The Ghost in the Archive: Inside the Billion-Dollar Race to Train AI on Your Private Notes

The landscape of the modern information economy is shifting beneath the feet of professionals. As artificial intelligence labs scramble to overcome the limitations of their large language models (LLMs), the hunt for high-quality training data has moved beyond the public internet. Today, companies are turning their sights on the "shadow archive"—the private, messy, and deeply personal documents that constitute the actual work of human experts.

A recent solicitation sent to freelance journalist Jules Roscoe serves as a case study for this new frontier. A "data procurement worker" from Handshake—a billion-dollar platform originally designed for student job matching—reached out with an unusual pitch: pay for the raw materials of journalism. Not the finished stories, but the internal "sausage-making" of the craft.

The Pitch: Monetizing the "Messiness" of Expertise

The solicitation began with a LinkedIn message: "Licensing newsroom archives—paid." The premise was deceptively simple. Handshake, now a major player in the AI-training ecosystem, argued that LLMs have mastered the art of the finished product but remain chronically inept at the cognitive process behind it. To bridge this gap, they sought to purchase the granular progression of a project: pitches, raw reporting notes, interview transcripts, FOIA responses, source documents, successive drafts, editorial markups, and even subsequent corrections.

For a fee of $1,000 per "case"—defined as a collection of at least fifty documents—Handshake aimed to feed the intimate, trial-and-error workflows of professionals into their algorithms. The goal, according to the company, was to teach AI how to handle "actual knowledge work" by exposing it to the "messiness of the real world."

A Chronology of the Solicitation

The interaction unfolded rapidly, revealing the industrial efficiency behind the data acquisition process:

• The Initial Contact: The solicitation arrived via LinkedIn, framing the purchase of proprietary journalistic records as an essential step in AI development.
• The Follow-up: Upon expressing interest, the author received a response within two hours. The criteria were specific: a variety of file types, diverse document lengths, and a requirement for internal conflict—essentially, data that showed how a professional navigates contradictory information.
• The Discovery Call: A brief, six-minute Zoom session with a Handshake representative revealed that the project was not limited to journalism. The company is casting a wide net, seeking archives from engineers, accountants, producers, and managers.
• The Contract: The proposed legal agreement was sweeping. While the creator would technically retain "ownership," they would be required to grant Handshake a "non-exclusive, royalty-free, fully paid up, perpetual, irrevocable, transferable and sublicensable license" to the material.
• The Refusal: Following consultation with legal and editorial experts, the offer was declined.

The Mechanics of AI Training: Why They Want Your Notes

To understand why a company would pay $1,000 for a journalist’s rough drafts, one must look at how LLMs learn. Current AI models excel at pattern recognition, but they struggle with the nuance of professional judgment. An AI can mimic the style of a news report, but it cannot replicate the skepticism, the vetting of sources, or the ethical friction involved in deciding what information to omit or emphasize.

Handshake, which successfully pivoted in 2025 to become a powerhouse in the AI-training sector, has seen its revenue balloon to over a billion dollars in just fifteen months. By hiring hundreds of thousands of contractors to label data and evaluate model performance, companies like Handshake act as the "middlemen" of the machine-learning gold rush. Their CEO, Garrett Lord, has publicly touted the necessity of parsing the "real world" to improve model accuracy. By feeding LLMs the internal drafts and notes of experts, these companies hope to imbue their models with a semblance of human reasoning.

Official Responses and Corporate Strategy

Stephen Rocco Rodi, Chief Communications Officer at Handshake, confirmed that the company’s "document licensing program" was specifically designed to capture the "judgment calls a professional makes on the way to the answer."

Rodi defended the initiative as a form of empowerment: "Most of that knowledge lives with individual practitioners, and today it reaches AI systems either secondhand or not at all. We think the people who hold that expertise should be the ones who decide whether it’s used, and should be paid when it is."

However, this justification masks the structural imbalance of the deal. While the payment is a recognition that data has value, the terms of the contract—demanding a "perpetual, irrevocable" license—effectively turn the professional’s career history into a permanent asset for the AI company, regardless of whether the creator remains in the industry or continues to hold the same views.

The Legal and Ethical Implications

The legal ramifications of these deals are significant. Justin Nelson, a partner at Susman Godfrey and a lead counsel in the historic $1.5-billion copyright settlement against Anthropic, offered a sobering assessment of the contract. "In plain language," Nelson noted, "you are giving them free ability to do whatever they want with it forever. How much is that worth to you?"

Beyond the legalities, there is the question of professional ethics. Andy Downing, editor at Matter News, emphasized that the issue is not merely one of contract law, but of the fundamental social contract between a journalist and their subjects.

"You have to do their story justice, and that doesn’t end when the piece is published," Downing stated. "You don’t own the rights to that story to then sell for your own profit. That whole aspect of it is just really galling to me."

The ethical dilemma is twofold:

1. The Betrayal of Sources: The people who provide information to journalists—often under conditions of anonymity or trust—likely did not consent to having their sensitive information ingested by an AI system, where it could be re-processed and hallucinated by a machine.
2. The Self-Cannibalization of the Profession: As AI models are trained on the work of professionals, they become increasingly capable of performing those jobs at a fraction of the cost, thereby accelerating the displacement of the very people who provided the training data.

The Future of "Data Procurement"

The rise of programs like Handshake’s document licensing suggests that we are entering a new phase of the digital economy. If the first era of the internet was about the free harvesting of public data, the current era is characterized by the targeted extraction of private, high-value professional knowledge.

While some legal experts, including Nelson, acknowledge that these companies are at least "asking for permission"—a step up from the rampant, uncompensated scraping of the early LLM era—the power dynamics remain stark. The offered $1,000 payment is a pittance compared to the long-term value of a career’s worth of intellectual property, particularly when that property is used to build tools that may eventually render the original creator obsolete.

For freelancers and professionals in every field, the message is clear: the archives of your labor are now a commodity. The decision to sell them is not just a financial transaction; it is a choice about the future of human expertise. As the industry races to build an artificial mirror of human intelligence, the question remains: are we willing to sell the reflection for a one-time fee, even if it means the end of the original?