors acquiring external training data, third-party risk management must go beyond basic string-matching scrubbing. Any AI vendor ingesting enterprise communications must implement differential privacy mechanisms, robust synthetic data generation, and comprehensive auditability controls to prevent training sets from becoming a permanent privacy exposure vector.