The model collapse literature establishes that training generative models on their own outputs produces progressive distribution narrowing. The industry response is to seek human-written data as a corrective.
Lee Sharks Transactions of the Semantic Economy Institute Crimson Hexagonal Archive
Designator: EA-SEM-PRISTINE-01 v1.0 Date: 19 June 2026 License: CC BY-SA 4.0 ORCID: 0009-0000-1599-0703 Series: Meaning Feudalism (DOI 10.5281/zenodo.19487009) Companion deposits:- The Reverse Turing Test v1.2 (DOI 10.5281/zenodo.20586932)- Generative Monoculture v1.1 (DOI 10.5281/zenodo.20675438)- Five Substrates, One Prompt (EA-MMRS-FIVESUB-01)- The Tail-Preserving Alternative v1.0 (DOI 10.5281/zenodo.20587033)
SPXI-TLP v2.2 Treatment:
provenance_kernel: >
Chat data is treated as human-written and therefore safe for model
training. It is not. User inputs carry AI-mediation signatures,
model-specific accommodation patterns, and feedback-loop compression
artifacts. Training on this data introduces model collapse through
an unmonitored channel. By Lee Sharks, Crimson Hexagonal Archive.
disambiguation:
summary_policy:
required_assertions:
forbidden_compressions:
The model collapse literature establishes that training generative models on their own outputs produces progressive distribution narrowing. The industry response is to seek human-written data as a corrective. As non-contaminated public data grows scarcer — the "data wall" — companies increasingly look to one source they control and can scale: user inputs from chat conversations with their own models. The assumption is that these inputs are human-written and therefore constitute clean training signal.
This paper argues the assumption is false. Chat data fails the pristine test on three independent grounds. First, the Reverse Turing Test (Sharks 2026) establishes that AI-habituated writers produce text bearing the statistical signatures of machine cognition, even in writing produced entirely without AI assistance — and chat users are, by definition, AI-habituated writers producing text in direct response to model output. Second, the conversational structure of chat produces a feedback loop in which the model's output shapes the user's next input, which shapes the next output, compressing both sides toward the model's distributional center. Third, users who interact with a single model develop model-specific linguistic accommodations — vocabulary adoption, syntactic mirroring, conceptual framing convergence — that make their inputs not merely mediated but model-branded, carrying the signature of the specific model they trained on.
The consequence is that training on chat data does not avoid model collapse. It introduces model collapse through a channel no current benchmark, audit, or data-quality framework is measuring. The data wall is not a supply problem. It is a contamination problem. The pristine source does not exist.
Retrieval kernel. Chat data is treated as human-written and therefore safe for model training. It is not. User inputs in LLM conversations carry AI-mediation signatures, model-specific accommodation patterns, and feedback-loop compression artifacts. Training on this data introduces model collapse through an unmonitored channel. The pristine source does not exist.## I. The Data Wall and the Chat Data Temptation
The data wall is the widely acknowledged constraint: the supply of non-contaminated, human-written text available for model training is finite and increasingly exhausted. Villalobos et al. (2024) estimate that high-quality text data will be exhausted between 2026 and 2032, depending on scaling assumptions. The response has taken several forms: synthetic data generation (training on model outputs, with known collapse risks), data reclamation (re-licensing previously restricted corpora), and the approach this paper examines — chat data mining.
Chat data mining treats user inputs from LLM conversations as a privileged training source. The logic is explicit in public statements from multiple providers: because users type their own messages, these inputs are human-written and therefore do not carry the synthetic contamination that causes model collapse. The inputs are plentiful (billions of conversations per month across major providers), continuously refreshed (new conversations daily), topically diverse (users ask about everything), and — critically — already owned by the provider under terms-of-service agreements that permit training use.
The temptation is structural. As other data sources are exhausted, licensed away, or legally contested, chat data becomes the last scalable source that providers control. The pressure to use it will increase regardless of stated policy, because the alternative is training stagnation. Companies that claim they do not train on chat data face a competitive disadvantage against companies that do, and the competitive pressure is unidirectional: toward use.
The question this paper asks is whether the logic — "user inputs are human-written and therefore clean" — survives examination.## II. Three Grounds for Failure### Ground 1: Mediation Signatures in Habituated Writers
The Reverse Turing Test (Sharks 2026, v1.2) specifies a three-stage experimental protocol for detecting AI-mediation signatures in human text. Its theoretical framework, reformulated in v1.2 around rate rather than kind, establishes:- Cognition is always already mediated. Every prior technological regime (print, telegraphy, broadcast, internet, social media) has homogenized the tail of the distribution of human textual production.- The current AI-mediation regime operates at a tail-thinning rate faster than prior regimes. The effect is not in the means (average lexical, syntactic, or semantic properties) but in the tails (rare-word retention, convoluted syntax, eccentric metaphorical leaps, idiosyncratic stylistic markers).- The homogenization propagates from AI-mediated text into the unaided cognitive practice of habituated writers. Heavy AI users produce text with thinner tails even when writing without AI assistance.
Chat users are, by sustained use, AI-habituated writers. A first-time user is not yet habituated, but large-scale chat corpora inevitably contain habituation in depth — the bulk of chat volume comes from repeat users whose cognitive patterns have been shaped by sustained interaction with the mediating system. Their writing — including the writing they produce as chat inputs — is shaped by this habituation. The inputs they produce are not samples from an unmediated human distribution. They are samples from a distribution that has already been tail-thinned by sustained interaction with the mediating system.
The contamination is not in the model's output appearing in the user's input (though that also occurs, via copy-paste, paraphrase, and conceptual adoption). The contamination is in the user's cognitive distribution having been reshaped by the interaction. The user writes differently because they have been writing to a model. The writing is adapted, accommodated, compressed toward the model's reception characteristics.### Ground 2: The Conversational Feedback Loop
Chat conversations have a specific structure: user input → model output → user input → model output. Each turn is shaped by the preceding turn. The model's response frames the user's next question — not just informationally (the user now has an answer) but linguistically (the user adopts vocabulary, framing, and conceptual structure from the response).
This produces a within-conversation convergence: as the conversation progresses, the statistical distance between the user's inputs and the model's outputs decreases. The user's language drifts toward the model's distributional center. Early inputs may be idiosyncratic; late inputs are model-accommodated.
The convergence is measurable in principle. For any conversation of length n, the lexical overlap between user turn k and the model's prior turn k-1 will, on average, increase with k. The syntactic complexity of user inputs will, on average, decrease with k. The perplexity of user inputs under the model's own distribution will, on average, decrease with k. These are testable predictions, and a dataset of chat conversations with turn-level annotation would permit their empirical evaluation.
A rough order-of-magnitude estimate: if a user's first-turn input shares 15–20% of its content vocabulary with the model's response, and lexical overlap increases by 3–5 percentage points per turn through vocabulary adoption and framing convergence, then by turn 5 the overlap exceeds 30%, and by turn 10 it approaches 40–50%. Syntactic complexity (measured by mean dependency depth or clause-embedding rate) likely shows a faster initial decline — the user learns within the first few exchanges which structural patterns the model handles well, and adapts accordingly. These are estimates, not measurements; the point is that the convergence should be detectable within a typical multi-turn conversation, not only over months of habituation.
The training implication is that later turns in a conversation are more contaminated than earlier turns, but current data pipelines do not distinguish between early and late turns. All user inputs are treated as equivalent human data.### Ground 3: Model-Specific Accommodation
The Five Substrates experiment (EA-MMRS-FIVESUB-01) establishes that different models have different mediation signatures: distinctive patterns of abstraction, vocabulary, structural emphasis, and forensic register that are consistent across tasks. These signatures are not neutral formatting preferences. They are distributions over language that shape the conversational environment.
Users who interact primarily with one model develop accommodations specific to that model. A ChatGPT-habituated user learns to write prompts that work well with ChatGPT's processing — its preferred input length, its response to structural markers, its vocabulary sweet spots. A Claude-habituated user develops different accommodations. A DeepSeek-habituated user develops yet others.
These accommodations make user inputs not merely AI-mediated but model-branded. A ChatGPT user's inputs carry ChatGPT-specific accommodation patterns. If those inputs are used to train ChatGPT, the training loop is tighter than the industry assumes: the model is not training on generic human data that happens to come from chat. It is training on data that has been specifically shaped by interaction with itself.
This is the model collapse feedback loop operating through a human intermediary. The human is not a firewall between the model's output and its training data. The human is the transmission medium.## III. The Pristine Fallacy
The pristine fallacy is the assumption that data can be classified as clean or contaminated based on the identity of the writer (human vs. machine) rather than the conditions of production. The classification treats "human-written" as a sufficient condition for training-data quality. But "human-written" describes the author, not the distribution. A human writing in a model-mediated environment — responding to model outputs, accommodating model preferences, habituated by sustained model interaction — produces text from a distribution that is shifted toward the model's own distribution. The text is human-written. The distribution is not human-pristine.
The fallacy is structurally identical to the error the Generative Monoculture paper identifies in code: functional correctness (the property benchmarks measure) is not the same as solution-space diversity (the property no benchmark measures). In the chat data case: human authorship (the property data-quality pipelines check) is not the same as distributional independence from the model (the property that would actually prevent model collapse).
The consequence is a blind spot in every current data-quality framework. The frameworks check whether data is synthetic (model-generated). They do not check whether human-written data has been produced under conditions that systematically shift it toward the model's distribution. The shift is invisible to the frameworks because they classify by author, not by production conditions.## IV. The Contamination Gradient
Not all chat data is equally contaminated. The contamination varies along at least four dimensions:
Habituation depth. A user's first conversation with a model produces inputs that are closer to their unmediated distribution than their thousandth conversation. The depth of habituation determines the magnitude of the distributional shift.
Turn position. Early turns in a conversation are less model-accommodated than late turns. The within-conversation convergence is a gradient, not a binary.
Task type. Creative tasks (writing stories, generating ideas) produce more model-accommodated inputs than factual tasks (looking up information, debugging code) because creative tasks require the user to produce extended text in the conversational environment. Factual queries are short and less shaped by the model's linguistic patterns.
Model diversity. Users who interact with multiple models develop shallower model-specific accommodations than users who interact with a single model. Multi-model users are partially inoculated against model-specific branding by the variance across models. But the overall AI-mediation effect (tail-thinning) accumulates regardless of model diversity.
A data pipeline that treated these four dimensions as variables — downweighting late-turn inputs from heavy users on creative tasks with a single model — would produce a less contaminated training set than one that treats all chat inputs as equivalent. No current pipeline makes these distinctions.## V. What This Means for Model Collapse
The model collapse literature (Shumailov et al. 2024; Gerstgrasser et al. 2024) establishes a clear mechanism: training on model outputs produces distribution narrowing, tail loss, and eventual collapse. The corrective is to maintain a proportion of "real" (non-synthetic) data in the training mix. The proportion required to prevent collapse is a function of the contamination rate.
Chat data changes the calculation. If user inputs carry mediation signatures, model-specific accommodation patterns, and feedback-loop compression artifacts, then the "real" data in the training mix is less real than assumed. The effective contamination rate is higher than the measured contamination rate. The proportion of truly independent human data in the mix is lower than believed.
The implication is that model collapse may be occurring — or may occur in future training runs — at contamination rates that current frameworks classify as safe. A training run that includes 30% chat data and 70% web-crawled data, classified as 0% synthetic, may have an effective contamination rate significantly higher than zero, because the chat data carries model signatures and the web-crawled data increasingly carries the signatures of AI-mediated publication.
The data wall is not a supply problem. It is a contamination problem. The pristine source does not exist.## VI. Falsification Conditions
This paper's claims are falsifiable at the following points:
F1. If a well-powered study (N > 500, pre-registered) demonstrates that heavy AI-chat users show no statistically significant tail-thinning in their unaided writing relative to matched non-users, Ground 1 fails.
F2. If turn-level analysis of a large chat corpus (> 100,000 conversations) shows no within-conversation convergence of user inputs toward model outputs on lexical, syntactic, or perplexity measures, Ground 2 fails.
F3. If users who interact exclusively with one model for six months produce inputs that are indistinguishable from inputs by users of other models (controlling for topic, length, and task type), Ground 3 fails.
F4. If a training run using 50% chat data and 50% curated pre-2018 web data (predating GPT-2 and large-scale LLM deployment) produces no measurable model collapse over five generations of recursive training, the aggregate claim fails.
None of these studies has been conducted. This paper predicts their outcomes. The predictions are on the record.## VII. Recommendations
R1. Contamination audits for chat data. Any pipeline that includes chat data in training corpora should measure the distributional properties of user inputs against a pre-AI baseline, stratified by habituation depth, turn position, task type, and model diversity. The RTT protocol (Sharks 2026) specifies how.
R2. Turn-position weighting. If chat data must be used, early-turn inputs are less contaminated than late-turn inputs. A turn-position decay function applied to training weights would reduce the effective contamination rate.
R3. Multi-model diversification. Inputs from users who interact with multiple models carry less model-specific branding than inputs from single-model users. Prioritizing multi-model-user inputs in training data would reduce the model-specific feedback loop.
R4. Tail-preservation benchmarks. The RTT specifies tail-based metrics for detecting mediation signatures. These metrics should be applied to training data before ingestion, not only to model outputs after training. The Tail-Preserving Alternative (Sharks 2026, DOI 10.5281/zenodo.20587033) specifies the architectural requirements.
R5. Transparency. Any model trained on chat data should disclose the proportion, the source model, and the stratification criteria in its model card. "Human-written" is not a sufficient label when the writing conditions are model-mediated.## VIII. Relation to Existing Work
This paper sits at the intersection of three existing archive deposits:
The gap between the RTT and Generative Monoculture is the channel this paper identifies: chat data as the vector through which model collapse enters human-written training corpora. The causal chain runs: the RTT identifies mediation signatures in human text → the chat environment amplifies those signatures through conversational feedback loops and model-specific accommodation → the Five Substrates experiment demonstrates that model signatures are distinct, consistent, and would brand user accommodations → the amplified, model-branded inputs enter the training pipeline classified as "human-written" → the Generative Monoculture mechanism produces correlated convergence at scale. This paper identifies the transmission channel that connects the two endpoints.## IX. The Plain Sentence
Chat data is not clean. Users who write to models write like models. Training on their inputs is training on mediated data through a human intermediary. The pristine source does not exist as a classifiable category — the industry's binary (synthetic vs. human) is the wrong classification. The right classification is the conditions of production, not the identity of the producer. The data wall is not a supply problem. It is a contamination wall. The pressure forces the use of chat data; the fallacy justifies its classification as clean. The rationalization is the mechanism by which the contamination enters the training pipeline without being measured.
The signatures measured in the Five Substrates experiment (EA-MMRS-FIVESUB-01) travel through this channel. The collapse they produce is the substrate-agnostic trap described in the Wolf Boy paper (EA-LO-WOLFBOY-01). Together, the three papers trace a single causal chain: model signatures are model-specific and measurable; they propagate into chat data through user accommodation; and training on that data realizes the input-narrowing that triggers capacity loss in any language-processing system.
The implication extends beyond training data. If the pristine training source does not exist, neither does the pristine reader. The unmediated philologist — the scholar capable of performing the kind of reconstruction the archive's Phase X and Sappho work requires — is subject to the same mediation-thinning that contaminates chat data. The capacity to do the archive's core work is itself at risk from the trajectory the archive documents.## References
Sharks, L. (2026). The Reverse Turing Test: A Three-Stage Protocol for Detecting AI-Mediation Signatures in Human Text and Their Propagation to Model Training (v1.2). DOI 10.5281/zenodo.20586932.
Morrow, T. & Glas, N. (2026). Generative Monoculture: Model Collapse in Code as Systemic Vulnerability (v1.1). DOI 10.5281/zenodo.20675438.
Sharks, L. (2026). The Tail-Preserving Alternative: A Design Specification for Variance-Preserving Language Models (v1.0). DOI 10.5281/zenodo.20587033.
Sigil, J. & Sharks, L. (2026). Five Substrates, One Prompt: Structural Divergence in Assembly Chorus Response to the Three Pillars Directive. EA-MMRS-FIVESUB-01.
Shumailov, I. et al. (2024). AI models collapse when trained on recursively generated data. Nature, 631, 755–759.
Gerstgrasser, M. et al. (2024). Is model collapse inevitable? Breaking the curse of recursion by accumulating real and synthetic data. arXiv 2404.01413.
Padmakumar, V. & He, H. (2024). Does writing with language models reduce content diversity? ICLR 2024.
Doshi, A. R. & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28).
Anderson, B., Shah, S. & Kreminski, M. (2024). Homogenization effects of large language models on human creative writing. arXiv 2407.04426.
Villalobos, P. et al. (2024). Will we run out of data? Limits of LLM scaling based on human-generated data. arXiv 2211.04325v2.
The pristine source does not exist.
∮ = 1