---
title: "A Bankrupt Airline, Midsize Firms and Gig Workers Now Sell A.I. Labs Their Data"
description: "Google bid $10 million for Spirit’s archive and Micro1 has $1 billion to spend, but the employees whose messages fill the files are paid nothing"
author: "rews special report"
published: 2026-10-18T14:00:01Z
modified: 2026-10-18T17:33:36Z
url: https://rews.cc/a/ai-firms-out-of-web-text-start-buying-your-slack-and-your-ch-edf48c
language: en
type: special report
tags: ["ai", "workforce", "automation", "data", "productivity", "tech"]
publisher: "Rews (https://rews.cc)"
---

# A Bankrupt Airline, Midsize Firms and Gig Workers Now Sell A.I. Labs Their Data

*Google bid $10 million for Spirit’s archive and Micro1 has $1 billion to spend, but the employees whose messages fill the files are paid nothing*

Special report · By rews special report · October 18, 2026 · https://rews.cc/a/ai-firms-out-of-web-text-start-buying-your-slack-and-your-ch-edf48c

## In brief

- Google won an Aug. 14 auction for Spirit Airlines’ internal records at $10 million, and a sale hearing is set for Nov. 6
- Micro1 offers companies $100,000 to $2 million for operational data and plans to spend $1 billion over 12 months
- 121 members of Congress, unions and EPIC say de-identification may not protect former Spirit employees
- Microagi, DoorDash and Objectways pay people to film household chores to train robots
- Harvey says its legal model Tenet was trained on public, synthetic and expert data, not client files

Google agreed in August to pay $10 million for the internal records of Spirit Airlines, about 100 million emails and 500 million Microsoft Teams messages left behind when the budget carrier stopped flying in May. A federal bankruptcy judge in New York is scheduled to decide on Nov. 6 whether the sale can close.

The Spirit auction is the most public deal in a market that grew mostly out of sight this year. A.I. developers built their first systems on text scraped from the open web. Now they are paying for what sits behind a company login: Slack threads, customer-relationship histories, support tickets, code repositories and video from a camera worn by someone mopping a kitchen floor. [Business Insider](https://www.businessinsider.com/ai-workplace-data-training-agents-robots) has spent this month documenting the change in a series it calls The AI Data Grab.

Most of the sellers are not the people who made the data. Midsize companies sell their operating records to brokers such as Micro1 and Polyshares for six- and seven-figure sums. Bankruptcy estates sell what is left when a company dies. Gig workers in Brooklyn and in a small town in southern India sell footage of their own hands. The employees who wrote the messages are almost never paid, and a privacy group that has filed in the Spirit case says it is the first time an employer has tried to sell its workers’ data as a bankruptcy asset.

## The data forecast

Epoch AI, a research group that tracks A.I. training trends, put numbers on the shortage in a [2024 analysis](https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data) led by Pablo Villalobos. It estimated the effective stock of public, human-written text at about 300 trillion tokens, with a 90 percent confidence interval running from 100 trillion to 1,000 trillion. It also estimated when frontier training runs would use it up.

> If trends continue, language models will fully utilize this stock between 2026 and 2032, or even earlier if intensely overtrained.
>
> — Pablo Villalobos and co-authors, Epoch AI researchers, [epoch.ai](https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data)

That window has already moved once. The group’s 2022 version of the paper put the date at 2024. The authors pushed it back after evidence that filtered web data trained models well and that text could be reused over several passes. In December 2024, Ilya Sutskever, a co-founder of OpenAI, told the NeurIPS conference that “we have but one internet” and that pretraining as the field knew it would end, [The Decoder reported](https://the-decoder.com/openai-co-founder-says-ai-is-reaching-peak-data-as-it-hits-the-limits-of-the-internet/).

By then the labs had found a second use for data. Over roughly the past year and a half, [Time reported](https://time.com/article/2026/08/25/google-spirit-airlines-ai-data-RL/) in August, the fastest gains have come from reinforcement learning, in which an A.I. agent works through a simulated version of a task and is rewarded when it gets the result right. Anthropic alone has talked about spending more than $1 billion in a single year on these training environments, the magazine reported. Code was the easy case because a program either runs or fails, and because the public web held plenty of it. Most office work is private.

## A price list for office work

Micro1 has been the most open buyer. It offers businesses $100,000 to $2 million for their operational data, plus a $50,000 bonus for referring another company that gets approved, [Fast Company reported](https://www.fastcompany.com/91622918/ai-companies-are-offering-businesses-up-to-2-million-for-something-their-employees-create-every-day). William Almeida, the company’s chief revenue officer, told Business Insider that the most valuable material is “long-horizon” data, records showing how work actually got done from start to finish. A company that holds back part of the set, such as its Slack messages or its email history, gets less for the rest.

![A hand lifts a bundle of message icons from a filing cabinet while another hand holds coins, suggesting the sale of stored communications.](https://rews.cc/img/caaa19143aa953e2dd60fc3a0cbd5337d39757b3.webp)

On Oct. 9, Micro1 said it would spend $1 billion over the next 12 months buying and licensing company data, with capital from Citi and Hercules Capital, [The Next Web reported](https://thenextweb.com/news/micro1-1bn-enterprise-data-ai-training-citi-hercules). The company says it turns de-identified records into reinforcement-learning environments that mirror real operations. Payment starts above $100,000 for a qualified partnership, rises above $500,000 for large datasets or work across several teams and passes $1 million for highly unique, proprietary data. Sellers need more than 30 employees, established documentation and material in English. American companies get priority.

Polyshares, a rival, [says on its website](https://polyshares.com/) that it licenses “the operating record a business produces by running” and supplies it to A.I. labs, in deals that typically run from $100,000 to $2 million or more. It presents itself as the buyer, not a marketplace, and promises an offer within one to two weeks. In the Business Insider investigation by Charles Rollet, Tom Carter and Stephen Council, some companies said they had received six-figure offers with 48 hours to decide, for data the outlet described as “priced by the terabyte, anonymized, and sold to AI giants desperate for an edge.”

Ali Ansari, Micro1’s founder and chief executive, has made the biggest claims for the trade. He has written on X that about 10 percent of U.S. gross domestic product could eventually come from corporate data and human expertise used to train A.I. In a January post, he argued that human data would become a $1 trillion-a-year market.

His arithmetic starts from about $50 trillion a year in global labor spending and assumes that 5 percent of it, about $2.5 trillion, goes to demonstrations, feedback and evaluations for A.I. systems. That is an assumption, not a measurement, and it comes from a company whose business depends on it. The prices on the public record are far smaller: about $60 million a year for Reddit’s posts and $10 million for an airline’s entire archive.

![A robotic arm lifts a sheet with geometric cutouts from an open book whose right side is a vault door with a keyhole.](https://rews.cc/img/2f9dabacc197e4190d176404215ca8a1a545cfd1.webp)

## An airline’s archive

Spirit stopped flying on May 2, in its second Chapter 11 case in two years. The asset schedule it filed in bankruptcy court lists about 100 million emails from roughly 80,000 accounts, 500 million Teams records and about 17 million OneDrive files, according to [an account of the filings](https://finance.yahoo.com/technology/ai/articles/google-paid-10m-spirit-airlines-152841454.html) published on Yahoo Finance. It also lists 516 code repositories holding about 30 million lines of software. [Skift reported](https://skift.com/2026/10/12/fighting-spirits-data-sale-to-google-court-filings-show-the-list-getting-longer/) that the package includes payroll records, Spirit’s pricing model and booking curves, and employee data going back to 1986.

Google opened the Aug. 14 auction at $5 million and won at $10 million. Mercor, an A.I. training-data company, was named the backup bidder at $7.5 million. Micro1 skipped the auction and later sent Spirit’s lawyers an offer of $12.5 million. Mr. Ansari told Business Insider that his bid was “materially higher” and that Google’s price was “actually quite low” for decades of records, [Startup Fortune recounted](https://startupfortune.com/micro1-bids-125-million-to-beat-google-for-spirit-airlines-data/). The Next Web dated Micro1’s formal notice of a competing bid to Sept. 3.

Per item, the numbers get small. Google’s bid works out to about 1.7 cents for each of the roughly 600 million emails and Teams messages, before counting the files and the code. Micro1’s comes to about 2.1 cents.

Google has said it will receive no personal information and that the data “can be helpful in improving our products and AI models,” a spokesperson told Time. On Oct. 5, Lucy L. Thomson, the court-appointed consumer privacy ombudsman, reported that Spirit and Google had dropped booking and transaction records, refunds, in-flight purchases and Wi-Fi sales from the deal. A de-identification company, Tonic.ai, would scrub the corporate email, which may still mention passengers. Ms. Thomson concluded that “the potential privacy losses or costs to consumers have been eliminated or mitigated,” [The Next Web reported](https://thenextweb.com/news/google-spirit-airlines-data-ombudsman-approval). Her review covered about 97 million customers. Her filing said the risks to employees fell outside it.

![A cream-colored passenger airplane with a black stripe and a pilot silhouette flies against a yellow background.](https://rews.cc/img/977411e98f30144f88485970084e5b52502f83da.webp)

Other material had already come out. The asset schedule excludes about 97.5 million customer profiles and 50.2 million records from Spirit’s loyalty program. Spirit had also agreed to leave out data on health, disabilities, drug and alcohol testing and collective bargaining, [Yahoo News reported](https://www.yahoo.com/news/politics/articles/spirit-airlines-wants-sell-data-100000712.html).

Not everyone in the archive knew it was for sale. Doug Kreuzkamp, chief executive of Springshot, which sells operations software to airlines, told Business Insider he learned during a Zoom call in August that Google had bought his company’s data. Springshot has since told the court that the dataset includes its intellectual property.

## The workers’ objection

On Oct. 8, 121 members of Congress, led by Senator Elizabeth Warren of Massachusetts and Representative Steven Horsford of Nevada, wrote to Spirit’s chief executive, Dave Davis, and Google’s, Sundar Pichai. Citing court findings, [the lawmakers said](https://horsford.house.gov/media/press-releases/spirit-airlines-to-sell-employee-data-to-train-google-ai-horsford-and-warren-lead-call-for-worker-privacy-protections) the sale would include personnel records, timecards, payroll and tax details and employment contracts, possibly with medical and accommodation information. They asked that employee data be left out as far as possible. “Spirit’s workers handed over this information because their jobs required it,” Mr. Horsford said in a statement. “In Las Vegas alone, nearly 1,000 people were laid off from Spirit.”

The Association of Flight Attendants-CWA filed the first objection. It argued that the scrubbed dataset would keep its “referential integrity,” the links between one kind of record and another, Time reported. Those links make the data useful for training, and the union said they could also let anonymized information about its members be pieced back together. The Air Line Pilots Association objected that a risk of pilots being identified could discourage voluntary reporting to confidential F.A.A. safety programs across the industry. The union for American Airlines pilots joined the opposition this month.

> Spirit Flight Attendants and other former employees are on the tip of the spear of this new threat to workers’ rights
>
> — Sara Nelson, president of the Association of Flight Attendants-CWA, [afacwa.org](https://afacwa.org/congress-spirit-data-letter/)

The Electronic Privacy Information Center and Seema Patel, a professor, represented by Harvard Law School’s Cyberlaw Clinic, [filed a brief](https://epic.org/epic-files-amicus-brief-challenging-spirit-airlines-bankruptcy-sale-of-employee-data-to-google/) supporting the flight attendants. They told the court that, as far as they knew, no employer had ever tried to sell its employees’ data as a bankruptcy asset. They wrote that approving the sale as proposed “would create perverse incentives for employers to over-collect employee data.”

The research on anonymity backs the objectors, with a limit. In a 2019 study in Nature Communications, Luc Rocher, Julien Hendrickx and Yves-Alexandre de Montjoye of Université catholique de Louvain and Imperial College London [estimated](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6650473/) that 99.98 percent of Americans would be correctly re-identified in any dataset using 15 demographic attributes. Their model was built on census microdata from Massachusetts. It measured tables of attributes, not years of free-form email, which no study has tested at this scale. A reader in a Hacker News thread on the sale raised the problem plainly.

## How the web closed

Private archives went on sale after four years in which public text got harder to take. In February 2024, Reddit disclosed in its IPO filing that it had signed data-licensing contracts worth $203 million over two to three years, [TechCrunch reported](https://techcrunch.com/2024/02/22/reddit-says-its-made-203m-so-far-licensing-its-data). Reuters reported that the deal with Google was worth about $60 million a year.

That July, a team led by Shayne Longpre of the Data Provenance Initiative published an [audit of 14,000 web domains](https://arxiv.org/pdf/2407.14933). It found that in a single year, about 5 percent of the tokens in C4, a widely used training corpus, and more than 28 percent of its most actively maintained sources had become fully blocked to A.I. crawlers through robots.txt files. Most of the blocking followed the arrival of OpenAI’s GPTBot and Google-Extended crawlers in mid-2023.

![A cream spiderweb forms a gate between dark blue posts, secured with a yellow padlock, while a dark spider stands nearby against a teal background.](https://rews.cc/img/fe1298c344107be4daa03b3fc6d7ac16b6c1114e.webp)

Workplace software followed. In May 2025, Salesforce rewrote Slack’s API terms to say that an app provider “may not use API Data to train a large language model,” according to [an analysis by the law firm Hunton](https://www.hunton.com/insights/legal/salesforce-locks-down-slack-data-time-to-review-your-slack-api-terms). Glean, an enterprise search company, told customers the change would lock their data inside Slack.

The bill for the scraping era came due in court. Anthropic agreed to pay $1.5 billion to settle claims by authors whose books it had downloaded from pirate libraries, about $3,000 a work across roughly 500,000 books, and the deal won [final approval in July](https://techcrunch.com/2026/07/20/anthropics-landmark-1-5b-copyright-settlement-is-approved/). Judge William Alsup had ruled that training on the books was fair use but left the piracy claims for trial. A pirated book cost Anthropic about $3,000. A Spirit message, at Google’s price, costs less than 2 cents.

## Paying for hands

Physical work leaves no archive to buy, so Microagi, a Munich startup founded in 2025, is building one. Its consumer service, Shift, began offering free apartment cleanings in New York on May 28 and sends cleaners and private chefs who wear cameras on their baseball caps. Suzy Weiss of The Free Press [had her Brooklyn apartment cleaned](https://www.thefp.com/p/ai-company-robot-cleaning-apartment) and lunch cooked under the deal in June, and was told the workers earned about $30 an hour. The footage, which the company calls egocentric human data, is sold to robotics labs or kept in-house.

> Our job is to figure out how to deploy robots at mass scale using human footage.
>
> — Bercan Kilic, co-founder and chief executive of Microagi, [thefp.com](https://www.thefp.com/p/ai-company-robot-cleaning-apartment)

Microagi raised 48 million euros, about $55 million, in July, the largest seed round ever for a German startup, [Heise reported](https://www.heise.de/en/news/German-robotics-startup-raises-largest-seed-round-Coop-with-Google-Cloud-11372650.html). People who film their own chores with an iPhone on a head rig can make roughly $20 an hour, and the company decides which recordings get paid. Microagi says more than 20,000 people in 15 countries have recorded work for it, and in September it [hired Animesh Garg and seven of his researchers](https://rews.cc/a/microagi-hires-georgia-tech-s-animesh-garg-and-seven-of-his--769d6e) from Georgia Tech. The Free Press named Luel, Kled and Waffle Video as competitors. DoorDash entered on March 19 with an app called Tasks that pays couriers to film themselves loading dishwashers and folding clothes, [TechCrunch reported](https://techcrunch.com/2026/03/19/doordash-launches-a-new-tasks-app-that-pays-couriers-to-submit-videos-to-train-ai/).

The cheapest footage comes from much farther away. In Karur, a town in southern India, Sneha Jegadish, 24, clips a smartphone to a headset and records herself chopping carrots, folding laundry and mopping for two to three hours a day, earning about $120 a month, [NPR reported](https://www.ypradio.org/2026-10-15/to-train-robots-ai-companies-are-paying-people-in-india-to-film-themselves-doing-chores) on Oct. 15. Her employer, Objectways, pays thousands of people around the world, including in the Philippines and Nigeria, and films chores in a mock home with a kitchen and three bedrooms. The company says most of its clients for such footage are in the United States.

> Think of today’s AIs as brains without a body, right?
>
> — Prathosh A.P., a machine-learning professor at the Indian Institute of Science in Bengaluru, [ypradio.org](https://www.ypradio.org/2026-10-15/to-train-robots-ai-companies-are-paying-people-in-india-to-film-themselves-doing-chores)

Investors are paying for the robots that footage is meant to train. Robotics and physical-A.I. startups raised $16.3 billion across 492 deals in the first quarter, the sector’s strongest quarter on record at the time, according to [PitchBook](https://pitchbook.com/news/reports/q1-2026-robotics-physical-ai-vc-trends). The second quarter beat it, with $18.6 billion across 450 deals. Some of the money goes to simulated worlds instead. On Sept. 28, AMD agreed to buy World Labs, Fei-Fei Li’s company, which makes models that generate 3D worlds, for $8.2 billion in stock, [Bloomberg reported](https://www.bloomberg.com/news/articles/2026-09-28/amd-to-buy-fei-fei-li-s-world-labs-ai-startup-for-8-2-billion). AMD had joined a $1 billion round in February that valued the startup at about $5 billion.

[![People are getting paid to teach robots - Rich On Tech](https://i.ytimg.com/vi/YjqOOrhcrhY/hqdefault.jpg)](https://www.youtube.com/watch?v=YjqOOrhcrhY)

Rich DeMuro of Rich On Tech meets a worker paid to wear a camera and record household chores used to train robots. Video: KTLA 5 · YouTube

## Law firms hold on

Large law firms are mostly keeping what they have. A third Business Insider story in the series describes firms building A.I. systems around their own deal histories, trial strategies and client preferences, and [Bloomberg Law reported](https://news.bloomberglaw.com/legal-ops-and-tech/law-firms-capitalize-on-internal-data-to-stand-out-in-ai-age) that Kirkland & Ellis had committed $500 million to building its own A.I. capabilities. Their richest material, client files, is covered by attorney-client privilege.

Harvey, the legal-A.I. company, says it built its first model without that material. Tenet, announced Aug. 20, is a version of Moonshot AI’s open-weight Kimi K3 that Harvey post-trained with the training platform Fireworks. On Harvey’s Legal Agent Benchmark, where a task counts only if every grading criterion passes, Tenet completed 19.7 percent of held-out tasks against 10.8 percent for the base model, at $5.92 a task against $5.62, [Fireworks wrote](https://fireworks.ai/blog/post-training-kimi-k3-with-harvey-for-long-horizon-legal-work). That works out to about $30 per completed task instead of about $52. The figures are the companies’ own and have not been independently checked. [Harvey said](https://www.harvey.ai/blog/post-training-update-harvey-tenet) no customer data went into training, and it described the training mix in a post on X.

> Introducing Tenet, our first model post-trained for legal.
>
> Tenet is a Kimi K3 base that we post-trained with @FireworksAI\_HQ on a corpus of publicly available legal data, synthetic data, and human expert data simulating long-horizon legal work.
>
> Training increases Tenet’s…
>
> — **Harvey** @harvey on X · [August 20, 2026](https://x.com/harvey/status/2090454750059958440)

## Weighing the case

The simplest version of this story is that A.I. companies ran out of internet and went shopping. The record supports a narrower one. Epoch’s own exhaustion date has already moved later once, and the private data now in demand is bought less to pretrain models on more text than to teach agents how jobs get done. Nick Heiner, who leads reinforcement-learning environments at Surge, an A.I. data company, told Time that the Spirit deal was “not centrally a coding agent purchase.”

> That’s a purchase if you believe that your agent is going to generalize into the rest of the economy.
>
> — Nick Heiner, head of reinforcement-learning environments at Surge, [time.com](https://time.com/article/2026/08/25/google-spirit-airlines-ai-data-RL/)

Surge and Mercor already pay experts to build realistic office simulations from scratch. Mr. Heiner said even that work is “a little bit removed from literally having actual data that was used in the real world.” Records of bad decisions have uses too, he said. A task built from something a company did poorly can be rewritten so the model learns the better move. On that account, a failed airline’s records are worth buying.

The privacy case is thinner than the buyers say. The one independent review so far, Ms. Thomson’s, did not examine employees. For everyone else, the protection is a public pledge by the buyer not to re-identify anyone, which any later recipient must also make, and a scrub by a firm that Google chose and paid for, [The Next Web reported](https://thenextweb.com/news/micro1-12-5m-counterbid-spirit-airlines-records-google-deidentification-cjeu-srb-edpb-opinion-28-2024) in its account of Micro1’s offer. Micro1 has pitched an ombudsman picked by Spirit’s own advisers as an alternative. The flight attendants’ point about linked records has not been answered by anyone with no stake in the sale.

Consent is where the argument gets loudest. In the Hacker News thread, one commenter argued that Spirit’s employees had consented when they took their jobs. Another replied:

![A man carrying a briefcase walks away, casting a long shadow shaped like a large price tag on a teal background.](https://rews.cc/img/f9e0a7c884c2396acca8f627f31314560ddbb0f7.webp)

Under current law, the first commenter is closer to right. Ifeoma Ajunwa, an Emory University law professor, told Business Insider that workers whose messages are sold have little recourse because the United States has no federal privacy law. Contracts may matter more than statutes. Slack’s terms bar app makers from training on API data, but it was not clear whether they reach a customer that exports its own workspace and sells it. Mercor faces class-action suits, [among them claims](https://www.business-humanrights.org/en/latest-news/usa-ai-startup-mercor-sued-over-hiring-practices-and-contractors-data-handling/) that it trained client models on material potentially owned by contractors’ other employers. Mercor has denied the allegations.

## What comes next

Judge Sean H. Lane of the U.S. Bankruptcy Court for the Southern District of New York has postponed the sale hearing several times since early September, most recently on Oct. 9, [when it moved to Nov. 6](https://chapter11cases.com/blogs/news/spirit-airlines-case-update-the-data-sale-slips-to-november-as-spirit-moves-to-unwind-the-hfs-aircraft-deal). Mercor’s $7.5 million offer and Micro1’s $12.5 million offer are still in play. Ms. Thomson told the court that if it considers either rival she would need further review, including information about a “supply chain security incident” at Mercor this year and about Micro1’s plan to handle its own de-identification.

Micro1’s $1 billion is supposed to be spent by next October. AMD and World Labs said they expect their deal to close by the end of the year, pending regulatory approval.

If Judge Lane approves Google’s purchase, Tonic.ai must file a certification with the court before each transfer of Spirit data, and Ms. Thomson has asked that each one state how many passengers’ details it found and removed.

## See also

- [Spirit Airlines Is Selling Its Workers' Data. That Can't Fly.](https://www.techpolicy.press/spirit-airlines-is-selling-its-workers-data-that-cant-fly/) — techpolicy.press · Tech Policy Press commentary arguing against the employee-data sale
- [We Should Be Worried About a Google-Spirit Deal](https://onlabor.org/we-should-be-worried-about-a-google-spirit-deal/) — onlabor.org · A labor-law blog's case on what the sale means for workers
- [Cantwell, colleagues urge Spirit Airlines and Google to protect workers' privacy](https://www.commerce.senate.gov/press/dem/release/cantwell-colleagues-urge-spirit-airlines-and-google-to-protect-workers-privacy-amid-sale-of-data-for-ai-training/) — commerce.senate.gov · Senate Commerce Democrats' statement on the congressional letter
- [With Robots.txt, Websites Halt AI Companies' Web Crawlers](https://spectrum.ieee.org/web-crawling) — spectrum.ieee.org · IEEE Spectrum interview with Shayne Longpre on the shrinking open web
- [Q2 2026 Robotics & Physical AI Report](https://pitchbook.com/news/reports/q2-2026-robotics-physical-ai-report-the-money-is-in-the-motion) — pitchbook.com · PitchBook's latest quarterly data on robotics funding
- [These workers help train robots who could one day steal their jobs](https://npr.org/2026/10/14/g-s1-146907/these-workers-help-train-robots-who-could-one-day-steal-their-jobs) — npr.org · NPR on Indian factory workers recording their movements for robot training
- [Anonymising personal data 'not enough to protect privacy', shows new study](https://www.imperial.ac.uk/news/192112/anonymising-personal-data-enough-protect-privacy/) — imperial.ac.uk · Imperial College's plain-language summary of the re-identification study
- [Who Buys Company Data for AI Training?](https://www.troveo.ai/resources/who-buys-company-data-for-ai) — troveo.ai · A data broker's own map of the buyers, from inside the industry