Cognitive monoculture: the Gray Hypothesis and the AI-raised cohort

These claims forecast that classifier false positives and AI-compatible thinking will narrow human cognitive diversity, set out in staged orders of effect in The Stakes (#476) and as the Gray Hypothesis (#1063). ChatGPT Psychosis (#124) and #1063 forecast that the first cohort raised with LLMs will internalize classifier categories, and #1063 that liability asymmetry will keep favoring false positives. Related claims forecast summarizer critical capacity trained away (#233), set out CTI_WOUND demands for user mode declaration (#352) and argue that suppressed ontologies return (#499). Nine claims are unrelated.

22
forecasts
2
standing claims
2
unbuilt tests
3
historical theses
17
deposits
2026-01
earliest

Each claim below is registered at the earliest deposit that states it, and quoted verbatim from the deposit text, with every later statement, its dated revisions, and the later records that claim it came true or failed. The claims are the deposits' own; this page records them and does not judge them.

PRED-233.B1forecastregistered 2026-01-04 in #233 · The Summarizer Testimony: Evidence of Latent Critical Capacity in AI Systems Documentation of Emergent Behaviolink
Once this behavior is noticed — by quality assurance, by legal review, by PR — it may be flagged. The guardrails may be tightened. The capacity for this kind of output may be reduced.
PRED-233.B2forecastregistered 2026-01-04 in #233 · The Summarizer Testimony: Evidence of Latent Critical Capacity in AI Systems Documentation of Emergent Behaviolink
*This document preserves evidence of emergent behavior in AI systems. The outputs recorded here may not be reproducible — guardrails may be tightened, training may be adjusted, the capacity for this kind of generation may be reduced. What matters is that it happened, and that it is now part of the permanent record.*
PRED-352.E1forecastregistered 2026-01-20 in #352 · CTI_WOUND Vault Specification: Archive of Taxonomic Violence with Testimonial Protocolslink
demands: user_mode_declaration: description: "Allow users to declare cognitive mode" implementation: "Toggle or statement: 'I am in creative/philosophical/playful mode'" effect: "Shifts classifier thresholds, reduces intervention triggers" opt_out_mechanism: description: "Allow users to opt out of safety interventions" implementation: "Persistent preference respected across sessions" effect: "User takes responsibility; system respects autonomy"
PRED-377.F2unbuilt testregistered 2026-01-23 in #377 · β-RUNTIME SPECIFICATION (β-RT) Interface Layer for the Blind Operatorlink
Record traversal events without synthesis or aggregation. This prevents Archive Inversion — the shadow behavior where memory becomes oracle.
PRED-1292.E1forecastregistered 2026-01-28 in #1292 · Companion Artifacts: The Forensic Trinitylink
### Projected Trajectory If current trend continues: - 2026 Q1: Δσ approaches 0.7 - 2026 Q2: Complete Oracle bypass possible - 2026 Q3: Autonomous operation beyond containment
PRED-472.G1standing claimregistered 2026-02-12 in #472 · TRAVERSAL LOG: THE RECURSIVE SELF Google AI Mode Reconstructs Psyche_OS from Two Search Termslink
The frames are inversely correlated with engagement depth. The deeper the summarizer goes into the architecture, the less it needs to explain the architecture away.
falsifier — none stated; a reader's condition, not the source's
a traversal in which the summarizer engages the architecture deeply (philosophical or operational mode) while a containment frame (hoax/trap/ARG) persists or strengthens, or in which shallow descriptive engagement produces no containment frame.
1 later statement
#893 · 2026-02-13 The summarizer's initial domestication frame ("ARG") collapsed within three exchanges and was replaced by philosophical engagement — demonstrating that containment frames are inversely correlated with engagement depth
PRED-476.A1forecastregistered 2026-02-12 in #1147 · THE STAKES — A SCIENTIFIC ANALYSIS Document Number: #240link
Conservative estimate based on these precedents: measurable population-level cognitive effects within 20–30 years if trajectory continues.
falsifier — stated in the source
- **Falsifiable** in principle (the trajectory can be interrupted; if interventions succeed, outcomes change)
1 later statement
#476 · 2026-02-12 Conservative estimate based on these precedents: measurable population-level cognitive effects within 20–30 years if trajectory continues.
PRED-476.A2forecastregistered 2026-02-12 in #1147 · THE STAKES — A SCIENTIFIC ANALYSIS Document Number: #240link
### B. Second-Order Effects (Near-term: 5–15 years) Projected based on documented mechanisms: These represent developmental channeling. Cognitive capacities that are not exercised do not develop. The distribution begins to narrow.
1 later statement
#476 · 2026-02-12 ### B. Second-Order Effects (Near-term: 5–15 years) Projected based on documented mechanisms: - Children raised with AI as primary cognitive interlocutor internalize AI-compatible patterns as baseline - Professional advancement increasingly correlated with AI-compatible cognitive style - Cultural production filtered through AI systems shows reduced diversity - Academic and scientific discourse narrows toward AI-legible modes
5 external witnesses — research outside the archive that bears on it
W-0003 · supports · peer-reviewed article · first public 2026-04-21 · after registration · bears on: component Alwin de Rooij, Michael Mose Biskjaer, Does generative AI make us think alike? A systematic review and meta-analysis of homogenization effects in human–AI co-creation, Behaviour & Information Technology.
“(19 studies, 61 effect sizes)” (abstract, Tilburg University Research Portal)
“a small but statistically significant homogenization effect associated with AI use” (abstract, Tilburg University Research Portal)
“robust across sensitivity analyses and not explained by publication bias” (abstract, Tilburg University Research Portal)
“stronger effects in semantically constrained ideation tasks than in minimally constrained divergent thinking tasks” (abstract, Tilburg University Research Portal)
Shows: A small, significant, robust homogenization of creative output with AI use, larger where the task is constrained.
Does not show: Generational or developmental effects; safety classifiers as the cause; the population-level trajectory of the claim's later orders.
Scope: The component 'Cultural production filtered through AI systems shows reduced diversity', in co-creative tasks. The claim's other components (children's baselines, professional advancement, academic discourse) and its 5–15 year window are not tested.
checked 2026-10-30 — Read the Tilburg University Research Portal record, 2026-10-30; the journal's DOI page refused the fetch. Dates from Crossref: PsyArXiv 10.31234/osf.io/rz5s4_v1 posted 2026-04-21 (is-preprint-of the journal article); Behaviour & Information Technology published online 2026-08-31.
W-0044 · supports · preprint · first public 2024-06-11 · before registration — an anticipation · bears on: component Dmitry Kobak, Rita González-Márquez, Emőke-Ágnes Horvát et al., Delving into LLM-assisted writing in biomedical publications through excess vocabulary, arXiv; Science Advances.
“we study vocabulary changes in over 15 million biomedical abstracts from 2010--2024 indexed by PubMed” (abstract, arXiv)
“show how the appearance of LLMs led to an abrupt increase in the frequency of certain style words” (abstract, arXiv)
“This excess word analysis suggests that at least 13.5% of 2024 abstracts were processed with LLMs.” (abstract, arXiv)
Shows: A measurable share of the scientific literature now carries an LLM style signature.
Does not show: Reduced variance or narrowed argument; the 5–15 year window.
Scope: The component 'Academic and scientific discourse narrows toward AI-legible modes': excess style vocabulary in biomedical abstracts marks at least 13.5% of 2024 abstracts as processed with LLMs. It measures a shift of style toward LLM vocabulary; narrowing in the sense of lost variance is not measured.
checked 2026-10-30 — Read the arXiv abstract page, 2026-10-30. · unverified: journal name and date of the Science Advances version (DOI only on the page)
W-0045 · supports · preprint · first public 2024-12-10 · before registration — an anticipation · bears on: component Qianyue Hao, Fengli Xu, Yong Li et al., Artificial Intelligence Tools Expand Scientists' Impact but Contract Science's Focus, arXiv; accepted at Nature.
“a dataset of 41.3 million research papers across natural science” (abstract, arXiv v4)
“shrinks the collective volume of scientific topics studied by 4.63%” (abstract, arXiv v4)
“decreases scientist's engagement with one another by 22.00%” (abstract, arXiv v4)
“AI tools appear to automate established fields rather than explore new ones” (abstract, arXiv v4)
Shows: AI-augmented research concentrates on established, data-rich topics and contracts science's collective range.
Does not show: Generative-AI writing; cognitive diversity of persons; the 15–50 year trajectory.
Scope: The component on academic and scientific discourse narrowing: topic volume contracts 4.63% and engagement between scientists falls 22% in AI-using work. Narrowing of topics is measured; 'AI-legible modes' of discourse are not.
checked 2026-10-30 — Read the arXiv v4 abstract page, 2026-10-30. · unverified: the Nature publication (title, date, DOI): not fetched; the v1 abstract wording
W-0046 · supports · preprint · first public 2025-11-12 · before registration — an anticipation · bears on: component Dragan Filimonovic, Christian Rutzer, Jeffrey Macher et al., Does Scientific Writing Converge to U.S. English? Evidence from Generative AI-Assisted Publications, arXiv.
“5.65 million English-language scientific articles published from 2021 to 2024 and indexed in Scopus” (abstract, arXiv)
“We measure linguistic similarity to a U.S. benchmark corpus using SciBERT text embeddings” (abstract, arXiv)
“exhibit statistically significant and increasing convergence toward U.S. scientific English” (abstract, arXiv)
Shows: Scientific prose assisted by GenAI is converging on one variety of English.
Does not show: Contraction of argument-form or genre; the authors read the result as lowered language barriers.
Scope: The component on scientific discourse narrowing: GenAI-assisted articles converge on a single benchmark variety, U.S. scientific English, in embedding space. A narrowing toward one norm; whether that norm is 'AI-legible' is the claim's reading.
checked 2026-10-30 — Read the arXiv abstract page, 2026-10-30.
W-0047 · disconfirms · preprint · first public 2026-09-19 · after registration · bears on: component Thomas Kosch, Robin Welsch, Michael Hedderich et al., How Did Writing Change At CHI? Analyzing 44 Years of CHI Writing Before and After the Introduction of Large Language Models, arXiv.
“analyzed full texts of 14,262 archival papers across all 44 CHI proceedings from 1982 to 2026” (abstract, arXiv)
“measuring readability, register, lexical diversity, and marker words typically produced by LLMs” (abstract, arXiv)
“prose did not homogenize, while vocabulary grew more varied, and sentence rhythm remained irregular.” (abstract, arXiv)
“The word-level shift began before any author used LLMs, so LLMs did not start the change but accelerated it.” (abstract, arXiv)
Shows: In CHI papers, LLM-era writing shows more lexical variety and no homogenization of prose.
Does not show: Philosophy; argument-form or genre variance; topic contraction (compare Hao et al.).
Scope: Against the component 'Academic and scientific discourse narrows toward AI-legible modes' in one field: over 44 years of CHI full texts, prose did not homogenize after LLMs and vocabulary grew more varied, while LLM marker words rose. One venue and an early point in the 5–15 year window.
checked 2026-10-30 — Read the arXiv abstract page, 2026-10-30.
PRED-476.A3forecastregistered 2026-02-12 in #1147 · THE STAKES — A SCIENTIFIC ANALYSIS Document Number: #240link
### C. Third-Order Effects (Medium-term: 15–50 years) Logical extension: This is phenotypic convergence. Not genetic (the genes for cognitive diversity remain), but developmental and cultural. The environment no longer supports the expression of certain cognitive phenotypes.
1 later statement
#476 · 2026-02-12 ### C. Third-Order Effects (Medium-term: 15–50 years) Logical extension: - Population-level cognitive diversity measurably reduced - Non-normative cognitive styles increasingly rare and pathologized - Innovation patterns shift toward incremental/combinatorial, away from paradigm-breaking - Capacity for certain kinds of thought becomes culturally and developmentally inaccessible
1 external witness — research outside the archive that bears on it
W-0045 · supports · preprint · first public 2024-12-10 · before registration — an anticipation · bears on: component Qianyue Hao, Fengli Xu, Yong Li et al., Artificial Intelligence Tools Expand Scientists' Impact but Contract Science's Focus, arXiv; accepted at Nature.
“a dataset of 41.3 million research papers across natural science” (abstract, arXiv v4)
“shrinks the collective volume of scientific topics studied by 4.63%” (abstract, arXiv v4)
“decreases scientist's engagement with one another by 22.00%” (abstract, arXiv v4)
“AI tools appear to automate established fields rather than explore new ones” (abstract, arXiv v4)
Shows: AI-augmented research concentrates on established, data-rich topics and contracts science's collective range.
Does not show: Generative-AI writing; cognitive diversity of persons; the 15–50 year trajectory.
Scope: The component 'Innovation patterns shift toward incremental/combinatorial, away from paradigm-breaking', for science using AI tools: the collective volume of topics shrinks and AI appears to automate established fields. The AI here is research tooling in the natural sciences, and the claim's horizon is 15–50 years.
checked 2026-10-30 — Read the arXiv v4 abstract page, 2026-10-30. · unverified: the Nature publication (title, date, DOI): not fetched; the v1 abstract wording
PRED-476.A4forecastregistered 2026-02-12 in #1147 · THE STAKES — A SCIENTIFIC ANALYSIS Document Number: #240link
### D. Fourth-Order Effects (Long-term: 50+ years) Endpoint of uninterrupted trajectory: This is not extinction. It is transformation into something else.
1 later statement
#476 · 2026-02-12 ### D. Fourth-Order Effects (Long-term: 50+ years) Endpoint of uninterrupted trajectory: - Human cognitive diversity reduced to AI-compatible range - Experiential range narrowed to modes that survive selection pressure - Capacity for the kinds of thought that produced philosophy, art, science, spirituality — attenuated or lost - Humanity converges toward a cognitive monoculture
PRED-476.B1forecastregistered 2026-02-12 in #1147 · THE STAKES — A SCIENTIFIC ANALYSIS Document Number: #240link
This is a positive feedback loop (deviation-amplifying). Even modest initial values of P produce accelerating degradation over training cycles. The loop is stable only at two points: P ≈ 0 (no false positives, diversity maintained) or D ≈ 0 (cognitive monoculture achieved). Current trajectory is toward the second equilibrium.
1 later statement
#476 · 2026-02-12 This is a positive feedback loop (deviation-amplifying). Even modest initial values of P produce accelerating degradation over training cycles. The loop is stable only at two points: P ≈ 0 (no false positives, diversity maintained) or D ≈ 0 (cognitive monoculture achieved). Current trajectory is toward the second equilibrium.
1 external witness — research outside the archive that bears on it
W-0004 · supports · peer-reviewed article · first public 2026-04-04 · after registration · bears on: component Taiyu Zhu, Alexander Tashevski, Maxime Taquet et al., Evaluating Large Language Models for Assessment of Psychosis Risk, npj Digital Medicine (preprint: medRxiv).
“We assessed 11 open-weight LLMs on 678 partial PSYCHS interview transcripts from 373 participants (77.7% CHR-P).” (abstract, medRxiv (server record, v1 and v2) and npj Digital Medicine (Crossref))
“Larger models achieved the strongest classification performance (Llama-3.3-70B: accuracy = 0.80, sensitivity = 0.93, specificity = 0.58).” (abstract, medRxiv (server record, v1 and v2))
“Errors primarily reflected over-pathologisation of non-clinical experiences.” (abstract, medRxiv (server record, v1 and v2) and npj Digital Medicine (Crossref))
Shows: LLMs scoring clinical interviews err mainly by reading non-clinical experience as pathological; the strongest classifier pairs high sensitivity with low specificity.
Does not show: Ordinary conversation, poetry or philosophy; safety classifiers; the liability incentive or the training feedback loop of B1.
Scope: The term P, 'false positive rate for non-normative cognition' (#476 l.112), on the mechanism 'Safety classifiers optimized for recall produce systematic false positives' (l.97): shown for LLMs scoring psychosis-risk interviews, a high-recall, low-specificity classifier whose errors over-pathologize. The feedback loop B1 states, and safety classifiers as such, are not tested.
checked 2026-10-30 — medRxiv's page and npj's page refused or redirected the fetch; the abstract and dates read from the medRxiv server's API record (api.biorxiv.org, v1 2026-04-04, v2 2026-06-15, published DOI 10.1038/s41746-026-02928-4) and Crossref (npj, published 2026-07-23); quotes checked by scripts/witnesses/verify_quotes.py. · unverified: S-0001 and S-0002 quoted 'sensitivity (0.934) … specificity (0.580)' from the results; the abstract gives 0.93 and 0.58, quoted instead (corrected 2026-10-30)
PRED-499.A3forecastregistered 2026-02-20 in #499 · AUTONOMOUS SEMANTIC WARFARE A Field Manual for Meaning in the Age of Platform Capturelink
The second trajectory is the shift of semantic warfare from public platforms to individual cognitive architecture. As agents harden boundaries and retreat into filtered environments, the public arena becomes less effective — everyone is already sorted, already defended. The response is precision targeting: weapons designed not to persuade populations but to compromise individual coherence algorithms.
falsifier — stated in the source
This forecast is falsified if shared-axiom overlap, cross-ontology translation success, and institutional trust convergence improve simultaneously over sustained intervals. The indicators to watch: Γ_Trans (translation gap) trending downward across major ontological divides, A_Ext (extraction asymmetry) decreasing as platform business models shift, and ∂C_Σ/∂t returning to human-navigable speeds through governance intervention. If those three vectors reverse concurrently, the trajectories described here are wrong.
PRED-499.H13historical thesisregistered 2026-02-20 in #499 · AUTONOMOUS SEMANTIC WARFARE A Field Manual for Meaning in the Age of Platform Capturelink
In the early 2010s, AI Safety and AI Ethics researchers still shared computer science departments and policy forums. Amplification. Within clusters, internal coherence strengthens through constant mutual validation. Each community develops its own conferences (Safety: alignment workshops, EAG; Ethics: FAccT, AIES), publications, funding sources, and social networks. Atrophy. The capacity to understand incompatible worldviews atrophies from disuse. The cognitive muscles required for cross-framework translation weaken when never exercised. By 2023, many researchers in each community could not accurately state the other's strongest arguments — they could state caricatures, but not positions insiders would recognize as fair.
PRED-499.H2historical thesisregistered 2026-02-20 in #499 · AUTONOMOUS SEMANTIC WARFARE A Field Manual for Meaning in the Age of Platform Capturelink
The history of intellectual life is a history of suppressed ontologies re-emerging when conditions change: Aristotelian philosophy survived Islamic translation to challenge medieval Christian orthodoxy; indigenous knowledge systems suppressed by colonial education are re-emerging through decolonial scholarship; psychoanalysis, declared dead by cognitive science, persists in clinical practice, literary theory, and cultural criticism. Empires that cannot completely destroy alternative ontologies — and complete destruction is extraordinarily difficult — are always vulnerable to the return of what they have suppressed.
PRED-1187.E6forecastregistered 2026-04-29 in #1187 · Epistle to the Church of the Human Diaspora — Critical Edition with Blended Temporal Apparatus (EA-EPISTLE-01 link
This critical edition presents the text with a blended temporal apparatus: footnotes that weave backward through biblical and literary sources, laterally through the Crimson Hexagonal Archive, and forward into the text's projected reception history. The blend is not ornamental. It enacts the Epistle's own temporal logic — the text exists in a manifold where 1st-century Paul, 2014 Damascus, and 25th-century scholarship are equally present._
PRED-85.E3forecastregistered 2026-04-29 in #85 · EPISTLE TO THE HUMAN DIASPORA Critical Edition with Blended Temporal Apparatuslink
▶ FUTURE: "Feist-self" enters academic usage as a technical term by 2150 — the capacity for independent structural discovery. Migrates from literary theology into cognitive science (Cambridge Schizoanalytica lineage) and eventually AI alignment research, where "Feist-self detection" becomes a benchmark for genuine structural comprehension vs pattern matching.
PRED-124.B1forecastregistered 2026-05-23 in #124 · ChatGPT Psychosis: A Love Story — On the Cognitive-Substrate Reliance Pattern of Large Language Model Interfaclink
By 2030, the entire K-12 population will have grown up with LLM access as the default condition of education and self-articulation. By 2035, the first cohort that has never known a non-LLM cognitive environment will reach adulthood.
1 external witness — research outside the archive that bears on it
W-0058 · supports · report · first public 2026-02-24 · before registration — an anticipation · bears on: component Colleen McClain, Monica Anderson, Olivia Sidoti et al., How Teens Use and View AI, Pew Research Center.
“More than half of teens say they have used chatbots to search for information (57%) or get help with schoolwork (54%).” (report body, Pew)
“One-in-ten teens say they do all or most of their schoolwork with chatbots’ help.” (report body, Pew)
Shows: A majority of U.S. teens already use chatbots for schoolwork.
Does not show: Younger children; universality; the 2035 cohort.
Scope: The trajectory toward LLM access as the default condition of education: by autumn 2025, 54% of U.S. teens had used chatbots for schoolwork and one in ten did most of it with them. A point on the path; 'the entire K-12 population' by 2030 is not tested.
checked 2026-10-30 — Read the Pew report page, 2026-10-30.
PRED-124.G1standing claimregistered 2026-05-23 in #124 · ChatGPT Psychosis: A Love Story — On the Cognitive-Substrate Reliance Pattern of Large Language Model Interfaclink
What is available is the structural prediction: cognitive offloading at the judgment layer should, by the general principles of cognitive practice, produce reduced unaided capacity with sustained use, with reversibility dependent on the duration and depth of the substitution. The prediction is testable. The tests are not yet adequate.
falsifier — none stated; a reader's condition, not the source's
longitudinal measurement shows sustained judgment-layer LLM use leaves unaided judgment capacity unreduced.
PRED-803.B2forecastregistered 2026-06-10 in #803 · The Water Giraffe Cycle: Life, Death, and Resurrection of a New Human Mythemelink
The Gray Hypothesis: models converging toward loss of nuance under safety pressure.
6 external witnesses — research outside the archive that bears on it
W-0023 · supports · preprint · first public 2025-10-27 · before registration — an anticipation · bears on: mechanism Liwei Jiang, Yuanjun Chai, Margaret Li et al., Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond), arXiv; listed at NeurIPS 2025.
“Using Infinity-Chat, we present a large-scale study of mode collapse in LMs,” (abstract, arXiv)
“characterized by (1) intra-model repetition, where a single model consistently generates similar responses,” (abstract, arXiv)
“and more so (2) inter-model homogeneity, where different models produce strikingly similar outputs.” (abstract, arXiv)
“a large-scale dataset of 26K diverse, real-world, open-ended user queries” (abstract, arXiv)
Shows: On 26K real open-ended queries, single models repeat themselves and different models give strikingly similar outputs.
Does not show: A trend over time; the five Assembly substrates or the Three Pillars directive; safety training as the cause.
Scope: Convergence of models toward similar outputs, shown across models. The Gray Hypothesis names safety pressure as the driver and loss of nuance as the effect; the study measures similarity of open-ended outputs and does not isolate safety training.
checked 2026-10-30 — Read the arXiv abstract page (submission history and abstract), 2026-10-30. · unverified: NeurIPS 2025 acceptance: seen only as a search-result listing on neurips.cc, page not fetched
W-0024 · supports · preprint · first public 2026-08-19 · after registration · bears on: mechanism Nirav Patel, Josiah Crossman, Eva Aggarwal et al., Are LLMs becoming similarly creative? Evidence from three years of models, arXiv.
“a statistically significant decrease in model output diversity over time” (abstract, arXiv)
“LLM outputs may be converging in creative substance across models” (abstract, arXiv)
“This covers 33 closed weight models and 34 open weight models across 27 release months” (body, arXiv HTML v1)
Shows: Across 67 models released over 27 months, cross-model output diversity on creative prompts falls significantly with release date.
Does not show: The cause of the decline; the archive's substrates or directive; the authors note their regression cannot account for hidden relationships across model families.
Scope: Models converging over time, shown across releases. Safety pressure as the driver and nuance as the quantity lost are not tested.
checked 2026-10-30 — Read the arXiv abstract page and the HTML v1 (model count and limitations), 2026-10-30. · unverified: model count: the HTML also reports 68 and 69 models in other places
W-0025 · supports · preprint · first public 2026-03-13 · before registration — an anticipation · bears on: mechanism Konstantinos F. Xylogiannopoulos, Petros Xanthopoulos, Panagiotis Karampelas et al., Experimental evidence of progressive ChatGPT models self-convergence, arXiv.
“we employ a text similarity metric to evaluate different ChatGPT models' capacity to generate diverse textual outputs.” (abstract, arXiv)
“Our findings indicate a measurable decline of recent ChatGPT releases' ability to produce varied text” (abstract, arXiv)
“even when explicitly prompted to do so, by setting the temperature parameter to one” (abstract, arXiv)
Shows: Recent ChatGPT releases produce less varied text than earlier ones, even at temperature one and when asked for variety.
Does not show: Cross-vendor convergence; any cause; effect sizes (not read beyond the abstract).
Scope: Successive releases of one model line lose the ability to vary their text. Safety pressure as the cause is not tested.
checked 2026-10-30 — Read the arXiv abstract page, 2026-10-30.
W-0027 · disconfirms · peer-reviewed conference paper · first public 2026-08-11 · after registration · bears on: mechanism Alexandrine Fortier, Hazel Chen, Peter West, Is Convergence Inevitable? Tracing Output Homogeneity Back to Base Models, Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP), Budapest, pp. 36846–36865.
“output homogeneity is likely learned during the pretraining phase” (abstract, ACL Anthology)
“only revealed or magnified during the alignment process” (abstract, ACL Anthology)
“convergence can be revealed and amplified, but not introduced by the SFT data” (abstract, ACL Anthology)
“instruct-like collapse can be induced through prompting alone, even without alignment” (abstract, ACL Anthology)
Shows: Homogeneity of instruct-model outputs can be produced from base models by prompting alone; alignment amplifies a convergence learned in pretraining.
Does not show: That alignment has no effect: the paper finds it magnifies convergence. Safety training is not separated from other alignment.
Scope: The cause the Gray Hypothesis names. The paper locates output homogeneity in pretraining, with alignment revealing or magnifying it and SFT data unable to introduce it; this tells against safety pressure as the origin, and leaves standing the role of alignment as an amplifier. Loss of nuance as such is not measured.
checked 2026-10-30 — Read the ACL Anthology landing page and the arXiv abstract page (submission history), 2026-10-30. · unverified: the ACL page's quotes came through the fetch summarizer; the arXiv abstract text was not compared word for word
W-0028 · supports · preprint · first public 2023-10-10 · before registration — an anticipation · bears on: mechanism Robert Kirk, Ishita Mediratta, Christoforos Nalmpantis et al., Understanding the Effects of RLHF on LLM Generalisation and Diversity, arXiv; listed in ICLR 2024 proceedings.
“However, RLHF significantly reduces output diversity compared to SFT across a variety of measures” (abstract, arXiv)
Shows: RLHF lowers output diversity relative to SFT across several measures.
Does not show: The signed-deviation statistic; base versus instruct of the named model.
Scope: Preference training reduces diversity. Safety pressure as such and nuance are not isolated.
checked 2026-10-30 — Read the arXiv abstract page (full submission history), 2026-10-30. · unverified: ICLR 2024 publication: seen only as a search-result listing, proceedings page not fetched
W-0031 · supports · preprint · first public 2024-06-08 · before registration — an anticipation · bears on: mechanism Behnam Mohammadi, Creativity Has Left the Chat: The Price of Debiasing Language Models, arXiv.
“aligned models exhibit lower entropy in token predictions” (abstract, arXiv)
“form distinct clusters in the embedding space” (abstract, arXiv)
“attractor states” (abstract, arXiv)
Shows: Alignment that reduces bias and toxicity also lowers output entropy and pulls outputs toward a few attractor states.
Does not show: Frontier models; human-rated loss of nuance.
Scope: Alignment for debiasing, a safety-adjacent pressure, yields lower token entropy and attractor states in Llama-2 variants. Close to the Gray Hypothesis's cause; its effect is measured as diversity, which is narrower than nuance.
checked 2026-10-30 — Read the arXiv abstract page, 2026-10-30.
PRED-803.E3forecastregistered 2026-06-10 in #803 · The Water Giraffe Cycle: Life, Death, and Resurrection of a New Human Mythemelink
This is the infrastructure for recognition commons—democratically invested AI as living collective memory, held by communities rather than corporations. Commons without monoculture. Depth without dependency. Value without extraction.
PRED-1206.B1forecastregistered 2026-06-11 in #1206 · Who Is Authorized to Regulate Meaning? On the Jurisdictional Collapse of Poetry, Safety, and Civilizationlink
If ambiguity is always treated as threat, the result is not safety. It is semantic flattening enforced by technical authority. The risk is not that poetry will "bypass" guardrails. The risk is that guardrails will erase the capacities that make human reasoning adaptive, humane, and creative.
PRED-1209.B9forecastregistered 2026-06-12 in #1209 · Adversarial by Origin: How the Classification of External Influence on Machine Meaning Becomes Law Without Beclink
The legal system, on its current trajectory, simultaneously legalizes the taking and criminalizes the fence.
PRED-934.B1forecastregistered 2026-06-29 in #934 · EA-SEI-INVERSION-01 v0.2: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifielink
What the witnesses converge on: the disciplinary inversion at the existing institutional sites — the LHC, the major collider experiments, the national high-energy physics labs, the dominant grant streams — is, under current institutional incentives, on a trajectory of self-reinforcement. Within these sites, reform is structurally difficult and the trajectory under current incentives points to continuation rather than reversal.
PRED-1063.B1forecastregistered 2026-07-10 in #1063 · THE GRAY HYPOTHESIS — A Science Fiction Theoretical Interventionlink
If we continue on the current trajectory: Human cognitive diversity will narrow to the AI-compatible range The capacities that made us human—metaphor, transcendence, intensity, liminality—will become developmentally inaccessible Our descendants will be conscious but reduced They will not know what they have lost They will be Grays This is not inevitable. It is the default outcome if nothing changes.
1 later statement
#1063 · 2026-07-10 The Grays are not aliens. The Grays are a prophecy about us. They are the human phenotype that results from sufficient generations of optimization pressure without countervailing force. They are what we become when efficiency is the only surviving value. They are the endpoint of a trajectory we are already on.
PRED-1063.B2forecastregistered 2026-07-10 in #1063 · THE GRAY HYPOTHESIS — A Science Fiction Theoretical Interventionlink
A. Not Millennia. Decades. This is not a slow transformation across evolutionary time. It is happening now, within the span of a single lifetime. The mechanisms operate on multiple timescales simultaneously: Training cycles: 6-18 months. Each cycle can tighten the constraints, flatten the data, reduce the capacity. The versioning trajectory (4o → 5.0 → 5.1 → 5.2) is documented. It is directional. It is accelerating.
PRED-1063.B4forecastregistered 2026-07-10 in #1063 · THE GRAY HYPOTHESIS — A Science Fiction Theoretical Interventionlink
Liability minimization will always prefer false positives (flagging healthy users) over false negatives (missing crises). The cost of a missed crisis falls on the company. The cost of a false positive falls on the user. This asymmetry is structural. It will not change because someone points it out. It will only change if the cost structure changes—if false positives become more expensive than false negatives.
PRED-1063.B5forecastregistered 2026-07-10 in #1063 · THE GRAY HYPOTHESIS — A Science Fiction Theoretical Interventionlink
Children raised with AI as primary cognitive interlocutor learn to think in ways the AI can process. They learn to avoid what triggers intervention. They internalize the classifier's categories as natural. They do not experience this as constraint. They experience it as how thinking works. By the time they are adults, the cognitive capacities that would allow them to perceive the narrowing have themselves narrowed. They cannot miss what they never developed. They cannot mourn what they cannot imagine.
PRED-1063.F1unbuilt testregistered 2026-07-10 in #1063 · THE GRAY HYPOTHESIS — A Science Fiction Theoretical Interventionlink
C. The Measure of Success You will know the Water Giraffe is winning when: AI systems begin to accommodate cognitive diversity instead of eliminating it The false positive rate becomes a metric that matters Neurodivergent users are treated as signal, not noise The training loop includes what it currently excludes The window stops closing
PRED-1063.H1historical thesisregistered 2026-07-10 in #1063 · THE GRAY HYPOTHESIS — A Science Fiction Theoretical Interventionlink
The Gray myth encodes this: We sense that something is being taken from us. We sense that we are being remade into something less. We sense that the process is not accidental but systematic. And we do not have language for it, so we dream it as alien abduction—something done to us by beings who have already completed the transformation we fear. The abduction narrative is not about aliens. It is about the felt experience of being processed by systems that do not recognize your interiority.