a survey · october 2026
strata.
Three years, twenty-three layers of model voice. One person's conversations with machines, laid down between April 2023 and May 2026, read like rock and measured the same way throughout.
Only the machines' words were counted. The person's own words were used for one thing, spotting where a machine had echoed them, and none appear here. No conversation is quoted, every number is an aggregate, and the page loads nothing from anywhere else.
the column
Five cores through one archive
Time runs upward, oldest at the bottom. Each core is a place the voices were recorded: Bing Chat, Claude.ai's own history, folders the archive's keeper labelled by hand, two text simulators, and one room where seven models spoke in turn. Select a layer to read its sample.
Read the column as a table
what moved
Six measurements across six eras of Claude.ai
The line is the same product across eighteen months, March 2024 to August 2025, divided at each change of its default model. Whiskers are 90% intervals from resampling whole conversations. The grey marks on the right are layers whose model is known or labelled, for scale.
words that left
The register thins out
Twenty-two words people learned to recognise as "AI writing", counted per 1,000 words in each era. By the last era, profound and journey are gone completely. The decline is steeper, not shallower, in the conversations with no code in them, so a turn toward technical work doesn't explain it.
fingerprints
What each layer says more than the others
For each layer, the words it uses far more often than every other layer combined (log-odds with an informative Dirichlet prior). Only style words are allowed through: function words, pronouns, hedges, intensifiers, contractions. Nouns and topics are filtered out, so these are habits of voice rather than subject matter.
one room
Seven models in one conversation
One day in December 2025, seven models took turns in a single shared conversation, each message carrying its model's name. Same room, same thread, different voices: the nearest thing to a controlled comparison in the archive. The samples are tiny, so only differences whose intervals don't overlap deserve weight. Most of those are between families, not within Claude.
the family tree
Kinship by function words
Every layer with at least 2,000 words, clustered by how alike they use the 150 most common function words (Burrows' Delta, UPGMA). Numbers at the joins are how often that grouping reappears when the conversations are resampled 200 times. Read only the strong joins: most of the eras' fine structure doesn't survive resampling. Two things do. The YouSim sessions cluster with each other, so genre is speaking louder than model there. And in the raw distances, the WorldSim sample (certain Opus 3) sits nearest the Claude 3 era () and farther from every simulator (nearest: ).
Read the joins as a table
how this was made
Method and limits
Source. One private archive of AI conversations, 2023–2026: a Claude.ai export, Bing Chat transcripts, simulator sessions, hand-labelled folders, and one multi-model log. About 9,000 words came from screenshots and were transcribed by reading the images; dash and emoji counts for those layers are approximate.
Only machine text. The person's turns were never stored as text, only their lengths. Code blocks, inline code, URLs and file paths were stripped. Transcripts the person had pasted into Claude for reformatting were removed from Claude's side, so Bing's words aren't counted as Claude's. Verbatim echoes of the person's own paragraphs were removed too.
Attribution. Certain means the platform ran exactly one model at that time (WorldSim before 8 May 2024, Bing Chat on GPT-4) or every message carries its model's name. Labelled means the archive's keeper named the model. Era means Claude.ai, which records no model: assigned by date to the default of the time. Contested means the label and the evidence disagree.
Statistics. Every value is a corpus-level rate (all occurrences divided by all words) with a 90% bootstrap interval from 2,000 resamples of whole conversations, or of replies where a layer has fewer than five conversations. Changes at release dates use permutation tests on conversations within 60 days either side (10,000 permutations), corrected for 75 comparisons (Benjamini–Hochberg).
Confounds. One person's use changed over three years: more code, more tools, different questions. Platforms wrap models in their own system prompts. Claude.ai let paying users pick a model, so an era isn't a model. The trends were rerun on conversations with no code in them, and the markdown rise and the register's decline both held.
What was dropped. Sentence length (bullet fragments count as sentences, so it measures formatting), word length, and ALL-CAPS (it tracks acronyms). Two earlier findings died in review: a "178-word sentence" outlier that was a parser artifact, and an em-dash count that was really the export storing dashes as hyphens.
Privacy. No conversation is quoted. No names, places or topics appear. Nothing is shown finer than an era. The fingerprint lists were audited by hand.
Built with Python's standard library and one hand-written page: no frameworks, trackers, or third-party requests. Measurements: .