top of page
検索

生成検索で『正統性』は不可逆に移るのか：ALS先行研究レビュー

1. イントロダクション：供給と需要の帰結としての「不可逆相」

本レポート（第3回）が扱うのは、その帰結である。 すなわち、供給と需要の相互作用によって生じたALS（Algorithmic Legitimacy Shift）が、本モデルで定義される「社会的前提・説明責任・正当化可能性が不可逆に変質する段階（Irreversible Regime）」に達しているかどうかの検証である。

本報告では、実証研究・公式調査・実務研究を統合し、現状がALSモデル上の「不可逆相」と整合するかどうか、Working Assessment（状態推定）を行う。

1.1 Methods (Evidence Collection and Integration)

Scope. This report integrates empirical, policy, and practitioner evidence relevant to ALS and the working state-estimation of the Irreversible Regime (IR). The scope is limited to (i) generative / LLM-based search behaviors, (ii) social frictions around AI use, and (iii) organizational decoupling of authority and responsibility under algorithmic mediation.

Search window. 2024–2026 (last accessed: 2026-01-24).

Primary sources prioritized. Peer-reviewed venues (PNAS, PNAS Nexus, CHI/ACM), official policy (OECD), and primary survey publishers (Pew Research Center, Ipsos). Practitioner UX research (Nielsen Norman Group) is included only as mechanism illustration.

Inclusion criteria. A source is included if it reports (a) user behavior changes under AI summaries/LLM search, (b) reliance/overreliance or learning-depth effects, (c) social evaluation penalties / concealment behaviors, or (d) institutional expansion of algorithmic management affecting responsibility/authority.

Exclusion criteria. Pure opinion pieces without traceable underlying surveys, sources lacking minimally stated methods/sample, and secondary summaries when primary sources are accessible.

Integration procedure. Evidence is clustered into three domains (2.1–2.3). Each item is recorded as: (i) Fact statement (source-anchored), (ii) ALS mechanism interpretation (model-level), and (iii) mapped model variables with declared measurement type (ratio/ordinal/binary/qualitative indicator). The final state-estimation (Section 3) is explicitly a Working Assessment, not a universal sociological claim.

1.2 Operational Definition (Working): Irreversible Regime (IR)

In this report, IR is treated as a model-state in which the default social/epistemic pipeline becomes path-dependent under generative/LLM-mediated information access. We operationalize IR as the co-presence of three conditions:

  • (i) Anchoring / verification suppression ($\alpha$): verification behaviors (e.g., external link clicks) are structurally reduced when AI summaries are available, implying that AI outputs become default priors.

  • (ii) Reliance reinforcement under known error risk ($\rho$): LLM-based search improves task convenience but increases overreliance or reduces depth of learning, implying reliance becomes behaviorally self-reinforcing.

  • (iii) Organizational decoupling ($\gamma$): algorithmic mediation expands in workplaces such that authority/decision formation becomes algorithmically structured while responsibility remains socially assigned to humans.

IR in this document is a Working Assessment: evidence can support “IR-consistent” or “IR-near-threshold” states without claiming irreversible finality in an absolute sociological sense.

Measurement Types for Model Variables:

  • $\alpha_{anchor}$: ratio / proportion indicator (e.g., click-rate differences)

  • $\rho_{reliance}$: empirical/behavioral indicator (ordinal/ratio depending on study metrics)

  • $\beta_{penalty}$: empirical/behavioral indicator (ordinal/ratio)

  • $H_{hide}$: proportion indicator (concealment rates)

  • $\gamma_{hollow}$: policy/organizational indicator (ordinal qualitative → mapped score)



2. 領域別分析と「不可逆相」への整合性検証

Evidence Tier (GMI Standard)

  • T1: Peer-reviewed empirical (experiments / observational with methods; journals or flagship conferences)

  • T2: Peer-reviewed non-empirical (editorial, extended abstract, position/theory; limited empirical weight)

  • T3: Official policy / intergovernmental report (OECD etc.)

  • T4: Primary survey / polling publisher with disclosed sampling or methodology (Pew, Ipsos)

  • T5: Practitioner / UX research (methods may be proprietary; used for mechanism illustration)

  • T6: Journalism / commentary (used only to reference otherwise inaccessible polling; never used as a sole basis for “Fact” claims)

Rule: “Fact statements” must be supported by T1–T4. T5–T6 are restricted to mechanism illustration or context and must be labeled as such.

2.1 エピステモロジーのヒステリシス（認識論的履歴効果）

～検索・LLM行動における「検証コスト」と「正解の固定化」～

No.

Study / Phenomenon

Evidence Tier

Study Design / Sample

Primary Outcome / Metric

ALS Interpretation (Mechanism)

Model Variable

Limitations

1

Pew Research Center (2025)


AI要約とクリック行動

T4

National survey (US adults)

CTR (Click-Through Rate)


AI summaryあり: 8%


なし: 15%

【正統性のアンカー効果】


AI要約提示時、外部検証行動（クリック）が構造的に抑制される。AI出力が初期値として固定化される初期条件。

$\alpha_{anchor}$


(Ratio)

Causal inference limited by observational nature

2

Nielsen Norman Group (2025)


検索行動の変化

T5

Qualitative UX study

Behavioral patterns (Attention allocation)

【前提化のプロセス（例示）】


情報のゲートキーパー権限がアルゴリズムへ移動する具体的メカニズムの例示。

$T_{attention}$


(Ordinal)

Non-random sampling; Qualitative only

3

Melumad & Yun (PNAS Nexus, 2025)


LLM vs ウェブ検索

T1

7 Experiments (n≈10,462)

Depth of learning / Understanding scores

【知識獲得コストのダンピング】


便利さが駆動要因となり、浅い学習プロセスがデフォルト化する経路依存性を支持。

$D_{depth}$


(Ratio)

Task-specific context (advice generation)

4

Spatharioti et al. (CHI 2025)


意思決定への影響

T1

Controlled Experiments

Speed, Accuracy, Reliance rates

【依存の制度化】


LLM検索は速度を上げるが、誤りに対する過信（Overreliance）も誘発する。単純な精度向上ではなく、判断委譲のリスク増大を示す。

$\rho_{reliance}$


(Ratio)

Lab setting may differ from wild usage

Fact (Evidence). Pew Research Center (2025) reports that when an AI summary is shown on Google search results, users are less likely to click through to external links compared with results without the AI summary (reported proportions: 8% vs 15%). It also notes that clicks on links inside the AI summary itself are rare (reported at ~1% in the report’s measurement).

Model-based interpretation (ALS / IR Working Assessment). The observed reduction in verification behaviors supports condition (i) Anchoring/Verification Suppression. $\alpha_{anchor}$ indicates a structural shift where AI outputs become default priors, consistent with an IR-near-threshold state.

2.2 正統性移動への摩擦と反作用（Social Friction）

～ヒューマン–AIインタラクションにおける規範的障壁～

No.

Study / Phenomenon

Evidence Tier

Study Design / Sample

Primary Outcome / Metric

ALS Interpretation (Mechanism)

Model Variable

Limitations

5

Reif et al. (PNAS, 2025)


社会的評価ペナルティ

T1

Experiments (n=4,400)

Perceived competence / Warmth scores

【正統性移動の二重性】


機能的正統性はAIへ移動するが、人間側には社会的ペナルティ（摩擦）が発生する。

$\beta_{penalty}$


(Ordinal)

Short-term evaluation focus

6

Ipsos (2025) / Reporting


職場での隠蔽行動

T4

Polling (UK workers)

Concealment rate (29%) / Anxiety rate (26%)

【隠蔽行動による前提化の強化】


AI利用を隠す（Black-box adoption）ことで、実質的支配が不可視化されつつ進行する。

$H_{hide}$


(Proportion)

Self-reported data

7

Sarkar (CHI EA 2025)


AI使用への蔑視語

T2

Discourse Analysis

Existence of classist slurs

【正統性境界の防衛戦】


正統性移動に対する最後の抵抗線（境界維持装置）としての言説機能。

$B_{boundary}$


(Qualitative)

Theoretical/Interpretive

8

Acut & Gamusa (2025)


教育現場のAI Shaming

T2

Qualitative / Reflection

Perception of academic integrity

【専門職権威の揺らぎ】


正しさの源泉がAIに移ることへの制度的拒絶反応。

$A_{authority}$


(Qualitative)

Context-specific (Teacher Ed)

Fact (Evidence). Reif et al. (2025) demonstrate a social penalty for AI use. Ipsos polling (2025) indicates that 29% of surveyed workers conceal AI use from colleagues, suggesting a decoupling between actual practice and stated norms.

Model-based interpretation (ALS / IR Working Assessment). Concealment behavior supports a “black-box adoption” pathway: outputs circulate as if human-authored while the decision substrate shifts algorithmically, consistent with IR condition (iii) Decoupling and the transition into a latent legitimacy-transfer phase.

2.3 責任の蒸発と主体性の形骸化（Decoupling）

～アルゴリズム管理と労働・組織～

No.

Study / Phenomenon

Evidence Tier

Study Design / Sample

Primary Outcome / Metric

ALS Interpretation (Mechanism)

Model Variable

Limitations

9

OECD (2025)


職場のアルゴリズム管理

T3

Employer Survey (6 countries, 6000+ firms)

Adoption rates / Management practices

【判断主体の形式化】


意思決定プロセスへのアルゴリズム介在拡大。判断の実質的正統性がAIへ移動。

$\gamma_{hollow}$


(Ordinal)

Employer-reported bias possible

10

Bowdler et al. (SJWEH, 2026)


心理社会的リスク

T2

Editorial / Lit Review

Risk pathway conceptualization

【知識→組織→身体への連鎖】


AI前提化が身体的・精神的負荷としてフィードバックされるリスク構造の整理。

$S_{stress}$


(Qualitative)

Non-empirical synthesis

Fact (Evidence). This editorial synthesizes emerging occupational safety and health concerns regarding algorithmic management and psychosocial risks, proposing the risk pathway rather than reporting new primary experimental data. OECD (2025) confirms wide adoption of algorithmic management tools.

Model-based interpretation (ALS / IR Working Assessment). Institutional adoption without clear responsibility frameworks supports condition (iii) Organizational Decoupling. The shift of authority to algorithms while responsibility remains with humans suggests the system is entering an IR-consistent state.


3. 結論：GhostDriftモデルへの実装に向けた状態推定

State Labels (Working)

  • S0: Pre-IR (insufficient evidence for any IR condition)

  • S1: Near-threshold (strong evidence for one condition + suggestive evidence for another)

  • S2: IR-consistent (co-presence of evidence supporting (i)(ii)(iii) within the scope of this report)

Note: These labels are internal to GMI’s ALS model and are not presented as universal sociological classifications.

Working Assessment

Based on the integrated evidence across (i) anchoring/verification suppression, (ii) reliance reinforcement under known error risk, and (iii) organizational decoupling under algorithmic mediation, the current environment is best labeled as S1–S2 (near-threshold to IR-consistent) within the scope of this report.

Model parameterization in GhostDrift’s ALS implementation should therefore assume non-trivial $\alpha$, persistent $\beta$ with concealment pathways, and institutionalized $\gamma$ as a baseline.


References

  • Pew Research Center. (2025, July 22). Google users are less likely to click on links when an AI summary appears in the results. Short Reads. (accessed 2026-01-24).

  • Moran, K., Rosala, M., & Brown, J. (2025, August 15). How AI Is Changing Search Behaviors. Nielsen Norman Group. (accessed 2026-01-24).

  • Melumad, S., & Yun, J. H. (2025). Experimental evidence of the effects of large language models versus web search on depth of learning. PNAS Nexus, 4(10), pgaf316. https://doi.org/10.1093/pnasnexus/pgaf316

  • Spatharioti, S. E., Rothschild, D., Goldstein, D. G., & Hofman, J. M. (2025). Effects of LLM-based Search on Decision Making: Speed, Accuracy, and Overreliance. CHI ’25. https://doi.org/10.1145/3706598.3714082

  • Reif, J. A., Larrick, R. P., & Soll, J. B. (2025). Evidence of a social evaluation penalty for using AI. Proceedings of the National Academy of Sciences, 122(19), e2426766122. https://doi.org/10.1073/pnas.2426766122

  • Ipsos. (2025, September 15). Nearly one in five Britons turn to AI for personal advice and support. Ipsos. (accessed 2026-01-24).

  • OECD. (2025). Algorithmic management in the workplace: New evidence from an OECD employer survey. OECD AI Papers, No. 24. Paris: OECD Publishing. https://www.google.com/search?q=https://doi.org/10.1787/287c13c4-en

  • Bowdler, M., et al. (2026; online 2025). Algorithmic management and psychosocial risks at work: An emerging occupational safety and health challenge. Scandinavian Journal of Work, Environment & Health, 52(1), 1–5. (Editorial). https://doi.org/10.5271/sjweh.4270

  • Sarkar, A. (2025). AI Could Have Written This: Birth of a Classist Slur in Knowledge Work. CHI EA ’25. https://doi.org/10.1145/3706599.3716239

  • Acut, D. P., & Gamusa, E. V. (2025). AI Shaming Among Teacher Education Students: A Reflection on Acceptance and Identity in the Age of Generative Tools. In Pitfalls of AI Integration in Education (IGI Global). https://doi.org/10.4018/979-8-3373-0122-8.ch005

 
 
 

コメント

bottom of page