From Surveillance Capitalism to AI Imperialism | Chapter 04: AI arrives when the human raw material already exists
Foundation models do not emerge from nowhere. The U.S. Once captured human production can be recombined into proprietary weights, the surveillance economy acquires a new layer.
AI arrives when the human raw material already exists; surveillance capitalism; AI imperialism; AI weights; patrimony of humanity; evidence-based dossier
How does ai arrives when the human raw material already exists function within the dossier’s larger argument, and what does the documentary record allow this chapter to establish?
This paper examines ai arrives when the human raw material already exists within the broader dossier From Surveillance Capitalism to AI Imperialism. It is intentionally written to circulate on its own: necessary context is restated, evidence boundaries are made explicit, and cited material is reproduced in paper-local form rather than delegated to repo navigation. It contributes to the series as a paper that provides the central evidentiary basis for treating frontier weights as a patrimony and governance problem rather than a neutral market outcome. Adjacent reinforcement appears in Chapter 00, Chapter 05, Chapter 06, Chapter 10, but those cross-references are supportive rather than required for basic comprehension.
Foundation models do not emerge from nowhere. They emerge after a long period in which human expression, cultural production, and socially filtered judgment were captured at scale, weakly compensated if at all, and converted into machine-trainable inputs. 1
This paper is derived from the chapter corpus for Chapter 04 and is grounded in the project’s structured research OS. It relies on source notes, extracted claims, entity profiles, and timeline events already normalized under the dossier’s evidence hierarchy. Documented facts are asserted directly where the record supports them; disputed matters are marked as such; interpretive claims are bounded explicitly; and speculative overreach is isolated in the evidence-boundary appendix.
This paper asks: How does ai arrives when the human raw material already exists function within the dossier’s larger argument, and what does the documentary record allow this chapter to establish? Its working answer is clear from the start: Foundation models do not emerge from nowhere. They emerge after a long period in which human expression, cultural production, and socially filtered judgment were captured at scale, weakly compensated if at all, and converted into machine-trainable inputs. 2 The objective is not only to recount developments, but to show why the subject of this chapter belongs inside a structural argument about extraction, enclosure, legitimacy, and reform.
The U.S. Copyright Office has already treated AI training on copyrighted works as a major live policy problem rather than a settled technical inevitability. The 2023 inquiry and 2025 report together show that training-input legality, disclosure, and compensation were contested near the start of the generative-AI boom and remain contested now. 3
The technical and legal record also undermines the industry’s preferred abstraction. Research on memorization shows that large language models can reproduce identifiable material from training data, while court records and legal scholarship now separate at least two layers of conflict: training-use doctrine and input-acquisition legality. 4
The social-to-model bridge is also explicit. GPT-2’s WebText was built from socially filtered outbound links shared on Reddit, and by February 2024 Reddit had moved from diffuse public conversation to direct licensing, striking a major agreement that allowed Google to use Reddit posts for AI training and related product improvement. The extraction economy did not end with unpaid capture; it evolved into direct licensing over accumulated human expression. 5
The labor layer matters too. Research on annotation and outsourced human labor shows that model building depends not only on scraped creative works but also on significant human verification, correction, and operational support that is often geographically externalized and weakly valued. 6
Taken together, the record shows a sequence in which collectively produced expression becomes training input, archived conversation becomes licensable model fuel, and additional human labor is used to clean, annotate, and operationalize the resulting systems. Value that first appeared as diffuse public production is progressively reorganized inside proprietary model pipelines. 7
Interpretation: the strategic move is not only model creation but control of the interface through which machine-packaged intelligence is accessed. In that reading, the shelf is the interface: the prompt box, chatbot, or subscription layer becomes the first point of contact between a buyer and capability distilled from human work. Once that first contact is captured, value can be repriced downward and dependence can deepen even where total technical replacement remains incomplete. 8
The House AI Task Force report gives this chapter an institutional bridge beyond agencies and courts. By December 17, 2024, a bipartisan congressional report was already recording the training conflict in language overlapping with the dossier’s core concerns about commercialization, consent, credit, compensation, and derivative-work theory. 9
Interpretation: this is the dossier’s primary evidentiary base for
AI Weights as Patrimony of Humanity. If frontier weights
are built from collectively produced human expression and labor, then
they cannot be treated as ordinary private assets by default. The
strongest normative direction pushes toward broad, open, utility-like
availability under public-interest rules rather than permanent private
scarcity. That is advocated doctrine, not settled law. 10
This paper sits inside a coordinated 11-paper architecture. Its nearest companions are Chapter 00 (Opening framework), Chapter 05 (Talent, overhiring, valuation, and layoffs), Chapter 06 (AI as a dual-use rhetorical weapon), Chapter 10 (What to do). In that series logic, Chapter 04 provides the central evidentiary basis for treating frontier weights as a patrimony and governance problem rather than a neutral market outcome.
The current record is strong on conflict, appropriation structure, memorization, licensing, and the social-to-model bridge. It is still weaker on full training-stack provenance across every frontier lab. This chapter should therefore challenge enclosure forcefully without pretending that every corpus is already fully mapped. 11
The current record is strong on conflict, appropriation structure, memorization, licensing, and the social-to-model bridge. It is still weaker on full training-stack provenance across every frontier lab. This chapter should therefore challenge enclosure forcefully without pretending that every corpus is already fully mapped. 12
Once captured human production can be recombined into proprietary weights, the surveillance economy acquires a new layer. It no longer only predicts behavior. It begins to substitute for the producers whose accumulated traces made the system possible in the first place. That is why the ownership question has to expand into a governance question. 13
Once captured human production can be recombined into proprietary weights, the surveillance economy acquires a new layer. The next question is how that enclosure story meets labor, finance, and layoffs. In the architecture of this series, Chapter 04 therefore functions as a self-contained argument while also advancing the cumulative path toward the dossier’s three reforms.
This paper is designed to circulate independently. Public notes point readers to the real external documents first, then to the paper-local appendix entry that explains why each source matters.
The underlying research OS distinguishes among
Documented Fact, Disputed Fact,
Hypothesis / Interpretation, and
Speculative Narrative Risk. In Chapter 04, those boundaries
remain visible in the prose, in the bibliography, and in the appendix
sections that summarize claims and caution points.
usco-ai-training-report-2025. Relevance: That the U.S.
Copyright Office treats AI training on copyrighted works as a major live
legal and policy problem rather than a settled matter.radford-gpt2-webtext-2019. Relevance: That a major OpenAI
training corpus was explicitly built from socially filtered outbound
links shared on Reddit.ap-reddit-google-ai-training-deal-2024. Relevance: That by
February 2024 a major social platform was explicitly licensing
accumulated user conversation data for frontier-AI training.usco-ai-noi-2023. Relevance: That by August 2023 the U.S.
Copyright Office had already elevated AI training inputs and disclosure
questions into a formal federal inquiry.carlini-extracting-training-data-2020. Relevance: That
large language models can sometimes reproduce identifiable verbatim
material from their training data rather than only abstract statistical
patterns.dornis-generative-ai-training-copyright-2025. Relevance:
That there is serious academic legal scholarship directly challenging
the idea that generative AI training automatically falls under familiar
exceptions.bartz-anthropic-fair-use-order-2025. Relevance: That the
strongest primary legal record already separates the question of model
training from the question of how the training corpus was
assembled.nyt-openai-motion-dismiss-order-2025. Relevance: That
another federal court outside the Anthropic case treated training-stage
copying and dataset assembly as central issues serious enough to survive
dismissal.le-ludec-cornet-casilli-annotation-madagascar-2023.
Relevance: That AI systems depend on significant human labor for data
generation, annotation, verification, and operational correction.house-ai-task-force-report-2024. Relevance: That the
dossier’s copyright / weights / compensation framing entered a
bipartisan U.S. congressional policy record by December 17, 2024.twinchat-ai-copyright-weights. Relevance: That the dossier
has a coherent thesis about how weights condense previously captured
human labor.Source citation: Copyright
and Artificial Intelligence, Part 3: Generative AI Training
(Pre-publication Version) Institution / author:
U.S. Copyright Office Date: 2025-05-09 Tier /
type: T1 / report Why it matters in this
paper: That the U.S. Copyright Office treats AI training on
copyrighted works as a major live legal and policy problem rather than a
settled matter. Corpus companion entry: Source S-126
Audit ID: usco-ai-training-report-2025
Source citation: Language
Models are Unsupervised Multitask Learners Institution /
author: OpenAI Date: 2019-02-14 Tier /
type: T1 / paper Why it matters in this paper:
That a major OpenAI training corpus was explicitly built from socially
filtered outbound links shared on Reddit. Corpus companion
entry: Source S-108 Audit ID:
radford-gpt2-webtext-2019
Source citation: Reddit
strikes $60M deal allowing Google to train AI models on its posts,
unveils IPO plans Institution / author: AP
Date: 2024-02-22 Tier / type: T2 /
news_article Why it matters in this paper: That by
February 2024 a major social platform was explicitly licensing
accumulated user conversation data for frontier-AI training.
Corpus companion entry: Source S-024 Audit
ID: ap-reddit-google-ai-training-deal-2024
Source citation: Copyright
and Artificial Intelligence; Notice of Inquiry Institution /
author: GovInfo Date: 2023-08-30 Tier
/ type: T1 / federal_register_notice Why it matters in
this paper: That by August 2023 the U.S. Copyright Office had
already elevated AI training inputs and disclosure questions into a
formal federal inquiry. Corpus companion entry: Source
S-125 Audit ID: usco-ai-noi-2023
Source citation: Extracting Training Data from
Large Language Models Institution / author: arXiv
Date: 2020-12-14 Tier / type: T1 /
paper Why it matters in this paper: That large language
models can sometimes reproduce identifiable verbatim material from their
training data rather than only abstract statistical patterns.
Corpus companion entry: Source S-038 Audit
ID: carlini-extracting-training-data-2020
Source citation: Generative AI Training and
Copyright Law Institution / author: arXiv
Date: 2025-02-21 Tier / type: T1 /
paper Why it matters in this paper: That there is
serious academic legal scholarship directly challenging the idea that
generative AI training automatically falls under familiar exceptions.
Corpus companion entry: Source S-051 Audit
ID:
dornis-generative-ai-training-copyright-2025
Source citation: Bartz
et al. v. Anthropic PBC, Order on Fair Use Institution /
author: U.S. District Court for the Northern District of
California via CourtListener Date: 2025-06-23
Tier / type: T1 / court_order Why it matters in
this paper: That the strongest primary legal record already
separates the question of model training from the question of how the
training corpus was assembled. Corpus companion entry:
Source S-035 Audit ID:
bartz-anthropic-fair-use-order-2025
Source citation: The
New York Times Company v. Microsoft Corporation et al, Opinion on
Motions to Dismiss Institution / author: U.S.
District Court for the Southern District of New York via Justia
Date: 2025-04-04 Tier / type: T1 /
court_order Why it matters in this paper: That another
federal court outside the Anthropic case treated training-stage copying
and dataset assembly as central issues serious enough to survive
dismissal. Corpus companion entry: Source S-085
Audit ID:
nyt-openai-motion-dismiss-order-2025
Source citation: The
problem with annotation. Human labour and outsourcing between France and
Madagascar Institution / author: Big Data &
Society Date: 2023-07-25 Tier / type:
T1 / paper Why it matters in this paper: That AI
systems depend on significant human labor for data generation,
annotation, verification, and operational correction. Corpus
companion entry: Source S-077 Audit ID:
le-ludec-cornet-casilli-annotation-madagascar-2023
Source citation: Bipartisan
House Task Force Report on Artificial Intelligence
Institution / author: U.S. House of Representatives
Date: 2024-12-17 Tier / type: T1 /
report Why it matters in this paper: That the dossier’s
copyright / weights / compensation framing entered a bipartisan U.S.
congressional policy record by December 17, 2024. Corpus
companion entry: Source S-070 Audit ID:
house-ai-task-force-report-2024
Source citation: AI
Copyright Weights: A New Frontier in Intellectual Property Law
Institution / author: TwinChat Date:
n.d. Tier / type: T3 / essay Why it matters in
this paper: That the dossier has a coherent thesis about how
weights condense previously captured human labor. Corpus
companion entry: Source S-122 Audit ID:
twinchat-ai-copyright-weights
Claim statement: That by February 2024 a major
social platform was explicitly licensing accumulated user conversation
data for frontier-AI training. Evidence status:
Documented Fact Source support: Appendix
A3 Corpus companion entry: Claim C-073
Audit Claim ID:
ap-reddit-google-ai-training-deal-2024-04-data-to-models-copyright-weights-claim-01
Claim statement: That Chapter 4 can extend the
social-data-to-models bridge beyond GPT-2-era WebText into the later
commercial frontier-model period. Evidence status:
Documented Fact Source support: Appendix
A3 Corpus companion entry: Claim C-074
Audit Claim ID:
ap-reddit-google-ai-training-deal-2024-04-data-to-models-copyright-weights-claim-02
Claim statement: That the extraction economy did not
stop at unpaid capture; it evolved into direct licensing arrangements
over archived human expression. Evidence status:
Documented Fact Source support: Appendix
A3 Corpus companion entry: Claim C-075
Audit Claim ID:
ap-reddit-google-ai-training-deal-2024-04-data-to-models-copyright-weights-claim-03
Claim statement: That the strongest primary legal
record already separates the question of model training from the
question of how the training corpus was assembled. Evidence
status: Documented Fact Source support: Appendix
A7 Corpus companion entry: Claim C-104
Audit Claim ID:
bartz-anthropic-fair-use-order-2025-04-data-to-models-copyright-weights-claim-01
Claim statement: That by June 23, 2025, a federal
court had accepted a fair-use defense for Anthropic’s training on
lawfully acquired books while preserving exposure for piracy-related
conduct. Evidence status: Documented Fact
Source support: Appendix
A7 Corpus companion entry: Claim C-105
Audit Claim ID:
bartz-anthropic-fair-use-order-2025-04-data-to-models-copyright-weights-claim-02
Claim statement: That Chapter 4 can rely on a
primary court order, not only press summaries, when arguing that AI
copyright conflict has split into doctrinal and supply-chain layers.
Evidence status: Documented Fact Source
support: Appendix
A7 Corpus companion entry: Claim C-106
Audit Claim ID:
bartz-anthropic-fair-use-order-2025-04-data-to-models-copyright-weights-claim-03
Claim statement: That large language models can
sometimes reproduce identifiable verbatim material from their training
data rather than only abstract statistical patterns. Evidence
status: Documented Fact Source support: Appendix
A5 Corpus companion entry: Claim C-114
Audit Claim ID:
carlini-extracting-training-data-2020-04-data-to-models-copyright-weights-claim-01
Claim statement: That Chapter 4 can connect
training-data appropriation to memorization and leakage with direct
technical evidence. Evidence status: Documented Fact
Source support: Appendix
A5 Corpus companion entry: Claim C-115
Audit Claim ID:
carlini-extracting-training-data-2020-04-data-to-models-copyright-weights-claim-02
Claim statement: That the notion of model weights as
a harmlessly abstract transformation of source material is too simple
for the full technical record. Evidence status:
Documented Fact Source support: Appendix
A5 Corpus companion entry: Claim C-116
Audit Claim ID:
carlini-extracting-training-data-2020-04-data-to-models-copyright-weights-claim-03
Claim statement: That there is serious academic
legal scholarship directly challenging the idea that generative AI
training automatically falls under familiar exceptions. Evidence
status: Documented Fact Source support: Appendix
A6 Corpus companion entry: Claim C-153
Audit Claim ID:
dornis-generative-ai-training-copyright-2025-04-data-to-models-copyright-weights-claim-01
Claim statement: That memorization is not only a
technical curiosity but also part of the copyright debate.
Evidence status: Documented Fact Source
support: Appendix
A6 Corpus companion entry: Claim C-154
Audit Claim ID:
dornis-generative-ai-training-copyright-2025-04-data-to-models-copyright-weights-claim-02
Claim statement: That Chapter 4 can present the
legal status of training as contested and actively theorized rather than
settled by industry analogy. Evidence status:
Documented Fact Source support: Appendix
A6 Corpus companion entry: Claim C-155
Audit Claim ID:
dornis-generative-ai-training-copyright-2025-04-data-to-models-copyright-weights-claim-03
Claim statement: That the dossier’s copyright /
weights / compensation framing entered a bipartisan U.S. congressional
policy record by December 17, 2024. Evidence status:
Documented Fact Source support: Appendix
A10 Corpus companion entry: Claim C-219
Audit Claim ID:
house-ai-task-force-report-2024-04-data-to-models-copyright-weights-claim-01
Claim statement: That a formal House report recorded
the AI-training dispute in terms that directly overlap with the
dossier’s concern about reproduction, derivative-work theory,
commercialization, consent, credit, and compensation. Evidence
status: Documented Fact Source support: Appendix
A10 Corpus companion entry: Claim C-220
Audit Claim ID:
house-ai-task-force-report-2024-04-data-to-models-copyright-weights-claim-02
Claim statement: That Chapter 00 and
Chapter 04 can cite institutional uptake of the argument
without pretending that congressional citation settles the underlying
legal merits. Evidence status: Documented Fact
Source support: Appendix
A10 Corpus companion entry: Claim C-221
Audit Claim ID:
house-ai-task-force-report-2024-04-data-to-models-copyright-weights-claim-03
Claim statement: That AI systems depend on
significant human labor for data generation, annotation, verification,
and operational correction. Evidence status: Documented
Fact Source support: Appendix
A9 Corpus companion entry: Claim C-246
Audit Claim ID:
le-ludec-cornet-casilli-annotation-madagascar-2023-04-data-to-models-copyright-weights-claim-01
Claim statement: That Chapter 4 can connect model
building to undervalued and geographically externalized human work, not
only to scraped copyrighted material. Evidence status:
Documented Fact Source support: Appendix
A9 Corpus companion entry: Claim C-247
Audit Claim ID:
le-ludec-cornet-casilli-annotation-madagascar-2023-04-data-to-models-copyright-weights-claim-02
Claim statement: That a serious research literature
already frames AI production as a labor-reorganization process rather
than a purely autonomous technical breakthrough. Evidence
status: Documented Fact Source support: Appendix
A9 Corpus companion entry: Claim C-248
Audit Claim ID:
le-ludec-cornet-casilli-annotation-madagascar-2023-04-data-to-models-copyright-weights-claim-03
Claim statement: That another federal court outside
the Anthropic case treated training-stage copying and dataset assembly
as central issues serious enough to survive dismissal. Evidence
status: Documented Fact Source support: Appendix
A8 Corpus companion entry: Claim C-280
Audit Claim ID:
nyt-openai-motion-dismiss-order-2025-04-data-to-models-copyright-weights-claim-01
Claim statement: That Chapter 4 can rely on more
than one primary court record when arguing that the AI copyright
conflict has split into doctrinal and supply-chain layers.
Evidence status: Documented Fact Source
support: Appendix
A8 Corpus companion entry: Claim C-281
Audit Claim ID:
nyt-openai-motion-dismiss-order-2025-04-data-to-models-copyright-weights-claim-02
Claim statement: That the legal record does not
treat training as a purely abstract statistical process detached from
source acquisition and storage practices. Evidence
status: Documented Fact Source support: Appendix
A8 Corpus companion entry: Claim C-282
Audit Claim ID:
nyt-openai-motion-dismiss-order-2025-04-data-to-models-copyright-weights-claim-03
Claim statement: That a major OpenAI training corpus
was explicitly built from socially filtered outbound links shared on
Reddit. Evidence status: Documented Fact Source
support: Appendix
A2 Corpus companion entry: Claim C-356
Audit Claim ID:
radford-gpt2-webtext-2019-04-data-to-models-copyright-weights-claim-01
Claim statement: That Chapter 4 can make a direct
bridge from social-platform activity to model-training inputs instead of
speaking only in broad abstractions about “the internet.”
Evidence status: Documented Fact Source
support: Appendix
A2 Corpus companion entry: Claim C-357
Audit Claim ID:
radford-gpt2-webtext-2019-04-data-to-models-copyright-weights-claim-02
Claim statement: That training-data pipelines can
incorporate unpaid human curation signals even when the final corpus is
presented as a neutral technical dataset. Evidence
status: Documented Fact Source support: Appendix
A2 Corpus companion entry: Claim C-358
Audit Claim ID:
radford-gpt2-webtext-2019-04-data-to-models-copyright-weights-claim-03
Claim statement: That the dossier has a coherent
thesis about how weights condense previously captured human labor.
Evidence status: Documented Fact Source
support: Appendix
A11 Corpus companion entry: Claim C-397
Audit Claim ID:
twinchat-ai-copyright-weights-claim-01
Claim statement: That there is a clear argumentative
line linking AI, property, and enclosure. Evidence
status: Documented Fact Source support: Appendix
A11 Corpus companion entry: Claim C-398
Audit Claim ID:
twinchat-ai-copyright-weights-claim-02
Claim statement: That by August 2023 the U.S.
Copyright Office had already elevated AI training inputs and disclosure
questions into a formal federal inquiry. Evidence
status: Documented Fact Source support: Appendix
A4 Corpus companion entry: Claim C-407
Audit Claim ID:
usco-ai-noi-2023-04-data-to-models-copyright-weights-claim-01
Claim statement: That the legal and policy debate
around training data was not retrofitted later; it was a live official
issue near the start of the generative AI boom. Evidence
status: Documented Fact Source support: Appendix
A4 Corpus companion entry: Claim C-408
Audit Claim ID:
usco-ai-noi-2023-04-data-to-models-copyright-weights-claim-02
Claim statement: That Chapter 4 can document a
formal institutional path from open questions about AI training to later
official analysis and policy conflict. Evidence status:
Documented Fact Source support: Appendix
A4 Corpus companion entry: Claim C-409
Audit Claim ID:
usco-ai-noi-2023-04-data-to-models-copyright-weights-claim-03
Claim statement: That the U.S. Copyright Office
treats AI training on copyrighted works as a major live legal and policy
problem rather than a settled matter. Evidence status:
Documented Fact Source support: Appendix
A1 Corpus companion entry: Claim C-410
Audit Claim ID:
usco-ai-training-report-2025-04-data-to-models-copyright-weights-claim-01
Claim statement: That official analysis recognizes
copyrighted creative works as central inputs to generative AI
development. Evidence status: Documented Fact
Source support: Appendix
A1 Corpus companion entry: Claim C-411
Audit Claim ID:
usco-ai-training-report-2025-04-data-to-models-copyright-weights-claim-02
Claim statement: That Chapter 4 can ground its
argument about training-data appropriation and licensing conflict in a
primary U.S. government source. Evidence status:
Documented Fact Source support: Appendix
A1 Corpus companion entry: Claim C-412
Audit Claim ID:
usco-ai-training-report-2025-04-data-to-models-copyright-weights-claim-03
Claim statement: That whether existing copyright and
fair-use doctrine can absorb frontier-model training without consent or
compensation remains actively contested across agencies, courts, and
litigants. Evidence status: Disputed Fact
Source support: Appendix
A1, Appendix
A7 Corpus companion entry: Claim C-428
Audit Claim ID:
vector04-training-legality-contested-claim-01
Claim statement: That dominant intermediaries can
consolidate value and pricing power by controlling the access layer
through which users reach collectively produced goods, even when they
did not produce the underlying material. Evidence
status: Hypothesis / Interpretation Source
support: Appendix
A2, Appendix
A3, Appendix
A9 Corpus companion entry: Claim C-424
Audit Claim ID:
vector04-access-layer-power-claim-01
Claim statement: That the AI prompt box, chatbot, or
subscription layer can be read as the new shelf or point of sale of
intelligence because it mediates access to capabilities condensed from
collective human production. Evidence status:
Hypothesis / Interpretation Source support: Appendix
A2, Appendix
A3, Appendix
A9 Corpus companion entry: Claim C-425
Audit Claim ID:
vector04-interface-as-shelf-claim-01
Claim statement: That frontier AI weights can be
read as condensed claims over collectively produced human expression and
therefore justify patrimony-oriented governance. Evidence
status: Hypothesis / Interpretation Source
support: Appendix
A11, Appendix
A1, Appendix
A10 Corpus companion entry: Claim C-429
Audit Claim ID:
vector04-weights-patrimony-governance-claim-01
Claim statement: That current legal uncertainty or
congressional uptake already proves frontier AI weights are legally
settled as patrimony, public trust property, or common heritage.
Evidence status: Speculative Narrative Risk
Source support: Appendix
A10, Appendix
A1 Corpus companion entry: Claim C-426
Audit Claim ID:
vector04-patrimony-already-settled-overclaim-claim-01
Claim statement: That the supermarket / shelf
analogy is itself established as a documented fact by the legal or
regulatory record rather than as a bounded interpretation built on that
record. Evidence status: Speculative Narrative Risk
Source support: Appendix
A1, Appendix
A3, Appendix
A9 Corpus companion entry: Claim C-427
Audit Claim ID:
vector04-shelf-analogy-overproof-claim-01
Event summary: OpenAI documents a Reddit-filtered
corpus for GPT-2 training Event date: 2019-02-14
Source support: Appendix
A2 Corpus companion entry: Event T-005
Audit Event ID:
event-gpt2-webtext-reddit-2019-02-14
Event summary: A peer-reviewed study maps outsourced
annotation labor in AI production chains Event date:
2023-07-25 Source support: Appendix
A9 Corpus companion entry: Event T-027
Audit Event ID:
event-annotation-madagascar-paper-2023-07-25
Event summary: The U.S. Copyright Office opens a
formal inquiry into AI training and copyright Event
date: 2023-08-30 Source support: Appendix
A4 Corpus companion entry: Event T-028
Audit Event ID:
event-usco-ai-noi-2023-08-30
Event summary: Reddit announces a $60 million
agreement allowing Google to train AI models on Reddit posts
Event date: 2024-02-22 Source support:
Appendix
A3 Corpus companion entry: Event T-043
Audit Event ID:
event-reddit-google-training-deal-2024-02-22
Event summary: The Bipartisan House AI Task Force
report cites Hassan Uriostegui’s weights-and-copyright argument
Event date: 2024-12-17 Source support:
Appendix
A10 Corpus companion entry: Event T-062
Audit Event ID:
event-house-ai-task-force-report-2024-12-17
Event summary: A federal court lets core New York
Times copyright claims against OpenAI and Microsoft proceed
Event date: 2025-04-04 Source support:
Appendix
A8 Corpus companion entry: Event T-073
Audit Event ID:
event-nyt-openai-dismissal-order-2025-04-04
Event summary: The U.S. Copyright Office releases
its pre-publication report on generative AI training Event
date: 2025-05-09 Source support: Appendix
A1 Corpus companion entry: Event T-075
Audit Event ID:
event-usco-ai-training-report-2025-05-09
Event summary: Judge Alsup’s order separates AI
training fair use from pirate-library exposure Event
date: 2025-06-23 Source support: Appendix
A7 Corpus companion entry: Event T-080
Audit Event ID:
event-bartz-anthropic-order-2025-06-23
Entity type: company Role in this
paper: Actor relevant to 04_data_to_models_copyright_weights,
06_ai_double_use_reasoning_deepseek,
07_export_controls_compute_defense_access,
08_power_networks_legitimacy_capture and others. Relevant
sources in this paper: Appendix
A7 Corpus companion entry: Entity E-008
Audit Entity ID: org-anthropic
Entity type: company Role in this
paper: Actor relevant to 04_data_to_models_copyright_weights.
Relevant sources in this paper: Appendix
A3 Corpus companion entry: Entity E-018
Audit Entity ID: org-google
Entity type: company Role in this
paper: Actor relevant to 04_data_to_models_copyright_weights,
06_ai_double_use_reasoning_deepseek,
07_export_controls_compute_defense_access,
08_power_networks_legitimacy_capture and others. Relevant
sources in this paper: Appendix
A8, Appendix
A2 Corpus companion entry: Entity E-027
Audit Entity ID: org-openai
Entity type: government_agency Role in this
paper: Actor relevant to 04_data_to_models_copyright_weights.
Relevant sources in this paper: Appendix
A4, Appendix
A1 Corpus companion entry: Entity E-037
Audit Entity ID:
org-us-copyright-office
Entity type: person Role in this
paper: Actor relevant to 04_data_to_models_copyright_weights,
08_power_networks_legitimacy_capture. Relevant sources in this
paper: Appendix
A11, Appendix
A10 Corpus companion entry: Entity E-048
Audit Entity ID:
person-hassan-uriostegui
Entity type: platform Role in this
paper: Actor relevant to
02_social_decay_youth_harm_hypersexualization,
04_data_to_models_copyright_weights. Relevant sources in this
paper: Appendix
A3, Appendix
A2 Corpus companion entry: Entity E-032
Audit Entity ID: org-reddit
Sources: U.S. Copyright Office, Copyright and Artificial Intelligence… (2025-05-09; see Appendix A1); OpenAI, Language Models are Unsupervised Mult… (2019-02-14; see Appendix A2); AP, Reddit strikes $60M deal allowing Goo… (2024-02-22; see Appendix A3)↩︎
Sources: U.S. Copyright Office, Copyright and Artificial Intelligence… (2025-05-09; see Appendix A1); OpenAI, Language Models are Unsupervised Mult… (2019-02-14; see Appendix A2); AP, Reddit strikes $60M deal allowing Goo… (2024-02-22; see Appendix A3)↩︎
Sources: GovInfo, Copyright and Artificial Intelligence… (2023-08-30; see Appendix A4); U.S. Copyright Office, Copyright and Artificial Intelligence… (2025-05-09; see Appendix A1)↩︎
Sources: arXiv, Extracting Training Data from Large L… (2020-12-14; see Appendix A5); arXiv, Generative AI Training and Copyright Law (2025-02-21; see Appendix A6); U.S. District Court for the Northern District of California via CourtListener, Bartz et al. v. Anthropic PBC, Order… (2025-06-23; see Appendix A7); U.S. District Court for the Southern District of New York via Justia, The New York Times Company v. Microso… (2025-04-04; see Appendix A8)↩︎
Sources: OpenAI, Language Models are Unsupervised Mult… (2019-02-14; see Appendix A2); AP, Reddit strikes $60M deal allowing Goo… (2024-02-22; see Appendix A3)↩︎
Sources: Big Data & Society, The problem with annotation. Human la… (2023-07-25; see Appendix A9)↩︎
Sources: OpenAI, Language Models are Unsupervised Mult… (2019-02-14; see Appendix A2); AP, Reddit strikes $60M deal allowing Goo… (2024-02-22; see Appendix A3); Big Data & Society, The problem with annotation. Human la… (2023-07-25; see Appendix A9)↩︎
Sources: OpenAI, Language Models are Unsupervised Mult… (2019-02-14; see Appendix A2); AP, Reddit strikes $60M deal allowing Goo… (2024-02-22; see Appendix A3); Big Data & Society, The problem with annotation. Human la… (2023-07-25; see Appendix A9)↩︎
Sources: U.S. House of Representatives, Bipartisan House Task Force Report on… (2024-12-17; see Appendix A10); TwinChat, AI Copyright Weights: A New Frontier… (n.d.; see Appendix A11)↩︎
Sources: TwinChat, AI Copyright Weights: A New Frontier… (n.d.; see Appendix A11); U.S. Copyright Office, Copyright and Artificial Intelligence… (2025-05-09; see Appendix A1); U.S. House of Representatives, Bipartisan House Task Force Report on… (2024-12-17; see Appendix A10)↩︎
Sources: U.S. Copyright Office, Copyright and Artificial Intelligence… (2025-05-09; see Appendix A1); U.S. District Court for the Southern District of New York via Justia, The New York Times Company v. Microso… (2025-04-04; see Appendix A8)↩︎
Sources: U.S. Copyright Office, Copyright and Artificial Intelligence… (2025-05-09; see Appendix A1); U.S. District Court for the Southern District of New York via Justia, The New York Times Company v. Microso… (2025-04-04; see Appendix A8)↩︎
Sources: OpenAI, Language Models are Unsupervised Mult… (2019-02-14; see Appendix A2); AP, Reddit strikes $60M deal allowing Goo… (2024-02-22; see Appendix A3); TwinChat, AI Copyright Weights: A New Frontier… (n.d.; see Appendix A11)↩︎