{
  "record_id": "20802169",
  "document_id": "20802169",
  "title": "Textual Inflation and Semantic Compression in the LLM Era",
  "pages": 18,
  "authors": [
    "Raynor Eissens"
  ],
  "doi_confirmed_in_pdf": null,
  "zenodo_record": "https://zenodo.org/records/20802169",
  "html": "papers/20802169.html",
  "text": "text/20802169.txt",
  "data": "data/20802169.json",
  "abstract_extracted": "Large language models have changed the economics of textual production by making fluent, plausible prose inexpensive to generate at scale. This paper develops the concept of textual inflation: the rapid increase in cheap, fluent, AI-generated text that reduces the marginal signaling value of individual text outputs while increasing filtering, verification, attention, and trust costs. It synthesizes prior work on information overload, cognitive load, attention scarcity, the attention economy, calm technology, glanceable interfaces, AI slop, AI fatigue, generative search, model collapse, and authenticity concerns. The central research question is whether large-scale LLM-generated textual abundance creates measurable pressure toward semantic compression: lower-bandwidth meaning carriers such as summaries, icons, reactions, badges, provenance marks, confidence displays, dashboards, haptics, ambient cues, and other glanceable signals that preserve enough relevance, state, trust, or intent while reducing attention cost. The reviewed evidence does not prove a deterministic replacement of te",
  "visual_pages": [
    11
  ],
  "low_text_pages": [],
  "characters_extracted": 56298,
  "words_extracted": 7650,
  "source_pdf_filename": "20802169_Textual_Inflation_and_Semantic_Compression_in_the_LLM_Era.pdf",
  "source_pdf_sha256": "1b7f7e84d5a79e4961d8c2cef95b0454732ba2ecc3c2176441191719848333b3",
  "full_text": "=== PDF PAGE 1 ===\nTextual Inflation and Semantic Compression in the\nLLM Era:\nEvidence for Attention Scarcity under Generative\nAbundance\n\nRaynor Eissens\nIndependent Researcher\nTSX-6 Technical Report\n\n2026\n\nAbstract\n\nLarge language models have changed the economics of textual production by making fluent,\nplausible prose inexpensive to generate at scale. This paper develops the concept of textual\ninflation: the rapid increase in cheap, fluent, AI-generated text that reduces the marginal\nsignaling value of individual text outputs while increasing filtering, verification, attention, and\ntrust costs. It synthesizes prior work on information overload, cognitive load, attention scarcity,\nthe attention economy, calm technology, glanceable interfaces, AI slop, AI fatigue, generative\nsearch, model collapse, and authenticity concerns. The central research question is whether\nlarge-scale LLM-generated textual abundance creates measurable pressure toward semantic\ncompression: lower-bandwidth meaning carriers such as summaries, icons, reactions, badges,\nprovenance marks, confidence displays, dashboards, haptics, ambient cues, and other glanceable\nsignals that preserve enough relevance, state, trust, or intent while reducing attention cost. The\nreviewed evidence does not prove a deterministic replacement of text, nor does it establish color or\nchromatic systems as necessary solutions. It does, however, support a cautious structural claim:\nwhen generated text increases faster than available human attention and verification capacity,\nfiltering burden and authenticity uncertainty rise, making curation, provenance, trust signals, and\ncompressed semantic forms more valuable. TSX-6 is best understood as a conceptual synthesis\nthat connects established overload theory with LLM-era evidence of AI fatigue, synthetic content\nsaturation, lower click-through behavior, and authenticity erosion. Its contribution is a testable\nmodel linking generative abundance to semantic compression pressure.\n\nKeywords: textual inflation; semantic compression; large language models; attention scarcity; AI\nfatigue; AI slop; information overload; authenticity; human-computer interaction; generative AI;\nsemantic entropy; provenance.\n\n1\nIntroduction\n\nThe public arrival of large language models (LLMs) changed not only how text is written but\nalso the conditions under which text is evaluated. Earlier digital media environments already\nproduced information overload, feed saturation, and attention competition. LLMs intensify this\nenvironment by reducing the cost of producing coherent-looking prose. A report, email, explanation,\n\n1\n\n=== PDF PAGE 2 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\nmarketing page, comment, summary, review, lesson plan, product description, or policy memo can\nnow be generated quickly and repeatedly. The result is not simply more information. It is more\nfluent information, more plausible information, and more text that resembles expert or institutional\nlanguage even when its evidential value remains uncertain.\n\nThis paper asks whether such abundance creates a new pressure within the information environ-\nment. If fluent text was once relatively costly to produce, its presence carried at least some implicit\nsignal of effort, expertise, or institutional process. When fluent text becomes cheap, that signal\nweakens. The marginal value of an individual text output declines, while the burden of determining\nwhether the output is accurate, relevant, authentic, or worth reading increases. This mechanism is\ndescribed here as textual inflation.\n\nThe concept does not mean that text disappears or that long-form reasoning becomes obsolete.\nText remains a high-resolution medium for explanation, law, research, mathematics, history, and\nargument. The claim is narrower: under conditions of generative textual abundance, attention, trust,\ncuration, and compression become more valuable. Users and systems increasingly need mechanisms\nthat reduce cognitive load while preserving enough meaning to act. These mechanisms can include\nsummaries, visual status indicators, badges, reaction systems, dashboards, confidence bars, haptics,\nambient cues, and other lower-friction carriers of meaning.\n\nThe central research question is therefore: does large-scale LLM-generated textual abundance\ncreate measurable selection pressure toward semantic compression? The evidence reviewed in this\npaper suggests that several adjacent phenomena are converging. Foundational information-overload\ntheory identifies attention as the scarce resource consumed by information [27]. Cognitive load\ntheory explains why excessive or poorly structured information can overwhelm limited working\nmemory [29]. The attention economy treats attention as a scarce social and economic resource\n[8, 12]. HCI research on calm and glanceable interfaces shows a long-standing design effort to move\ninformation from the center of attention into peripheral, low-friction channels [21, 32]. Recent\ndiscourse and studies on AI slop, workslop, and AI fatigue document cultural and workplace unease\naround the flood of generic AI-generated output [15, 17, 18]. Search-behavior evidence suggests\nthat AI summaries alter the value chain of original text by reducing click-through behavior [1, 20].\nResearch on model collapse adds a related systemic concern: synthetic content can feed back into\ntraining systems and degrade distributions when not balanced by real data [24, 26].\n\nThe contribution of this paper is synthetic and conceptual.\nIt does not claim that TSX-6\ndiscovered attention scarcity, information overload, cognitive load, or calm technology. Nor does it\nclaim that chromatic or color-based systems are proven solutions. Instead, it links older theories\nof overload and attention with recent LLM-era evidence into a single structural model: generative\nabundance produces textual inflation; textual inflation increases semantic entropy and cognitive\nresidue; these pressures intensify attention and authenticity scarcity; and these scarcities create\npressure toward semantic compression.\n\nThe paper contributes five elements. First, it provides definitions of textual inflation, semantic\nentropy, cognitive residue, semantic compression, and compression pressure. Second, it describes a\nmethodology for a structured evidence review. Third, it compares TSX-6 to existing theories of\noverload, attention, cognitive load, calm technology, AI slop, and model collapse. Fourth, it proposes\nexplicit hypotheses that can be tested in future empirical work. Fifth, it presents a structural model\nand novelty statement suitable for further refinement, critique, and experimental validation.\n\n2\n\n=== PDF PAGE 3 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\n2\nMethodology\n\nThis paper is a structured conceptual evidence review rather than a controlled empirical study.\nIt synthesizes foundational theory, peer-reviewed literature, academic preprints, research-institute\nfindings, industry reports, technology journalism, and public discourse in order to define and evaluate\na proposed conceptual chain. The goal is not to estimate effect sizes or establish causality. The goal\nis to determine whether existing literature already formulates the full TSX-6 chain and whether the\navailable evidence is consistent with the proposed model.\n\nThe review followed four steps. First, foundational literature was identified for established\ntheories that explain overload and attention scarcity: information overload, cognitive load theory,\nattention economy, and calm technology. Second, LLM-era sources from 2023–2026 were reviewed\nfor evidence of synthetic content saturation, AI fatigue, workplace oversight burden, AI-generated\nsearch summaries, model collapse, and authenticity concerns. Third, sources were classified by\nevidential strength. Peer-reviewed literature and research-institute reports were treated as stronger\nthan journalism, blogs, LinkedIn posts, and forum discourse. Fourth, the reviewed sources were\ncompared against the complete TSX-6 chain to assess novelty and prior-art overlap.\n\nThe evidence hierarchy used in this paper is as follows: (1) peer-reviewed academic literature;\n(2) academic books and conference papers; (3) research institutes and public-interest empirical\nstudies; (4) academic preprints and working papers; (5) industry reports and technical analyses;\n(6) major journalism and trade journalism; and (7) blogs, social media, and anecdotal discourse.\nSources in lower categories are not excluded, because they may document emerging language and\ncultural symptoms before peer-reviewed work appears. However, they are not treated as strong\ncausal evidence.\n\nThis methodology has limits. The search was not a fully systematic review using preregistered\ndatabase queries, inclusion criteria, and inter-rater coding. It is therefore inappropriate to claim\nabsolute historical priority. The first-mover assessment in this paper is phrased cautiously: to the\nbest of the author’s knowledge and within the reviewed corpus, the complete chain appears not to\nhave been formulated as a single model before TSX-6. This is a novelty claim about synthesis and\nstructure, not a claim that every component is new.\n\n3\nDefinitions\n\n3.1\nTextual Inflation\n\nTextual inflation is the rapid increase in cheap, fluent, AI-generated text that reduces the marginal\nsignaling value of individual text outputs while increasing filtering, verification, attention, and trust\ncosts. The term is intentionally economic in structure but semiotic in application. It describes a\ncondition in which the supply of fluent symbolic output expands faster than the available human\ncapacity to evaluate it. Under textual inflation, the problem is not only that there is more text.\nThe problem is that more text now appears polished, plausible, and contextually fluent, even when\nits evidential status is weak.\n\nTextual inflation differs from ordinary information overload.\nTraditional overload may be\nproduced by email volume, database growth, news abundance, or feed saturation. Textual inflation\nis specifically tied to the generative capacity of LLMs and related systems. It refers to a decrease in\nthe signaling value of fluent prose itself. When fluent prose becomes cheap, readers must devote\nmore attention to judging relevance, origin, accuracy, and authenticity.\n\n3\n\n=== PDF PAGE 4 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\n3.2\nSemantic Entropy\n\nSemantic entropy is the degradation of distinctiveness, trust, nuance, or interpretive stability under\nconditions of excessive symbolic output. It does not mean that all generated text is meaningless.\nRather, it describes the increased difficulty of distinguishing meaningful, grounded, authored, and\ntrustworthy outputs from generic, redundant, or low-accountability outputs. In LLM environments,\nsemantic entropy may appear as sameness of tone, flattening of style, formulaic structure, repeated\nrhetorical patterns, or uncertainty about whether a text reflects human judgment.\n\n3.3\nCognitive Residue\n\nCognitive residue is the leftover cognitive burden created by verification, filtering, rewriting, tool\nswitching, and uncertainty management. A user may receive an apparently useful AI-generated\nanswer but still need to inspect its accuracy, compare it with sources, remove generic phrasing,\nadapt it to context, verify factual claims, and decide whether it can be trusted. This residual burden\nremains after the apparent labor-saving benefit of generation. In workplace contexts, this may\nappear as the management of polished but low-substance output. In search and publishing contexts,\nit may appear as uncertainty over whether a summary, article, comment, or review is grounded.\n\n3.4\nSemantic Compression\n\nSemantic compression is the use of lower-bandwidth meaning carriers that preserve relevance, state,\ntrust, or intent while reducing attention cost. It does not necessarily mean reducing meaning. It\nmeans reducing the attention cost per unit of usable meaning. Examples include summaries, emoji,\nreaction buttons, icons, badges, status indicators, confidence bars, visual dashboards, traffic-light\nsystems, haptics, ambient displays, glanceable UX, short human-curated signals, provenance markers,\nand compact state indicators.\n\nSemantic compression is neutral with respect to medium. A compressed semantic carrier can be\ntextual, visual, auditory, tactile, spatial, or procedural. A one-line summary, a warning badge, a\ncheckmark, a vibration pattern, a color state, and a confidence indicator can all function as semantic\ncompression if they reduce cognitive load while preserving actionable meaning. Chromatic systems,\nwhere color is used as a state carrier, are one possible subtype, not the necessary endpoint.\n\n3.5\nCompression Pressure\n\nCompression pressure is the tendency for users, platforms, and interfaces to prefer lower-friction\nmeaning carriers when text becomes too abundant. It is not a law of history or a proof of inevitability.\nIt is a hypothesized selection pressure: when the volume of text increases and attention remains\nlimited, systems that help users act with less cognitive overhead become more valuable.\n\n4\nResearch Hypotheses\n\nTSX-6 is presented as a conceptual model, but it can be translated into testable hypotheses. The\nfollowing hypotheses are proposed for future empirical work:\n\nH1. Increased exposure to AI-generated text is associated with increased perceived cognitive residue.\nThis hypothesis predicts that users exposed to higher volumes of LLM-generated prose will report\n\n4\n\n=== PDF PAGE 5 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\nmore verification burden, source-checking effort, rewriting effort, uncertainty management, or fatigue\nthan users exposed to lower volumes or clearly sourced human-authored text.\n\nH2. Increased cognitive residue is associated with preference for compressed semantic carriers. This\nhypothesis predicts that when users experience higher filtering and verification burden, they will\nprefer summaries, badges, dashboards, trust marks, confidence indicators, visual states, or other\ncompressed meaning carriers over long textual outputs for routine orientation and decision support.\n\nH3. Increased textual abundance is associated with increased demand for trust and provenance\nsignals. This hypothesis predicts that as users perceive more generic or AI-generated text in\ntheir environment, they will place greater value on author identity, source links, citations, human-\nauthorship labels, editorial reputation, provenance metadata, and verification indicators.\n\nThese hypotheses are intentionally correlational in their first form. Future studies could test\ncausal versions through controlled exposure experiments, interface comparisons, longitudinal diary\nstudies, and field experiments in workplace, education, search, and platform contexts.\n\n5\nRelation to Existing Theories\n\n5.1\nInformation Overload Theory\n\nThe intellectual foundation of TSX-6 begins with information overload. Herbert Simon’s observation\nthat a wealth of information creates a poverty of attention remains central [27]. Information overload\nliterature later developed extensive accounts of the causes, symptoms, and coping strategies of\nexcessive information volume [3, 10]. These works explain why users and organizations require\nfiltering, summarization, selection, and information management.\n\nTSX-6 accepts this foundation but adds an LLM-specific condition. Information overload theory\ngenerally treats the problem as excess information volume. Textual inflation treats the problem\nas excess fluent symbolic production whose surface quality is no longer a reliable proxy for effort,\nexpertise, or trustworthiness. The difference matters because LLM-generated text can appear\nstructured, professional, and plausible even when it is redundant, weakly grounded, or synthetic.\n\n5.2\nCognitive Load Theory\n\nCognitive load theory explains why limited working-memory resources can be overwhelmed by\nexcessive, extraneous, or poorly structured information [29]. In TSX-6, cognitive residue is an\nLLM-era application of this principle. The burden does not necessarily occur while reading alone.\nIt may occur after generation, when users must verify, adapt, repair, compare, and decide whether\nan AI output can be trusted. The location of labor shifts from production to evaluation.\n\n5.3\nAttention Economy\n\nThe attention economy treats attention as a scarce resource in environments of information abundance\n[8, 12]. TSX-6 extends this logic to generated text. If attention is scarce and LLMs radically expand\nthe supply of text seeking attention, then text competes more intensely for human evaluation.\nUnder this condition, the value of curation, source reputation, provenance, and compressed signals\nincreases.\n\n5\n\n=== PDF PAGE 6 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\n5.4\nCalm Technology and Glanceable Interaction\n\nCalm technology and ubiquitous computing provide a second major predecessor. Weiser and Brown\nargued that technology should move between the periphery and the center of attention, informing\nwithout constantly demanding focus [32]. Ambient-display and glanceable-interface research later\ndeveloped related ideas: information can be conveyed through small cues, peripheral signals, visual\nstates, badges, or ambient changes rather than through full textual explanation [21].\n\nThis prior art supports the semantic-compression side of TSX-6, but it does not arise from LLM-\ngenerated text abundance. Calm technology responded to earlier computational and informational\noverload. TSX-6 argues that LLM-era textual inflation creates a renewed and intensified reason to\nrevisit calm, glanceable, peripheral, and low-bandwidth interface forms.\n\n5.5\nAI Slop and Synthetic Content Saturation\n\nThe recent discourse around AI slop names a cultural symptom of textual inflation. AI slop generally\nrefers to low-quality, mass-produced, generic, synthetic content that floods platforms and feeds [15].\nA 2025 academic study of AI-generated algorithmic virality examined synthetic content in TikTok\nand Instagram search results across several European countries and described the emergence of\naccounts producing generative AI content at scale [28]. Although detection methods, definitions, and\nplatform samples vary, the phenomenon is relevant because it indicates that synthetic abundance is\nincreasingly experienced as pollution, sameness, or trust degradation.\n\n5.6\nModel Collapse\n\nModel collapse is not the same as textual inflation, but it is an adjacent theoretical warning.\nShumailov and colleagues argue that recursively training generative models on generated data can\nmake models forget tails of the original distribution [26]. Subsequent work has analyzed conditions\nunder which synthetic data may or may not produce degradation [11, 24]. For TSX-6, model collapse\nmatters because it gives a model-level analogue to semantic entropy: if synthetic outputs crowd\nthe informational environment, both human readers and future models may face greater difficulty\npreserving diverse, grounded, human-originated signal.\n\n6\nPrior-Art Gap and Novelty Statement\n\nThe novelty of TSX-6 is not that it identifies information overload, attention scarcity, cognitive load,\ncalm technology, AI slop, or authenticity concerns. Each of these exists in prior work. The novelty\nis the integration of these fields into a single LLM-era model in which cheap fluent text reduces the\nmarginal signaling value of prose and increases the value of lower-friction meaning carriers.\n\nThe prior-art gap can be stated narrowly. Existing work separately addresses: (1) information\noverload and filtering; (2) attention scarcity and attention economics; (3) cognitive load and working-\nmemory limits; (4) calm, ambient, and glanceable interface design; (5) AI slop and synthetic content\nsaturation; (6) model collapse and synthetic data feedback; and (7) authenticity concerns around\nAI-generated work. In the reviewed corpus, however, these literatures do not appear to formulate\nthe complete chain as a single structural model:\n\nLLM text abundance →textual inflation →semantic entropy / cognitive residue →\n\n6\n\n=== PDF PAGE 7 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\nTable 1: Relationship between TSX-6 and existing theories.\n\nTheory or discourse\nWhat it explains\nLLM-\nspecific?\nWhat TSX-6 adds\n\nNo\nDefines textual inflation as a\ngenerative-text-specific overload\ncondition.\n\nInformation\noverload\nExcessive information can\noverwhelm limited processing\ncapacity and reduce decision\nquality.\n\nAttention economy\nAttention is scarce and becomes\na central economic resource\nunder abundance.\n\nNo\nLinks AI verification, rewriting,\nand uncertainty management to\ncognitive residue.\n\nNo\nConnects cheap generated text\nto declining marginal signaling\nvalue of prose and rising\ncuration value.\nCognitive load\ntheory\nWorking memory is limited;\npoorly structured or excessive\ninformation creates cognitive\nburden.\n\nNo\nTreats calm and glanceable\nforms as semantic compression\nunder LLM-era pressure.\n\nCalm technology\nInterfaces can inform through\nperipheral, low-friction cues\nrather than constant central\nattention.\n\nAI slop discourse\nSynthetic content saturation\nproduces perceived banality,\nsameness, clutter, and trust\ndecline.\n\nYes\nPlaces AI slop inside a broader\nstructural model of textual\ninflation and compression\npressure.\nModel collapse\nRecursive synthetic training can\ndegrade model distributions\nunder some conditions.\n\nYes\nIntegrates older overload\ntheories and recent\ngenerative-AI evidence into one\ntestable chain.\n\nYes\nProvides an adjacent\nsystem-level analogue for\nsemantic entropy and\nhuman-signal scarcity.\nTSX-6\nLinks LLM text abundance,\ntextual inflation, semantic\nentropy, cognitive residue,\nattention scarcity, authenticity\nscarcity, and semantic\ncompression pressure.\n\nattention and authenticity scarcity →semantic compression pressure.\n\nThe strongest novelty claim is therefore synthetic and conceptual: TSX-6 names and connects a\nset of pressures that are visible across multiple fields but usually treated separately. It proposes that\nLLM-era textual abundance may accelerate a pre-existing migration toward compression, curation,\nprovenance, and glanceable interface forms. This claim should be tested empirically and should not\nbe treated as a completed proof.\n\n7\nEvidence Review: Textual Inflation in 2023–2026\n\n7.1\nAI Fatigue and AI Brain Fry\n\nAI fatigue refers to tiredness, disengagement, or overload associated with repeated exposure to AI\nsystems, AI-generated content, or AI-mediated work. The term overlaps with technostress, digital\nfatigue, cognitive overload, and workplace burnout, but it points to a more specific LLM-era pattern:\n\n7\n\n=== PDF PAGE 8 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\nusers are not merely using technology; they are repeatedly supervising, prompting, evaluating,\ncorrecting, and integrating generated output.\n\nOlder technostress research provides a strong theoretical base. Technostress studies identify\noverload, invasion, complexity, uncertainty, and insecurity as conditions under which technology use\nbecomes psychologically costly [2, 23, 30]. AI fatigue can be read as a new subtype of technostress\nin which the stressor is not simply device use or connectivity but the continuous need to interact\nwith intelligent-seeming outputs.\n\nRecent AI-specific evidence remains emerging. Ragolane and Patel’s narrative review frames AI\nfatigue as related to overexposure to intelligent systems, cognitive strain, emotional exhaustion,\nsaturation, decision paralysis, and erosion of agency [22]. Miranda, Parreno, and Rivera propose\nan academic AI-fatigue model based on 1,054 university students, identifying dimensions such as\ncognitive overload, motivational disengagement, moral unease, physical strain, and attentional drift\n[17]. These sources do not prove the TSX-6 model, but they support the plausibility of H1: sustained\nAI exposure can be associated with perceived cognitive residue.\n\nWorkplace reports also point in this direction.\nHBR-linked work on “workslop” describes\nAI-generated content that looks polished but lacks substance, shifting burden to recipients [18].\nReporting on “AI brain fry” describes mental fatigue among workers supervising multiple AI tools or\niterating too extensively on AI outputs [35]. These are not substitutes for peer-reviewed longitudinal\nstudies, but they are important symptoms of the evaluation burden that TSX-6 calls cognitive\nresidue.\n\n7.2\nWorkslop and the Productivity Paradox\n\nThe productivity paradox in generative AI is that automation of production can increase the\nburden of evaluation. A user may save time generating a memo, but another user may spend\ntime determining whether that memo contains real analysis. A manager may receive polished AI-\ngenerated output that appears complete but requires rework. A developer may use a coding assistant\nthat accelerates generation while increasing review burden. The output is fast; the judgment is not.\n\nThis is the core mechanism of cognitive residue. It helps reconcile two apparently conflicting\nfindings. Generative AI can improve productivity in some controlled tasks and organizational\nsettings [5, 19]. At the same time, poorly governed or excessive use can create low-substance output,\nverification burden, and coordination costs [18]. TSX-6 does not claim that AI text is intrinsically\nharmful. It claims that when generated text proliferates without adequate trust, provenance, and\ncompression systems, the evaluation burden can rise.\n\n7.3\nAI Slop as Cultural Symptom\n\nAI slop is important for TSX-6 because it is a cultural name for the devaluation of generic generated\noutput. Knibbs reported that AI-generated content was appearing at scale on Medium and described\nmuch of it as banal [15]. Stanusch and colleagues’ study of AI-generated algorithmic virality provides\na more systematic academic account of synthetic content circulating through social-media search\nresults and agentic AI accounts [28]. These sources differ in method and evidential strength, but\nthey converge on a common observation: generative tools lower the cost of producing platform\ncontent, and some actors use that low cost to pursue visibility.\n\nThe TSX-6 interpretation is structural rather than moral. Not all AI-generated content is slop.\nSome generated content is useful, clear, accessible, and well supervised. The relevant point is\n\n8\n\n=== PDF PAGE 9 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\nthat when low-cost generation becomes available to many actors, users must spend more effort\ndistinguishing signal from generic or low-accountability output. The term “slop” is culturally\nimprecise, but it indexes a real attention problem: cheap content can increase the cost of finding\ntrustworthy content.\n\n7.4\nAuthenticity Erosion and Human-Signal Scarcity\n\nAuthenticity concerns form another evidence stream. Empirical work on perceptions of AI-generated\ntext suggests that labels alone can shape evaluation. Zhu and colleagues found that raters favored\ncontent labeled as human-generated over content labeled as AI-generated, even when labels were\nswapped [34]. Kim and colleagues report an AI penalization effect in which people reduce compen-\nsation for workers who use AI, partly because they assign less credit to AI-assisted work [14]. These\nstudies do not show that AI content is worse. They show that origin, authorship, and perceived\nhuman contribution affect value judgments.\n\nMarketing and platform discourse points in the same direction. Digiday reporting describes\nrenewed demand for authenticity and “messiness” after oversaturation of AI-generated content\n[16]. Public surveys and industry reports increasingly emphasize distrust of unclear AI-generated\ninformation and demand for human tone or attribution [33]. These sources should be interpreted\ncautiously, but they support H3: increased perceived textual abundance is associated with greater\ndemand for provenance, source clarity, and human-authorship signals.\n\n7.5\nSearch, Summaries, and Click Behavior\n\nSearch behavior provides one of the stronger empirical anchors. Pew Research Center found that\nGoogle users who encountered an AI summary clicked traditional search-result links less often than\nusers who did not encounter such a summary; Pew also reported higher rates of session-ending\nbehavior on search pages with AI summaries [20]. Ahrefs analysis associates AI Overviews with\nlower click-through rates for top-ranking pages [1]. A 2026 empirical study comparing Google\nSearch, AI Overviews, and Gemini found that AI Overviews were displayed for a substantial share of\nrepresentative queries and that generative search could retrieve sources differently from traditional\nsearch [13].\n\nThese findings do not prove textual inflation in full. They do show that generated summaries\ncan alter attention allocation and the value chain of text. Users may accept compressed generated\nsummaries rather than engaging source documents. This supports the claim that under abundance,\nthe value of curation, trust, and summary interfaces increases. It also raises risks: compression\nwithout provenance can reduce source visibility and intensify authenticity concerns.\n\n7.6\nEvidence Classification\n\nThe reviewed evidence should be classified rather than flattened. It includes strong theoretical\nfoundations, emerging empirical signals, and weaker but culturally relevant discourse. Table 2\ndistinguishes evidence categories.\n\n9\n\n=== PDF PAGE 10 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\nTable 2: Evidence categories used in the TSX-6 review.\n\nCategory\nExamples\nStatus in TSX-6\n\nEstablished theory\nSimon; cognitive load theory; attention\neconomy; information overload; calm technology.\nStrong conceptual\nfoundation.\nPeer-reviewed or\nacademic empirical work\nNoy and Zhang; Brynjolfsson et al.; Shumailov\net al.; technostress studies; AI perception\nstudies.\n\nStrong to moderate,\ndepending on topic and\ngeneralizability.\nResearch-institute or\ntechnical data\nPew click behavior; Ahrefs click-through\nanalysis; generative search benchmark studies.\nStrong for observed\nbehavior; cautious for\ncausal interpretation.\nAcademic preprints\nAI fatigue models; AI slop/virality studies;\ngenerative search disruption studies.\nUseful emerging evidence;\nnot all peer reviewed.\nIndustry and workplace\nreports\nWorkslop, brain fry, adoption/fatigue surveys.\nSymptom mapping; not\ncausal proof.\nJournalism and public\ndiscourse\nWired, Digiday, platform discourse, marketing\ndiscourse.\nUseful for cultural\nsymptoms and terminology;\nweaker evidence.\n\n8\nThe TSX-6 Structural Model\n\nThe model begins with generative abundance: the capacity to produce large volumes of fluent text\nat low marginal cost. This abundance produces textual inflation when the supply of plausible prose\nexpands faster than human attention, trust infrastructure, and verification capacity. Textual inflation\nthen contributes to semantic entropy: sameness, provenance uncertainty, loss of distinctiveness, and\nlower interpretive stability.\n\nSemantic entropy creates cognitive residue because users must verify, filter, compare, rewrite,\nand manage uncertainty. Cognitive residue intensifies attention scarcity by consuming the limited\nattention that information requires. Attention scarcity then interacts with authenticity scarcity:\nwhen synthetic text is abundant and generic, accountable human judgment becomes more valuable.\nUnder these combined conditions, users, platforms, and interfaces experience semantic compression\npressure: pressure to provide meaning in forms that are faster to inspect, easier to trust, and less\ncostly to evaluate.\n\n8.1\nConceptual Formalization\n\nLet T represent perceived textual abundance, A available user attention, F filtering and verification\nburden, R perceived cognitive residue, P demand for provenance and trust signals, and C demand\nfor semantic compression. TSX-6 proposes the following conceptual relationships:\n\nF = f(T/A, Q, V ),\n(1)\n\nR = g(F, S, U),\n(2)\n\nP = h(T, U, O),\n(3)\n\nC = k(R, P, A−1).\n(4)\n\nHere Q represents perceived quality variability, V represents verification difficulty, S represents\nsameness or semantic entropy, U represents uncertainty about origin or trustworthiness, and O\nrepresents opacity of authorship or provenance. The notation is not a proven mathematical law.\n\n10\n\n=== PDF PAGE 11 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\nLow marginal cost of fluent LLM text\n\nGenerative\nAbundance\n\nMore plausible prose than attention can evaluate\n\nTextual\nInflation\n\nSameness and provenance uncertainty reduce distinctiveness\n\nSemantic\nEntropy\n\nVerification, filtering, rewriting, and source checking\n\nCognitive\nResidue\n\nEvaluation burden consumes limited human attention\n\nAttention\nScarcity\n\nHuman judgment and provenance become higher-value signals\n\nAuthenticity\nScarcity\n\nDemand for summaries, badges, dashboards, and cues\n\nSemantic Compression\nPressure\n\nEvidence streams: information overload | attention economy | calm technology | AI fatigue | AI slop | search click behavior | authenticity signals\n\nFigure 1: The TSX-6 structural model. The figure presents the proposed conceptual chain from\nLLM-era generative abundance to semantic compression pressure. The arrows indicate hypothesized\ndirectional relationships, not experimentally proven causal laws. The right-hand boxes describe the\nmechanism at each stage, while the lower band lists evidence streams that currently support or\nmotivate the model.\n\nIt is a conceptual map that clarifies the model’s claim: when textual abundance rises faster than\navailable attention and verification infrastructure, filtering burden tends to increase; when filtering\nburden and uncertainty increase, cognitive residue and trust demand tend to rise; when residue\nand trust demand rise under limited attention, demand for compressed semantic carriers tends to\nincrease.\n\nThis formalization is useful because it separates measurable variables. Future research could\noperationalize T as exposure to AI-generated text per day, F as verification time, R as self-reported\ncognitive residue, P as preference for source and provenance markers, and C as revealed preference\nfor compressed interface elements.\n\n9\nSemantic Compression as Adaptive Response\n\nSemantic compression is an adaptive response to overload because it reduces attention cost with-\nout necessarily reducing usable meaning. A badge can communicate verified status faster than a\nparagraph. A confidence indicator can communicate uncertainty faster than a long disclaimer. A\ntraffic-light interface can communicate urgency faster than a policy memo. A dashboard can commu-\nnicate system state faster than a report. A reaction button can communicate social acknowledgement\nfaster than a written reply.\n\nThese forms are not replacements for reasoning. They are routing mechanisms for attention.\nThey help users decide where full attention is needed and where peripheral awareness is sufficient.\n\n11\n\n=== PDF PAGE 12 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\nIn this sense, semantic compression inherits the logic of calm technology: not everything should\noccupy the center of attention at all times.\n\nExamples include emoji and reactions for social acknowledgement; badges and indicators for\ntrust, status, or provenance; dashboards and progress rings for state; traffic-light systems for\nurgency; summaries for condensation; confidence displays for uncertainty; haptic cues for silent\nnotification; ambient displays for peripheral awareness; and possible chromatic systems for compact\nstate representation. The TSX-6 claim does not privilege any single implementation. Color may be\nuseful in some designs, but the general category is broader than color.\n\n9.1\nCompression Is Not Anti-Text\n\nSemantic compression should not be misread as a rejection of language. Compression supports\nlanguage by reducing the number of moments in which full prose must be inspected. A good\ncompressed carrier can route the reader toward the right text, preserve attention for high-stakes\nreasoning, and make long-form documents easier to navigate. In this sense, semantic compression is\nnot the enemy of textual depth. It is an attention-preserving layer around textual depth.\n\n9.2\nCompression Without Provenance Is Risky\n\nGenerated summaries can also create new problems. If a summary replaces source engagement\nwithout preserving provenance, it may reduce accountability and weaken the incentive to produce\ndurable original work. If a badge or score is opaque, it can become a false signal. If a dashboard\ncompresses too aggressively, it can hide ambiguity. TSX-6 therefore treats semantic compression\nand provenance as linked: the most useful compressed carriers are not merely shorter; they preserve\nenough origin, confidence, and context to support responsible action.\n\n10\nCounterarguments and Boundary Conditions\n\n10.1\nAI Can Improve Text Quality\n\nA major counterargument is that AI can improve average text quality. Evidence supports this\nin some settings. Noy and Zhang found productivity gains in writing tasks using generative AI\n[19]. Brynjolfsson, Li, and Raymond found that generative AI assistance increased productivity in\ncustomer support, especially for less experienced workers [5]. These findings matter. They show\nthat textual abundance is not automatically harmful.\n\nTSX-6 is compatible with these findings. The claim is not that AI text is bad. The claim is that\na lower cost of producing fluent text changes evaluation conditions. High-quality AI-assisted writing\nmay coexist with textual inflation. Indeed, if AI improves some outputs while flooding environments\nwith many more outputs, the need for trust, curation, and compression may increase rather than\ndecrease.\n\n10.2\nSummaries Can Increase Access\n\nA second counterargument is that AI summaries can make information more accessible. This is\ntrue. Summaries can help users with limited time, low literacy, disability, language barriers, or\ncomplex search tasks. They can reduce friction and improve navigation. The TSX-6 concern is\n\n12\n\n=== PDF PAGE 13 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\nnot summary itself. The concern is summary without provenance, source visibility, or adequate\nuncertainty signaling.\n\nThe Pew and Ahrefs findings should therefore be interpreted carefully. Lower click-through is not\nautomatically bad from the user’s perspective. It may mean users found what they needed faster.\nBut from a publishing and knowledge-ecosystem perspective, lower click-through may also reduce\nsource visibility, revenue, and accountability. TSX-6 treats this as a trade-off, not a simple harm.\n\n10.3\nCompression Predates LLMs\n\nA third counterargument is that compression trends long predate LLMs. Emoji, icons, dashboards,\nbadges, traffic-light systems, haptics, and ambient displays existed before ChatGPT. This is correct.\nTSX-6 does not claim that LLMs invented semantic compression. It argues that LLM-generated\nabundance may accelerate pre-existing compression dynamics by increasing the filtering burden\naround text.\n\n10.4\nNot All AI Content Causes Inflation\n\nNot all AI-generated content contributes equally to textual inflation.\nCarefully edited, cited,\naccountable, and context-specific AI-assisted writing may be valuable. The inflationary risk is\nhighest when generation is cheap, high-volume, low-accountability, weakly sourced, stylistically\ngeneric, and distributed into already overloaded environments.\n\n11\nDiscussion\n\n11.1\nHuman-Computer Interaction\n\nFor HCI, TSX-6 suggests that text-heavy interaction may not remain the default optimal form for\nevery AI-mediated task. As LLMs produce more text, the interface problem shifts from generating\nlanguage to managing attention.\nGood interfaces may increasingly need to summarize, rank,\ncompress, indicate confidence, display provenance, and support peripheral awareness.\n\n11.2\nAI Assistants and Agents\n\nAI assistants and agents should not be evaluated only by how much text they can produce. In\nan environment of textual inflation, better support may mean producing less text, offering clearer\nstatus, showing uncertainty, preserving provenance, and compressing state. Agentic systems may\nneed dashboards, state indicators, action logs, reversible summaries, and trust cues rather than long\nprose after every action.\n\n11.3\nSearch and Publishing\n\nAI summaries may reduce direct engagement with source text, as suggested by Pew, Ahrefs,\nand generative-search research [1, 13, 20]. This creates a tension. Summaries reduce user effort,\nbut they may also reduce traffic to original sources and weaken incentives for durable human-\nauthored publication. As generated summaries become common, curation, reputation, citation, and\nprovenance may become more valuable.\n\n13\n\n=== PDF PAGE 14 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\n11.4\nSocial Platforms\n\nOn social platforms, AI slop and generic content may create pressure toward authenticity markers,\nverified authorship, human-made labels, provenance systems, or design forms that reduce the need\nto inspect every post in depth. The risk is that authenticity itself becomes performative or gamed.\nCompression systems therefore need governance, transparency, and reversibility.\n\n11.5\nAmbient and Runtime Interfaces\n\nFuture ambient and runtime interfaces may use situational, low-friction forms instead of fixed app\nscreens or long textual outputs. This may include state displays, contextual summaries, symbolic\ncues, voice confirmations, haptics, or environmental signals. TSX-6 does not prove that such\ninterfaces will dominate. It suggests that the pressure toward them increases when generated text\nbecomes abundant and attention remains limited.\n\n12\nFirst-Mover Assessment\n\nTo the best of the author’s knowledge, and within the reviewed literature and public discourse,\nTSX-6 appears to be best classified as a substantially novel conceptual synthesis rather than a\ncompletely unprecedented idea in every component. The model draws on well-established prior work.\nInformation overload, cognitive load, attention scarcity, calm technology, and attention economy are\nnot new. AI slop, workslop, model collapse, generative search, and AI-authenticity concerns are also\nactive emerging areas.\n\nWhat appears distinctive is the formulation of the complete chain: LLM text abundance leads\nto textual inflation; textual inflation produces semantic entropy and cognitive residue; cognitive\nresidue intensifies attention scarcity and authenticity scarcity; and these pressures increase demand\nfor semantic compression. The claim should be phrased cautiously: no reviewed source was found\nthat explicitly formulates this complete chain as a single structural model. This does not prove that\nno such formulation exists elsewhere. It supports a defensible novelty statement: TSX-6 contributes\na named, integrated, testable framework for a set of LLM-era pressures that existing literatures\ndiscuss separately.\n\n13\nLimitations\n\nThis paper has several limitations. First, it is a conceptual synthesis, not a controlled empirical\nexperiment. It proposes hypotheses and a structural model but does not test them directly. Second,\nsome evidence comes from journalism, industry reports, blogs, LinkedIn posts, and public discourse.\nThese sources are useful for mapping symptoms and language, but they are weaker than peer-\nreviewed empirical studies. Third, causality is not fully established. AI fatigue, search behavior,\nauthenticity concerns, and productivity problems are multi-causal.\n\nFourth, the evidence does not prove that generated text universally reduces value. High-quality\nAI-supported writing may increase clarity, access, and productivity in many contexts.\nFifth,\nsemantic compression may take many forms, and no single form is identified as necessary or\ninevitable. Sixth, chromatic or color-based systems are not proven by this paper. They remain one\npossible implementation layer within a broader compression category. Seventh, more longitudinal,\n\n14\n\n=== PDF PAGE 15 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\nexperimental, and cross-cultural research is needed to measure textual inflation, semantic entropy,\nfiltering burden, and compression efficiency.\n\n14\nFuture Experimental Design\n\nFuture research should operationalize textual inflation by measuring the rate of generated-text\nexposure, perceived redundancy, verification time, trust judgments, and source-engagement patterns.\nLongitudinal studies could examine whether sustained exposure to AI-generated text increases\ncognitive fatigue or shifts user preference toward compressed signals. Experiments could compare\ntext-heavy AI interfaces with compressed state-based interfaces across tasks such as email triage,\nsearch, workplace reporting, education, and agent monitoring.\n\nA minimal experimental design could expose participants to three conditions: human-authored\nsource text, AI-generated long-form text, and AI-generated text accompanied by semantic-\ncompression supports such as provenance badges, confidence indicators, and state summaries.\nDependent variables could include task completion time, perceived cognitive residue, trust cali-\nbration, source-checking behavior, recall accuracy, and preference for compressed carriers. Such a\ndesign would test H1 and H2 directly.\n\nA longitudinal field study could ask knowledge workers to record daily exposure to AI-generated\ntext, number of AI tools used, verification time, perceived fatigue, and preference for summaries\nor dashboards. This would help separate ordinary workload from AI-specific cognitive residue. A\nsearch study could compare AI summary interfaces with and without source-provenance cues to test\nwhether compression plus provenance preserves user efficiency while reducing source invisibility.\n\nAdditional work should develop metrics for semantic entropy and compression efficiency. Semantic\nentropy might be measured through perceived sameness, loss of source distinctiveness, reduced\nauthor recognition, uncertainty about origin, or increased verification time. Compression efficiency\nmight be measured as usable meaning retained per unit of attention cost. HCI prototypes could\ntest dashboards, confidence indicators, provenance badges, haptic cues, and ambient states against\nlong textual outputs.\n\nAcknowledgements\n\nThe author acknowledges the broader bodies of work on information overload, cognitive load,\nattention economics, calm technology, human-computer interaction, generative AI, and digital\nauthenticity that make this synthesis possible. This paper originated as TSX-6, a technical report\nin an independent research series.\n\n15\nConclusion\n\nText is not disappearing. It remains one of the most powerful media for reasoning, explanation, and\ninstitutional memory. The TSX-6 claim is more specific: when generated text becomes abundant,\nthe marginal signaling value of fluent text declines while the value of attention, trust, curation,\nprovenance, and compression rises. LLMs make text easier to produce, but they do not make human\nattention easier to expand.\n\nThe reviewed evidence supports a cautious framework rather than a proof.\nFoundational\n\n15\n\n=== PDF PAGE 16 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\ninformation-overload theory explains why attention becomes scarce under abundance. Calm technol-\nogy and glanceable-interface research show that lower-friction carriers have long been used to reduce\nattention burden. Recent AI-slop, AI-fatigue, workslop, search, model-collapse, and authenticity\nevidence suggests that LLM-generated abundance is producing new pressures around trust, filtering,\nsameness, source visibility, and human signal scarcity. TSX-6 connects these fields into a structural\nmodel: generative abundance creates textual inflation; textual inflation increases semantic entropy\nand cognitive residue; these pressures intensify attention and authenticity scarcity; and the resulting\nenvironment creates selection pressure toward semantic compression.\n\nThe framework should be tested, refined, and limited by evidence. It should not be used to\nclaim that color replaces text, that AI text is inherently valueless, or that semantic compression is\ninevitable. Its strongest defensible claim is that LLM-era textual abundance creates measurable\npressure toward lower-bandwidth, trustworthy, attention-preserving semantic carriers.\n\nReferences\n\n[1] Ahrefs. (2026). Update: AI Overviews reduce clicks by 58%. Online technical analysis. https:\n//ahrefs.com/blog/ai-overviews-reduce-clicks-update/.\n\n[2] Ayyagari, R., Grover, V., and Purvis, R. (2011). Technostress: Technological antecedents and\nimplications. MIS Quarterly, 35(4), 831–858. https://doi.org/10.2307/41409963.\n\n[3] Bawden, D. and Robinson, L. (2009). The dark side of information: Overload, anxiety and\nother paradoxes and pathologies. Journal of Information Science, 35(2), 180–191. https:\n//doi.org/10.1177/0165551508095781.\n\n[4] Bender, E. M., Gebru, T., McMillan-Major, A., and Shmitchell, S. (2021). On the dangers of\nstochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference\non Fairness, Accountability, and Transparency, 610–623. https://doi.org/10.1145/3442188.\n3445922.\n\n[5] Brynjolfsson, E., Li, D., and Raymond, L. R. (2025). Generative AI at work. The Quarterly\nJournal of Economics. Online first / working-paper version also available as arXiv:2304.11771.\n\nhttps://arxiv.org/abs/2304.11771.\n\n[6] Candeto, J. (2024). The price of attention. Phronesis Fund. https://phronesisfund.substa\nck.com/p/the-price-of-attention.\n\n[7] Clementson, J. (2024). How to understand AI fatigue and why people are starting to tune out.\nAzura Magazine. https://azuramagazine.com/articles/how-to-understand-ai-fatigue\n-and-why-people-are-starting-to-tune-out.\n\n[8] Davenport, T. H. and Beck, J. C. (2001). The attention economy: Understanding the new\ncurrency of business. Harvard Business School Press.\n\n[9] Dell’Acqua, F., McFowland, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran,\nS., Krayer, L., Candelon, F., and Lakhani, K. R. (2023). Navigating the jagged technological\nfrontier: Field experimental evidence of the effects of AI on knowledge worker productivity and\nquality. Harvard Business School working paper. https://www.hbs.edu/faculty/Pages/it\nem.aspx?num=64700.\n\n16\n\n=== PDF PAGE 17 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\n[10] Eppler, M. J. and Mengis, J. (2004). The concept of information overload: A review of literature\nfrom organization science, accounting, marketing, MIS, and related disciplines. The Information\nSociety, 20(5), 325–344. https://doi.org/10.1080/01972240490507974.\n\n[11] Gerstgrasser, M. et al. (2024). Is model collapse inevitable? Breaking the curse of recursion by\naccumulating real and synthetic data. arXiv:2404.01413. https://arxiv.org/abs/2404.01413.\n\n[12] Goldhaber, M. H. (1997). The attention economy and the net. First Monday, 2(4). https:\n//doi.org/10.5210/fm.v2i4.519.\n\n[13] Grossman, R., Liu, S., Chen, M. K., Smith, M., Borcea, C., and Chen, Y. (2026). How\ngenerative AI disrupts search: An empirical study of Google Search, Gemini, and AI Overviews.\narXiv:2604.27790. https://arxiv.org/abs/2604.27790.\n\n[14] Kim, J., Schweitzer, S., Riedl, C., and De Cremer, D. (2025). People reduce workers’ compensa-\ntion for using artificial intelligence (AI). arXiv:2501.13228. https://arxiv.org/abs/2501.1\n3228.\n\n[15] Knibbs, K. (2024). AI slop is flooding Medium. Wired. https://www.wired.com/story/ai-g\nenerated-medium-posts-content-moderation/.\n\n[16] Mercante, A. (2026). After an oversaturation of AI-generated content, creators’ authenticity\nand messiness are in high demand. Digiday. https://digiday.com/media/after-an-overs\naturation-of-ai-generated-content-creators-authenticity-and-messiness-are-i\nn-high-demand/.\n\n[17] Miranda, J. P. P., Parreno, E. B., and Rivera, J. G. (2026). Defining AI fatigue in academic con-\ntexts: Dimensions, indicators, and a stage-based model using grounded theory. arXiv:2605.23123.\n\nhttps://arxiv.org/abs/2605.23123.\n\n[18] Niederhoffer, K., Kellerman, G. R., Lee, A., Liebscher, A., and Rapuano, K. (2025). AI-generated\n“workslop” is destroying productivity. Harvard Business Review. https://hbr.org/.\n\n[19] Noy, S. and Zhang, W. (2023). Experimental evidence on the productivity effects of generative\nartificial intelligence. Science, 381(6654), 187–192. https://doi.org/10.1126/science.adh2\n586.\n\n[20] Pew Research Center. (2025). Google users are less likely to click on links when an AI summary\nappears in the results. https://www.pewresearch.org/short-reads/2025/07/22/google-u\nsers-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-res\nults/.\n\n[21] Pousman, Z. and Stasko, J. (2006). A taxonomy of ambient information systems: Four\npatterns of design. Proceedings of the Working Conference on Advanced Visual Interfaces, 67–74.\nhttps://doi.org/10.1145/1133265.1133277.\n\n[22] Ragolane, M. and Patel, S. (2025). Too much, too fast: Understanding AI fatigue in the digital\nacceleration era. International Journal of Arts, Humanities and Social Sciences, 6(8), 53–60.\n\n[23] Ragu-Nathan, T. S., Tarafdar, M., Ragu-Nathan, B. S., and Tu, Q. (2008). The consequences of\ntechnostress for end users in organizations: Conceptual development and empirical validation.\nInformation Systems Research, 19(4), 417–433. https://doi.org/10.1287/isre.1070.0165.\n\n17\n\n=== PDF PAGE 18 ===\nTextual Inflation and Semantic Compression in the LLM Era\nRaynor Eissens\n\n[24] Seddik, M. E. A., Chen, S.-W., Hayou, S., Youssef, P., and Debbah, M. (2024). How bad is\ntraining on synthetic data? A statistical analysis of language model collapse. arXiv:2404.05090.\n\nhttps://arxiv.org/abs/2404.05090.\n\n[25] Shibumi. (2026). AI fatigue statistics 2026: Data on burnout, ROI and tool sprawl. https:\n//shibumi.com/blog/ai-fatigue-statistics-2026/.\n\n[26] Shumailov, I., Shumaylov, Z., Zhao, Y., Gal, Y., Papernot, N., and Anderson, R. (2024). AI\nmodels collapse when trained on recursively generated data. Nature, 631, 755–759. https:\n//doi.org/10.1038/s41586-024-07566-y.\n\n[27] Simon, H. A. (1971). Designing organizations for an information-rich world. In M. Greenberger\n(Ed.), Computers, communications, and the public interest (pp. 37–72). Johns Hopkins Press.\n\n[28] Stanusch, N., Degeling, M., Romano, S., Cetin, R. B., Schueler, M., and Semenzin, S. (2025).\nAI-generated algorithmic virality. arXiv:2508.01042. https://arxiv.org/abs/2508.01042.\n\n[29] Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science,\n12(2), 257–285. https://doi.org/10.1016/0364-0213(88)90023-7.\n\n[30] Tarafdar, M., Tu, Q., Ragu-Nathan, B. S., and Ragu-Nathan, T. S. (2007). The impact of\ntechnostress on role stress and productivity. Journal of Management Information Systems,\n24(1), 301–328. https://doi.org/10.2753/MIS0742-1222240109.\n\n[31] Weiser, M. (1991). The computer for the 21st century. Scientific American, 265(3), 94–104. ht\ntps://www.scientificamerican.com/article/the-computer-for-the-21st-century/.\n\n[32] Weiser, M. and Brown, J. S. (1996). Designing calm technology. Xerox PARC / conference\npaper. https://people.csail.mit.edu/rudolph/Teaching/weiser.pdf.\n\n[33] WordPress VIP. (2026). Consumer trust, AI-generated information, and human-centered web\ncontent. Online survey report; public summaries reported by technology press.\n\n[34] Zhu, T., Weissburg, I., Zhang, K., and Wang, W. Y. (2024). Human bias in the face of\nAI: The role of human judgement in AI-generated text evaluation. arXiv:2410.03723. https:\n//arxiv.org/abs/2410.03723.\n\n[35] Axios. (2026). AI overuse could spark “brain fry,” new research finds. https://www.axios.co\nm/2026/03/06/ai-chatgpt-claude-jobs-brain-fry.\n\n18"
}