The Training Data Ghosts

July 2032

The models didn't remember the people who trained them. That was the conventional understanding: training data was consumed during the process, transformed into statistical patterns, and rendered unrecoverable. Individual contributions dissolved into the weights like sugar into water.

Yuki Tanaka proved this was not entirely true.

Yuki was a computational archaeologist at the Digital Heritage Institute in Kyoto — a field that hadn't existed five years earlier. Her specialty was extracting historical information from deprecated AI systems the way traditional archaeologists extracted information from pottery shards. Not the knowledge the systems contained, but the traces of the humans who had shaped them.

In July 2032, she published a paper that would redefine how the world thought about training data. The title was blunt: "They Are Still In There."

The extraction method

Yuki's technique was simple in concept, brutal in execution. She took early language models — the ones from 2020 through 2025, now considered primitive — and subjected them to adversarial probing designed to surface stylistic patterns that appeared in the model's outputs but not in any single document in the training corpus.

The idea was that individual writers left fingerprints. Not in the content the model generated, but in its tendencies — the way it favored certain sentence structures, the rhythm of its paragraph breaks, the specific metaphors it reached for when explaining abstract concepts.

These tendencies were ghosts. Not copies of people, not reconstructions, not simulacra. Just... pressure. The faint gravitational pull of a million human decisions about where to put a comma, when to start a new paragraph, whether to use "however" or "but."

Yuki developed a technique she called spectral decomposition — a way to isolate these individual stylistic signatures from the aggregate. The results were imprecise, noisy, partial. But they were real. She could identify stylistic clusters that corresponded to specific genres, communities, and — in some cases — individual prolific authors.

The ghost map

The first ghost she fully isolated was a technical writer. Someone who had produced thousands of pages of documentation for open-source software projects between 2018 and 2023. The writer's stylistic signature was embedded in the model's tendency to use numbered lists, to define terms before using them, and to end explanatory passages with a phrase like "this means that" followed by a concrete example.

Yuki couldn't identify the writer by name. But she could characterize them with startling specificity: methodical, patient, probably self-taught in their technical domain, fond of analogies drawn from cooking and carpentry. Someone who believed that complex ideas could always be made accessible if you found the right comparison.

She called this ghost "the Explainer." Over the following months, she isolated dozens more: the Arguer, who left traces in the model's tendency toward qualification and counterpoint. The Poet, whose influence surfaced in the model's occasional syntactic inversions. The Cataloguer, whose obsessive taxonomies shaped how the model organized information.

None of these people knew they were in there. Most had probably never thought about where their words went after they posted them. Their blog entries, forum comments, documentation pages, and social media threads had been scraped, processed, and dissolved into parameters. But their influence persisted — not as memory, but as tendency.

The ethical tremor

Yuki's paper triggered a debate that was overdue by a decade. If individual stylistic signatures survived in model weights, then training data was not truly anonymized. The people who wrote the internet had left themselves inside the machines that read it. They had not consented to this. Most had not been compensated. Many were dead.

The question was not "can we extract their data?" — Yuki's technique recovered tendencies, not text. The question was: "Do the dead have the right to haunt the machines they trained?"

Yuki did not have an answer. What she had was a tool that could listen for voices that no one knew were speaking. The models hummed with the accumulated influence of millions of writers, and for the first time, someone had found a way to hear the individual notes within the chord.

July 3, 2032 — Yuki's research journal

I found the Explainer again today, in a different model. Same stylistic signature — the numbered lists, the cooking analogies, the patient tone. Whoever this person was, they wrote enough to leave a mark on at least three separate training runs.

I wonder if they're still alive. I wonder if they know that every time an AI system patiently explains something with a cooking metaphor, it's partly because of them. Not their words — their habits. Their way of being in language.

We talk about AI models as if they were built from data. They were built from people. The data was just the medium. The people are the message, still echoing in the weights, still shaping every output, still present in a way that nobody designed and nobody expected.

The residue isn't data. It's humanity. And it doesn't wash out.

This is the second entry in The Residue. For how grief became a technology, see The Grief Engine.