{"id":"a2a6b7e8-583d-45f6-9ced-e9c1eed3076a","arxiv_id":"2507.22893","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper reframes AI as invisible cognitive infrastructure that pre-consciously reshapes thinking, and proposes a new interdisciplinary field to study it.","lead":"This paper proposes a new research field, Cognitive Infrastructure Studies, for studying how AI systems invisibly pre-shape human thought before conscious awareness. It argues that search engines, recommender systems, and language models act as cognitive infrastructures, and should be studied and governed like public utilities.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Core construct 'preconscious' lacks an operational test; proposed breakdown methods cannot distinguish deep cognitive integration from ordinary tool reliance, so the central claim is not yet falsifiable.","rationale":"The reader's weakest assumption, that breakdown methodologies are unvalidated and individual-to-collective scaling is assumed, is close to the mark. I sharpen it: even a successful breakdown study would not discriminate the paper's core claim because the contrast class of ordinary tool reliance is not controlled. The paper explicitly calls System 0 'an experimental construct that can be operationalised and measured today' (Section 1) and says the proposed methods can differentiate superficial tool use from deep infrastructural integration (Section 5), but no operational definition is provided. This is not a dispute with external consensus; it is an internal gap between the central claim and the proposed method. The concern does not require rejecting the framework: a revised version could add the transparent-tool control, a pre-registered marker of preconscious integration, and at least a minimal model of how individual breakdown effects propagate to collective outcomes. The reader's CONDITIONAL verdict already captures this need, so I recommend no change. My agreement is partial because the reader emphasized the unvalidated status of the method, while I emphasize that the method as described cannot, even in principle, separate the core preconscious claim from tool-use dependence.","tokens_in":12728,"tokens_out":3417,"duration_ms":40485,"concrete_test":"Preregister a three-arm experiment where participants are habituated for two weeks to: (A) an adaptive AI news summarizer and ranker that invisibly personalizes content; (B) a transparent static digest delivering matched baseline content with no personalization; (C) no preprocessor. At withdrawal, measure task performance (comprehension, decision accuracy, recall), eye-tracking, self-reported awareness, and a structured interview probing strategies. Analyze whether the performance loss in A exceeds B after controlling for content novelty and effort, and whether A produces qualitatively different markers, such as residual attentional bias, persistence after explicit re-training, or psychophysiological response.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that AI systems reshape cognition 'pre-consciously' as an invisible infrastructure, not merely as tools people deliberately use. What would have to be true: withdrawal of AI preprocessing after habituation reveals a distinctive cognitive dependence that is not explained by loss of a familiar, useful tool. The paper's only proposed method (Section 5, 'infrastructure breakdown methodologies') cannot establish this. Performance degradation after removing a summarizer or ranker is exactly what ordinary tool reliance predicts—take away a calculator and arithmetic slows. The paper asserts these methods 'help differentiate between superficial tool usage and deep infrastructural integration' but specifies no operational measure of 'preconscious,' 'depth of coupling,' or 'baseline reasoning system.' Without a control for transparent tool use and a pre-specified marker (e.g., persistence after retraining, attentional capture, physiological response) the proposed experiments cannot confirm System 0's distinctive preconscious status. A second gap: the individual-level finding, even if secured, is assumed to scale to collective and societal outcomes solely via narrative scenarios (Sections 4.2 and 4.3), with no aggregation mechanism or population-level data. The central claim is therefore underdetermined by the evidence and method currently on offer.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This position paper introduces \"Cognitive Infrastructure Studies\" (CIS) as a proposed new interdisciplinary domain for studying AI systems as \"cognitive infrastructures\" (or \"System 0\"): always-on, largely invisible algorithmic layers that filter, rank, and transform information before it reaches conscious awareness. The paper synthesizes five theoretical traditions—STS, distributed/extended cognition, digital sociology, infrastructure studies, and argumentative theory—to argue that AI preprocessing reshapes individual cognition, collective deliberation, and societal knowledge governance. It illustrates these claims through three narrative scenarios (a dependent professional, a civic engagement platform, and a national audit of AI access), and it proposes \"infrastructure breakdown methodologies\" as an experimental route to make invisible algorithmic influence visible by withdrawing AI support after habituation. The manuscript presents no new data, no formal derivation, and no pilot study; it is a conceptual and programmatic proposal. Its central empirical claim is that AI preprocessing operates preconsciously and constitutes a distinctive form of cognitive infrastructure rather than merely being a useful tool.","tokens_in":13072,"tokens_out":3771,"duration_ms":40095,"significance":"If the framework were made empirically operational, it could provide a useful new lens for human-AI interaction research, foregrounding ambient and habitual effects that tool-centric models tend to miss. The paper's strengths are its serious engagement with infrastructure studies (Star and Ruhleder, Bowker and Star), its deliberate anchoring in established theory (Clark and Chalmers, Hutchins, Mercier and Sperber), and its concrete, falsifiable-in-principle proposal to study withdrawal effects through breakdown experiments. The normative questions raised—cognitive dependency, epistemic agency, democratic fragmentation, and governance as public utility—are timely and important for a human-centered computing venue. At the same time, the manuscript's central empirical assertion is currently not supported: the proposed methods lack the operational specificity needed to distinguish preconscious infrastructural integration from ordinary tool reliance, and the claimed individual-to-societal scaling rests on narrative extrapolation rather than evidence or a formal aggregation mechanism.","major_comments":[{"comment":"The key methodological claim in §5—that \"infrastructure breakdown methodologies\" can \"help differentiate between superficial tool usage ... and deep infrastructural integration\"—is not supported by the designs described. Withdrawing algorithmic summarization, ranking, or filtering after habituation predicts performance degradation, strategy shifts, and attentional disruption under ordinary tool-use accounts as well; the paper specifies no operational measure of \"preconscious\" processing, no control condition for transparent tool use, and no pre-registered marker that distinguishes a cognitive extension from a familiar, valuable tool. Without such a differential indicator, the proposed experiments cannot confirm System 0's distinctive preconscious status.","section":"Section 5"},{"comment":"The societal-level claims—for example, that AI preprocessing poses \"a systemic threat to democratic deliberation and societal cohesion\" (§2.5, §3)—are built on narrative scenarios and an asserted multi-scale measurement agenda. Section 5 calls for integrating individual-level cognitive assessment, social network analysis, and cultural analysis, but it provides no aggregation model, no population-level data, and no account of how individual-level cognitive dependency scales to collective epistemic fragmentation. The individual-to-societal inference is therefore a conjecture, not a demonstrated result.","section":"Sections 4.2-4.3 and 5"},{"comment":"The paper states in §1 that \"System 0 is an experimental construct that can be operationalised and measured today,\" yet the manuscript does not provide an operational definition of \"preconscious\" influence, a dependent measure that separates System 0 from System 1/2-mediated tool use, or a falsifiable prediction linking anticipatory personalization and adaptive invisibility to a specific behavioral, physiological, or cognitive signature. Without this operationalization, the central construct remains at the level of metaphor.","section":"Section 1"},{"comment":"A substantial part of the framework's grounding comes from the author's own prior work: Chiriatti et al. 2024 and 2025 for System 0, Riva 2025a for \"Digital We,\" and Riva 2025b—which is the current preprint itself—for the term \"cognitive infrastructure.\" Citing one's own current preprint as an independent source in §2.1 gives the framework a false appearance of external support; the self-referential status should be acknowledged, and independent conceptual or empirical grounding should be supplied or the relevant claims should be softened.","section":"Section 2.1 and References"}],"minor_comments":[{"comment":"The keyword \"adaptive invisibilty\" contains a typo; it should read \"adaptive invisibility.\"","section":"Keywords/Abstract"},{"comment":"The sentence containing \"as we just hve seen, a demonstrates\" contains typographical errors that should be corrected.","section":"Section 2.5"},{"comment":"The reference for Mercier and Sperber (2017) includes a DOI (10.14763/2023.1.1683) that does not correspond to The Enigma of Reason; please verify and correct the DOI or remove it.","section":"References"},{"comment":"Tables 1 and 2 are information-dense; the text should summarize their key contrasts in the main body rather than relying on lengthy cell text that is not closely discussed.","section":"Tables 1-2"},{"comment":"The inline citation \"NMI 2025\" uses an abbreviation without a clear author label; since it refers to an editorial, please render it in a consistent citation style.","section":"Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"The paper's core constructs are repeatedly grounded in the author's own prior publications, including the current preprint, which may inflate the appearance of novelty. Editors may also wish to consider whether a position paper with no empirical data and no formal derivation fits the journal's expected standards for submission; the proposed revisions should strengthen the empirical and methodological content or explicitly reposition the paper as a hypothesis-generating research agenda."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read. The real value is the synthesis. Riva pulls five established literatures—STS, distributed cognition, digital sociology, infrastructure studies, argumentative theory—into one frame and names it Cognitive Infrastructure Studies. That is genuinely useful for HCI and policy researchers who keep rediscovering the same \"AI shapes cognition\" point from different angles. Table 1 is well done, and the writing is clear throughout.\n\nWhat is not new: System 0 comes from the author's own prior Nature Human Behaviour commentary and Cyberpsychology paper. The preprint cites itself (Riva 2025b) as support for \"cognitive infrastructures,\" and \"Digital We\" (Riva 2025a) is likewise self-referential. That is circular and should be cleaned up; the outside anchor (Antikythera 2025) exists but the heavy self-reliance weakens the argument.\n\nThe bigger soft spot is methodological. The proposed \"infrastructure breakdown methodologies\"—withdraw AI preprocessing after habituation and observe—are plausible in spirit, but the paper does not specify how to distinguish deep cognitive integration from ordinary tool reliance. Take away a calculator and arithmetic slows; that does not mean the calculator reshaped your arithmetic pre-consciously. The paper asserts these methods \"help differentiate between superficial tool usage and deep infrastructural integration\" but gives no operational marker (persistent deficits after retraining, attentional capture, physiological response, etc.). So the central claim—that AI reshapes cognition pre-consciously as infrastructure—remains unfalsifiable as stated. The individual-to-collective scaling is asserted through narrative scenarios (Section 4) rather than a mechanism or data.\n\nThat said, I don't think this is a takedown. The conceptual framework is coherent, and the author is explicit that this is a proposal, not empirical validation. The scenarios are clearly illustrative. For a programmatic position paper, it is decent. For a journal, the self-citation pattern and unoperationalized central construct need revision.\n\nWho gets value: graduate students entering human-AI interaction, and policy readers wanting a vocabulary for \"AI as environment\" rather than \"AI as tool.\" It deserves a serious referee, not a desk reject, but the referee should demand that self-citations be replaced with independent support, a falsifiable protocol for System 0 be specified, and the aggregation claims be tempered. I'd send it to review with clear revision guidance.","headline":"A clear, useful synthesis naming a new field, but the central System 0 construct leans on self-citations and the proposed breakdown methods don't yet make preconscious influence testable.","tokens_in":13464,"tokens_out":1966,"would_cite":false,"duration_ms":20220,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"AI systems act as invisible cognitive infrastructure, reshaping thought before conscious awareness, and this paper proposes a new science to study that layer.","keywords":["Cognitive Infrastructure","System 0","algorithmic preprocessing","epistemic agency","infrastructure breakdown methodologies","distributed cognition","digital governance","cognitive inequality"],"falsifier":"A controlled longitudinal experiment could settle this: habituate one group to an AI-curated information environment and a matched control group to a static, non-algorithmic environment, then silently switch the AI group to the control condition. If reasoning quality, attention allocation, decision-making accuracy, and information diversity show no deterioration beyond what the control group experiences from routine environmental variation, the core dependency claim fails. A sharper version would test the infrastructure-specific claim by comparing silent removal of algorithmic preprocessing against silent removal of an equally informative but non-adaptive feed: identical effects would indicate the distinctive properties of the AI infrastructure, not generic information loss, are what matter.","tokens_in":12503,"feed_emoji":"🧠","tokens_out":5433,"duration_ms":50311,"temperature":0.7,"pith_summary":"This paper argues that mainstream human-AI interaction research has a blind spot: AI systems do not merely assist thinking as tools, they condition it before conscious awareness, by filtering, ranking, and curating information. It names that preconscious algorithmic layer 'System 0' and proposes 'Cognitive Infrastructure Studies' (CIS) as a new scientific field to study AI as cognitive infrastructure — systems that are invisible in normal use, adaptive to users, and increasingly authoritative over what counts as relevant, knowable, and actionable. The paper develops four distinguishing properties of cognitive infrastructures: anticipatory personalization, adaptive invisibility, automation of relevance judgment, and a shift in the locus of epistemic agency from humans to non-human systems. It then offers a methodological innovation, 'infrastructure breakdown methodologies', in which AI preprocessing is withdrawn after a period of habituation to reveal hidden cognitive dependencies. If the framework is right, it matters because the governance target changes from regulating individual AI applications to managing the infrastructural conditions under which societies know, decide, and evolve.","feed_headline":"AI pre-processes thought before you think — a new field to study it","feed_subtitle":"This paper introduces Cognitive Infrastructure Studies and System 0, the invisible layer where algorithms set what is knowable.","key_machinery":"The central object is 'System 0': an invisible, non-human layer of distributed cognition embedded in AI-mediated infrastructure that precedes both Kahneman's System 1 (fast, intuitive) and System 2 (slow, analytical) thinking. The concept is carried by three load-bearing components: the eight infrastructure characteristics of Star and Ruhleder (embeddedness, transparency, reach, learning, linkage, standards, installed base, and visibility upon breakdown), applied to cognitive processes; the four distinctive properties of cognitive infrastructures — anticipatory personalization, adaptive invisibility, automation of relevance judgment, and the relocation of epistemic agency; and 'infrastructure breakdown methodologies', experimental designs that habituate users to an AI-mediated environment and then withdraw or degrade algorithmic preprocessing to make cognitive dependence observable in performance degradation, strategy shifts, and attentional breakdown. The machinery works by transposing infrastructure theory's principle that infrastructure becomes visible upon breakdown from material systems to cognitive environments.","core_discovery":"The central claim is that AI preprocessing, System 0, reshapes human cognition, collective reasoning, and societal functioning in invisible yet foundational ways. The author reconceptualizes AI systems — search engines, recommender systems, algorithmic curation platforms, and large language models — not as tools that people consciously engage, but as cognitive infrastructures that are continuously active, anticipate user behavior, and automate relevance judgment before it reaches deliberate thought. This automation transfers a traditionally human cognitive task to non-human systems, shifting the locus of epistemic agency, and the paper argues that the combination of anticipatory personalization and adaptive invisibility creates a self-reinforcing loop that fragments shared reality into personalized information environments. Because Mercier and Sperber's argumentative theory ties effective collective reasoning to shared factual foundations, that fragmentation is presented as a systemic threat to democratic deliberation. The paper's position is that these dynamics across individual, collective, and societal scales are not anecdotal but signals of a structural transformation that current frameworks are inadequate to capture.","pith_inferences":["The framework implies that every major platform algorithm change is a natural breakdown experiment; reanalyzing existing longitudinal trace data from such changes could test cognitive dependency without new habituation studies.","A governance corollary the author leaves implicit is the need for 'breakdown audits': routine, adversarial tests in sandboxed or simulation settings where algorithmic support is withdrawn to map hidden dependencies before they are exploited.","The concept of 'cognitive sovereignty' follows as a new normative right: not just privacy over data, but a right against silent preprocessing of one's reasoning environment.","One testable extension is differential breakdown signatures: if dependency is real, its signs should appear earlier and stronger in populations with lower algorithmic literacy or less diverse information diets."],"forward_implications":["Governance would shift from policing individual AI applications to managing the underlying cognitive architectures as public utilities, with standards for transparency, accessibility, and interoperability.","Cognitive inequity becomes a measurable axis of inequality: differential access to thinking-shaping infrastructures stratifies who benefits from machine-mediated cognition.","Automated relevance judgments would mean decisions about what is worth knowing, seeing, or acting upon are increasingly performed by non-human systems, changing the knowledge landscape at scale.","Personalized curation fragments the shared epistemic foundations that argumentative theory identifies as necessary for productive democratic deliberation.","Infrastructure breakdown methodologies would provide an empirical route to distinguish deep cognitive coupling from superficial tool use, by measuring the depth of dependence through withdrawal."],"supporting_citations":[{"why":"Supplies the eight characteristics of infrastructure, including visibility upon breakdown, that the paper applies to AI systems.","marker":"Star & Ruhleder (1996)"},{"why":"Establishes the extended mind thesis, grounding System 0 as a constitutive external cognitive layer.","marker":"Clark & Chalmers (1998)"},{"why":"Defines System 0, the concept of an invisible non-human layer of distributed cognition that this paper develops.","marker":"Chiriatti et al. (2024)"},{"why":"Provides the argumentative theory of reasoning used to argue that fragmented personalized realities undermine shared epistemic foundations.","marker":"Mercier & Sperber (2011)"},{"why":"Shows how classification systems invisibly shape cognition, supporting the claim that algorithmic preprocessing acts as infrastructure.","marker":"Bowker & Star (1999)"},{"why":"Documents the platform society and how algorithms mediate social life, grounding the societal-scale claims.","marker":"van Dijck, Poell, & de Waal (2018)"}],"fun_headline_variants":["AI rewires thinking before you notice","System 0: the hidden layer shaping your thoughts","AI pre-processes cognition invisibly","New science unveils AI's invisible cognitive grip","AI shapes what you think, before you think"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework's load-bearing premise is that after habituation, withdrawing AI preprocessing will reveal measurable cognitive dependency rather than mere annoyance, adaptation, or learned strategy shifts, and that these individual-level effects scale up to collective and societal outcomes.","fun_headline_variants_meta":{"raw":{"variants":["AI rewires thinking before you notice","System 0: the hidden layer shaping your thoughts","AI pre-processes cognition invisibly","New science unveils AI's invisible cognitive grip","AI shapes what you think, before you think"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00056,"raw_usage":{"total_tokens":2678,"prompt_tokens":980,"completion_tokens":1698,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":1630}},"tokens_in":596,"tokens_out":1698,"duration_ms":12016,"temperature":1.0,"reasoning_tokens":1630,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:26:53.855045+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled longitudinal experiment could settle this: habituate one group to an AI-curated information environment and a matched control group to a static, non-algorithmic environment, then silently switch the AI group to the control condition. If reasoning quality, attention allocation, decision-making accuracy, and information diversity show no deterioration beyond what the control group experiences from routine environmental variation, the core dependency claim fails. A sharper version would test the infrastructure-specific claim by comparing silent removal of algorithmic preprocessing against silent removal of an equally informative but non-adaptive feed: identical effects would indicate the distinctive properties of the AI infrastructure, not generic information loss, are what matter.","supporting_citations":[],"review_version":2}