What would a company like Google want with data from a failed airline?
Google won a bankruptcy auction this month with a bid to spend $10 million on the corporate data of Spirit Airlines, which stopped flying in May. Google’s offer beat a $7.5 million bid for the data from an AI data company, Mercor.
The cache includes 100 million emails and 500 million Microsoft Teams messages, and around 30 million lines of code.
AI companies have been battling it out this year to make the best AI coding agent—and they have benefited from an abundance of publicly available code to train these models. But teaching agents to do other kinds of white-collar work may require data that is largely private, buried in corporate emails and chats. If that effort succeeds, it could send ripples across the labor market.
“Buying Spirit Airlines—that’s not centrally a coding agent purchase,” says Nick Heiner, the head of reinforcement learning environments at Surge, a data company that supplies the frontier AI labs. “That’s a purchase if you believe that your agent is going to generalize into the rest of the economy.”
Reinforcement learning
One of the biggest trends in AI over the last 18 months or so has been the rise of reinforcement learning from verifiable rewards—a method of AI training that has led to rapid improvements in models that can work independently for long periods of time.
Anthropic alone has discussed spending more than $1 billion in a year on so-called reinforcement learning (RL) environments, which are essentially simulations of commonly used software, where AI agents can take strings of actions independently. If these actions lead to a desired outcome, the AI is rewarded and will be more likely to take similar actions in the future.
Until now, the biggest jumps from this type of training have come from coding models, mostly because code has a useful property: it either works or it doesn’t, meaning that the reward signal is immediate, so improvement can happen in a fast loop. (It’s also helpful that there was plenty of coding data already out there on the internet, meaning models were good coders to begin with.)
But AI companies’ long-term goal is to automate large swathes of the economy. That’s where Spirit’s data likely comes in.
What makes the data useful
RL environments are only as good as the data that populates them, says Heiner of Surge AI. Companies like Surge and Mercor often hire human workers who are tasked with populating these environments with realistic data, either from scratch or in partnership with AI tools. “But even that is a little bit removed from literally having actual data that was used in the real world,” Heiner says. “That’s where deals like Spirit come in.”
Even though Spirit was a failed company, its data probably still encodes many useful things about how humans carry out their jobs, like how workers collaborate toward shared goals, or the norms of how the aviation industry operates. Even data depicting bad decisions can be helpful, Heiner says. “Maybe you see something where they’re actually doing something poorly, and you build a task out of that, and you change what they were doing to be something better.”
But unlike coding, which can be quickly and easily determined to either work or not, office tasks are fuzzier. Human experts can still provide that feedback—a service that companies like Mercor and Surge offer to AI companies—but that work is relatively expensive, time-consuming, and subjective. It’s therefore unclear that even populating RL environments with real-world data like Spirit’s will allow AI companies to move into white-collar fields as quickly as they have ripped through the software industry.
Privacy drama
A Google spokesperson declined to comment on the specifics of how Spirit’s data would be used, but said the data “can be helpful in improving our products and AI models.” The spokesperson added that personal information would be “rigorously scrubbed” by a third party before the dataset’s delivery to Google, and would include no customer data.
After Google submitted its bid, a union of flight attendants filed an objection, saying that although the deal’s terms would result in personal data being removed, the dataset would maintain “referential integrity,” meaning that links between one type of data and another would be kept. That integrity is what makes the data useful for training AI, but the union said it may also allow for even anonymized information about its members to be reconstructed. The group is seeking additional privacy protections; a judge is set to rule on the matter on Sept. 9.
Another AI data company, Micro1, submitted a higher bid of $12.5 million after the auction closed.
Whichever way the deal goes, it looks like the AI industry is setting its sights on white-collar work. “Google is willing to pay $10 million for a failed budget airline, because if you believe this is really generalizable, you see a path to all these different industries—not just the ones that Spirit was involved in—being something you can replace pretty quickly,” Heiner says.
The post Why Google Bid $10 Million for a Failed Airline’s Data appeared first on TIME.
