A census of the corpora that carry writing into machine training, the tools that test whether a work is in a training set or trace an output back to training text, the readers, models and practices that make new writing from what they read, the literature composed for a machine or other carrier to read, and the field's history and indexes. Each resource is read from its own pages, and each states what it is to this census: a direct case, a historical lineage, an empirical comparator, infrastructure, a theoretical interlocutor or a peripheral comparator. Every field is quoted from a saved copy of the resource's own pages, with the date read and a hash of the copy. Compiled 2026-11-01; editor Lee Sharks.
Infrastructure
a feature in the Ai2 Playground that traces language model outputs back to their full training data in real time; its paper calls it a system, released as open source · Ai2 (The Allen Institute for Artificial Intelligence), as a feature in the Ai2 Playground; the paper's authors are Jiacheng Liu and 30 others
infrastructureInspectionReception becoming compositionactive
What the apparatus makes from it: the model's response, of which only long, unique, self-contained spans with a verbatim match are highlighted (K = 0.05 times the response's token count); Ai2 shows it tracing 'creative' expressions, arithmetic and a false knowledge-cutoff claim to training documents, in its own hedged terms ('may have learned', 'could have easily learned it through memorization')
artifacts that record training data and allow downstream inspection, proposed for widespread adoption and implemented with data sketching, in a research paper (NeurIPS 2023 Datasets and Benchmarks) · the paper's two authors, Marc Marone and Benjamin Van Durme; the saved source names no institution, and the live interface's operator is unread
infrastructureInspectionThe corporahistorical
What the apparatus makes from it: not recorded as a product of the model; the paper states the tool 'enables answering questions about test set leakage and model plagiarism', which checks material against the recorded training data
free, open repository of web crawl data · the Common Crawl Foundation, a 501(c)(3) non-profit founded in 2007; Wikipedia names Gil Elbaz as founder and Rich Skrenta as executive director
infrastructureThe corporaInspectionactive
What the apparatus makes from it: not recorded on its own pages; per Wikipedia, other organizations refine its raw data into datasets (FineWeb, DCLM, C4), and AI companies train large language models on the data
online library of free eBooks · the Project Gutenberg Literary Archive Foundation, which operates Project Gutenberg and accepts donations; per Wikipedia, a 501(c)(3) chartered in Mississippi in 2000, with Eric Hellman acting CEO after Gregory Newby died in 2025
infrastructureThe corporaactive
What the apparatus makes from it: not recorded on its own pages; downstream, per Wikipedia, the Pile carries a 'Gutenberg (PG-19)' component, and NaNoGenMo's rules allow a generated 'novel' that grabs 'a random novel from Project Gutenberg'
non-profit digital library (Wikipedia's description; the own about page saved no text) · the Internet Archive, an American 501(c)(3) non-profit library founded in 1996 by Brewster Kahle, headquartered in San Francisco (per Wikipedia; its own about page was unread)
infrastructureThe corporaactive
What the apparatus makes from it: not recorded: the saved sources name no training set or model made from its holdings
large-scale collaborative repository of digital content from research libraries (Wikipedia's description) · HathiTrust, a university consortium of more than 219 research libraries (2024) under shared governance, administered by the University of Michigan; executive director Mike Furlough (per Wikipedia; the own about page was refused)
infrastructureThe corporaactive
What the apparatus makes from it: per Wikipedia, a full-text search index across the repository, which the courts held transformative and fair use; computational analysis of the volumes was the separate service of the HathiTrust Research Center, ended 30 September 2026
research center dedicated to enabling computational analysis of the HathiTrust Digital Library; its analytics portal · the HathiTrust Research Center, established in 2011 as a partnership between Indiana University, the University of Illinois Urbana-Champaign and HathiTrust
infrastructureThe corporaended
What the apparatus makes from it: derived datasets computed from the volumes (Extracted Features, BookNLP dataset, word frequencies, geographic locations), HathiTrust+Bookworm and the Gendered Characterizations visualization; the saved sources record no model training
open source language modelling data set: an English text corpus targeted at training large-scale language models · EleutherAI (the home page footer reads 'EleutherAI 2021'; the paper's authors are Leo Gao and 11 others)
infrastructureThe corporahistorical
What the apparatus makes from it: language models trained on it: its own page scores models on Pile BPB, and the paper reports that models trained on the Pile improve over Raw CC and CC-100 on all its components
an English corpus for language model pretraining research, released with its documentation and an open-source data curation toolkit · the paper's 36 authors, Luca Soldaini and 35 others (arXiv:2402.00159); the dataset card sits under the allenai namespace on Hugging Face, and its text was rate-limited (HTTP 429) and is unread, so the saved sources name no institution
infrastructureThe corporaactive
What the apparatus makes from it: not recorded in the saved abstract beyond the corpus's purpose, research on language model pretraining; the paper reports analyses and experiments on intermediate states of the corpus
not recorded in the saved sources: on 2026-11-01 haveibeentrained.com served a static Spawning maintenance page that names a Do-Not-Train registry · Spawning (the wordmark on the page served at haveibeentrained.com)
infrastructureThe corporapaused
What the apparatus makes from it: not recorded
Historical lineage
a combination museum, library, and preservation space for collections of digital art and writing, in its own words; Wikipedia (third party) calls it a repository of electronic literature · the Electronic Literature Organization (ELO), managed by the Electronic Literature Lab at Washington State University Vancouver; Dene Grigar, Managing Director and Curator; the ELO Repository names the university library as a partner
historical lineageField history and indexes of electronic, generative and computational literatureWriting addressed to a receiving apparatusactive
What the apparatus makes from it: no machine composition is recorded; what the preservation apparatus makes from a work is recorded (Wikipedia, third party): re-created and re-imagined versions on newer systems, video playthroughs, traversal recordings, and metadata standing in for works that cannot yet be run
in the publisher's words, 'An anthology of seven decades of English-language outputs from computer generation systems' · edited by Lillian-Yvonne Bertram and Nick Montfort; published by Counterpath, co-published with MIT Press (the Counterpath page is the saved source; its title line prints the first editor's name as 'Lillian Yvonne-Bertram')
historical lineageField history and indexes of electronic, generative and computational literatureReception becoming compositionended
What the apparatus makes from it: the anthology's object is what generation systems made: outputs across conversations, humor, letters, poetry, prose and sentences, from implemented systems only
a Columbia RBML exhibit, online in eight cases, on the long shared history of literature and computation through the Columbia Library's holdings · a collaboration between the Columbia English Department, the Columbia University Rare Books & Manuscript Library and Columbia University's Digital Scholarship department; the site is copyright Columbia University Libraries
historical lineageField history and indexes of electronic, generative and computational literaturehistorical
What the apparatus makes from it: in the exhibit's terms, 'algorithmic composition, such as prose and poetry written by machines', plots produced by plot generators with 'spinning devices or automatic number generators', and analyses that sort narrative into types; it records these as production and analysis
a publication of the Electronic Literature Organization: curated, edited volumes of electronic literature, published on the Web and as a physical version · the Electronic Literature Organization (ELO), c/o Washington State University Vancouver as the collection's mailing address; Wikipedia (third party) states the ELO is currently hosted at York University, Toronto
historical lineageField history and indexes of electronic, generative and computational literatureWriting addressed to a receiving apparatusactive
What the apparatus makes from it: not recorded
a research resource for electronic literature, now preserved as a frozen archive · the University of Bergen Electronic Literature Research Group, as an outcome of the ELMCIP project; built with a community of international researchers and writers
historical lineageField history and indexes of electronic, generative and computational literaturehistorical
What the apparatus makes from it: not recorded
The census as data
Resources listed and held for the next batch, whose own pages could not be read on fetch: Books3 search (The Atlantic).
Pages that could not be read on fetch, each recorded with its status, and their fields left unquoted: books3-search--atlantic (http 403); books3-search--wikipedia (http 404); dolma--card (http 429); elmcip--about (http 404); elmcip--home (http 200); hathitrust--about (http 403); htrc--about (http 200); internet-archive--about (http 200); nanogenmo--org (http 403); nanogenmo--wikipedia (http 404); output-anthology--mit (http 000); spawning-hibt--spawning (http 200); the-next--curatorial (http 200).
census.json (schema 1.1). The MMRS, provenance, Labor Vector and left-census core (slug, name, aka, url, operator, kind, unit, count, lifecycle, version_and_time, negative_tags, evidence, relations, unchecked, sources) extended for the training layer (1.1, revised 2026-11-01): host_relation, record_type and question as in the authorship census; areas (one or more of six), evidence_kind, record_object (what it records), record (in_layer: whose works it holds or carries into training, or how it bears on that; name: whether a work keeps its author's name and attribution there, as the resource itself shows, or 'not recorded'; readback: whether and how a work can be found there and read back: search, membership test, provenance, or 'not recorded'; makes: what a receiving apparatus makes from the work there, as the resource shows, or 'not recorded'), method, coverage, holdings, stated_limits, answers and does_not_answer, and tll_reading (where it meets the question of #186 and #1543: the archive's reading, kept apart from the resource's own terms). Every quote is verbatim from a saved source.