pith. sign in

arxiv: 2605.27396 · v1 · pith:YNS543KLnew · submitted 2026-04-21 · 💻 cs.CY · cs.AI

Agentic Literacy Debt: A Structural Problem the AI Literacy Field Has Not Yet Named

Rohith Nama This is my paper

Pith reviewed 2026-07-05 07:26 UTC · model glm-5.2

classification 💻 cs.CY cs.AI
keywords literacyagenticagentsdebtproblembehalfcontextsdecide
0
0 comments X

The pith

Naming the debt AI agents create

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper introduces the concept of _agentic literacy debt_: the accumulating societal deficit that grows when autonomous AI agents are deployed at scale without the literacy infrastructure people need to govern them. The author argues that every existing AI literacy framework was built for a world where humans evaluate AI outputs and decide whether to act. Agentic AI — systems that plan, decide, and execute actions on a user's behalf across email, payments, healthcare, and more — breaks all three assumptions those frameworks depend on: that users can see outputs (evaluation), that bad decisions can be reversed, and that the human remains the agent of action. When an AI agent acts autonomously, the user becomes a principal who has delegated authority to a system whose actions may be invisible, irreversible, and uncontrollable. The debt compounds through three reinforcing channels: normalization (each opaque delegation habituates users to granting permissions without scrutiny, creating a ratchet effect on access), ecosystem complexity (each new agent interacts with previously deployed agents, producing multi-agent chains harder to oversee than any single system), and institutional path dependence (organizations that skip literacy infrastructure for one deployment build no capacity for the next, and retrofitting costs grow with each iteration). The debt is incurred by the organizations that deploy agents but paid by the users, patients, and citizens the agents act upon — an asymmetry the author identifies as what makes this an ethics problem rather than merely an educational one. The paper draws on evidence from healthcare (where agentic AI markets are growing at 45% annually and trust calibration failures are already documented), financial fraud (where attacks increasingly target the AI agent rather than the human user, with prompt injection rated the top risk in production LLM applications), and global equity (where populations least served by literacy research are most exposed to agentic deployment). The author contends this gap is structural, not temporary, because national curriculum cycles take five to seven years while agentic AI capabilities evolve on product cycles measured in months. The paper calls for a reframing of AI literacy from an evaluative capability to a governance capability, identifying six new principal-side competencies: delegation, o[

Core claim

The paper's central contribution is naming and structuring a problem the AI literacy field has not yet articulated: that the shift from generative AI (which produces outputs humans evaluate) to agentic AI (which takes actions humans may never observe) invalidates the three foundational assumptions of every existing literacy framework — evaluation, reversibility, and control. The author defines _agentic literacy debt_ as the compounding societal deficit that accumulates through three reinforcing channels (normalization of opaque delegation, multi-agent ecosystem complexity, and institutional path dependence) when agentic systems are deployed without corresponding literacy infrastructure. The贷

What carries the argument

The argument rests on a structural analogy to technical debt in software engineering: expedient short-term deployment decisions create compounding future costs. The debt metaphor is extended from two precedents — Ladson-Billings's 'education debt' (reframing annual achievement gaps as cumulative structural deficits) and Petrozzino's 'ethical debt' in AI (costs incurred by developers but paid by marginalized communities). Agentic literacy debt extends both: it is incurred at the point of deployment (not design), compounds with every user interaction lacking literacy infrastructure, and is paid by users rather than deployers. The three compounding channels operate as a self-reinforcing system:

Load-bearing premise

The paper's load-bearing premise is that the speed mismatch between agentic AI deployment (months) and institutional curriculum adaptation (five to seven years) is permanent — that no institutional delivery mechanism can adapt fast enough to close the gap. If institutions develop faster, technology-embedded literacy delivery (such as contextual micro-learning at the point of risk, which the paper itself mentions), the 'permanent structural condition' framing weakens and the债务

What would settle it

If institutions or deploying organizations successfully embed literacy infrastructure at the point of deployment — through contextual micro-learning, transparency-by-design, or other mechanisms that adapt at product-cycle speed — and the compounding channels (normalization, ecosystem complexity, path dependence) are demonstrably interrupted, the 'permanent structural debt' framing collapses into a temporary lag that conventional adaptation could close.

read the original abstract

Autonomous AI agents now plan, decide, and act on behalf of users across healthcare, financial services, and workplace contexts, often without step-by-step human approval. Existing AI literacy frameworks were built for a world in which humans evaluate AI outputs and decide whether to act; they have no vocabulary for the user who has delegated decision-making authority to an agent whose actions may not be observable, reversible, or controllable. This paper names the resulting problem agentic literacy debt: the accumulating societal deficit that grows when agentic AI systems are deployed at scale without corresponding literacy infrastructure. The debt compounds through three reinforcing channels (normalization of opaque delegation, multi-agent ecosystem complexity, and institutional path dependence), and it is incurred by the organizations that deploy agents but paid by the users, patients, and citizens on whose behalf the agents act. Evidence from healthcare, financial fraud, and global equity contexts suggests the gap is already consequential. The problem is structural, not a temporary lag that curriculum reform will close. It demands a reframing of AI literacy as a governance capability, not an evaluative one.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 7 minor

Summary. This paper introduces the concept of 'agentic literacy debt'—the accumulating societal deficit that arises when autonomous AI agents are deployed at scale without corresponding literacy infrastructure. The author argues that existing AI literacy frameworks, built for systems that produce outputs for human evaluation, fail to address the principal-agent relationship created by agentic AI, where users delegate decision-making authority to systems whose actions may not be observable, reversible, or controllable. The debt is framed as structural rather than temporary, compounding through three channels (normalization of opaque delegation, multi-agent ecosystem complexity, and institutional path dependence), and is incurred by deploying organizations but paid by users, patients, and citizens. The paper draws on real-world examples (OpenClaw prompt injection, EchoLeak CVE, FTC fraud data) and prior conceptual work (Ladson-Billings, Petrozzino, Floridi) to motivate a reframing of AI literacy as a governance capability.

Significance. The paper identifies a genuine gap in the AI literacy literature: the transition from evaluative to delegated interactions with AI systems is not well captured by existing frameworks, and naming this gap has conceptual value for the field. The framing as 'debt'—with the incurred-by/paid-by asymmetry—is a productive extension of Ladson-Billings and Petrozzino. The six proposed principal-side competencies (p.6) are a concrete, falsifiable contribution that future empirical work could operationalize. The paper is honest about its evidence being illustrative rather than systematic (p.4), which is appropriate for a conceptual essay. The policy hook (EU AI Act Article 4) is well-timed. However, the central claim of structural permanence has an internal tension that needs resolution before the contribution is fully secured.

major comments (1)
  1. p.5, 'Why the Gap Is Structural, Not Temporary': The paper's headline claim is that the deployment-literacy gap is a 'permanent condition' (p.5), and the argument for permanence rests on a single premise: curriculum cycles take 5–7 years while agentic AI evolves in months. However, p.7 acknowledges that 'the technology creating the debt is capable of helping close it, whether through transparency-by-design that embeds literacy into agent interactions, AI tutoring that simulates agentic scenarios at scale, or contextual micro-learning at the point of risk.' If technology-embedded literacy delivery can operate at technological speed, the pace mismatch is not structurally permanent—it is contingent on whether deploying organizations prioritize literacy as a design objective. The paper's response (that such features 'do not emerge organically from product roadmaps optimized for task完成') conb
minor comments (7)
  1. Reference [21] (companion paper, p.6) is listed as arXiv:XXXX.XXXXX, a placeholder. This must be resolved before publication.
  2. p.2: The OpenClaw reference [1] is dated 2026 and cited as a CrowdStrike blog post. The author should verify this source is publicly available and accurately described, as the details (150,000 GitHub stars, walletdrain attacks) are load-bearing for the motivating example.
  3. p.4: The healthcare market figure ($538M in 2024, 45.56% CAGR) is cited to a Grand View Research report. The precision of the CAGR figure (45.56%) implies a specificity that may not be warranted for a market forecast; consider rounding or noting the uncertainty.
  4. p.4: The fraud loss figures ($12.5B FTC, $16.6B FBI IC3) are for all fraud, not agentic AI specifically. The paper acknowledges this ('These figures cover all fraud, not agentic AI specifically'), but the framing could be tightened to make clearer that the figures establish the scale of the adjacent problem, not the target problem.
  5. p.3: The three compounding channels (normalization, ecosystem complexity, institutional path dependence) are asserted rather than argued. For instance, the claim that 'permission grants in production agentic systems are typically inherited across sessions and rarely revoked' is stated without citation. If this is an empirical observation, a source would strengthen it; if it is a conceptual claim, it should be flagged as such.
  6. p.6: The six proposed competencies (delegation, oversight, accountability attribution, attack surface awareness, agent-specific informed consent, calibrated trust) overlap with concepts in Feng et al. [15] and Kasirzadeh and Gabriel [16]. The paper should clarify what is novel beyond these works at the competency-specification level, not just at the claim of novelty.
  7. p.5: The statement that 'none include delegation, oversight, or accountability attribution for autonomous agent action' regarding UNESCO, MAILS, and the AI Literacy Heptagon is a strong negative claim. A brief table or appendix showing the specific competency domains of each framework would make this verifiable rather than assertoric.

Simulated Author's Rebuttal

1 responses · 0 unresolved

The referee identifies a genuine internal tension between the paper's claim of structural permanence (p.5) and its acknowledgment that technology-embedded literacy could close the gap at technological speed (p.7). We agree the language needs sharpening and will revise.

read point-by-point responses

  1. Referee: p.5, 'Why the Gap Is Structural, Not Temporary': The paper's headline claim is that the deployment-literacy gap is a 'permanent condition' (p.5), and the argument for permanence rests on a single premise: curriculum cycles take 5–7 years while agentic AI evolves in months. However, p.7 acknowledges that 'the technology creating the debt is capable of helping close it, whether through transparency-by-design that embeds literacy into agent interactions, AI tutoring that simulates agentic scenarios at scale, or contextual micro-learning at the point of risk.' If technology-embedded literacy delivery can operate at technological speed, the pace mismatch is not structurally permanent—it is contingent on whether deploying organizations prioritize literacy as a design objective. The paper's response (that such features 'do not emerge organically from product roadmaps optimized for task完成') [cut

    Authors: The referee is correct that there is an internal tension between the claim of permanence on p.5 and the concession on p.7 that technology-embedded literacy could operate at technological speed. We accept this criticism and will revise the manuscript to resolve it. The revision will make the following distinction explicit: the debt is not permanent in a logical or physical sense — it is structurally permanent under current institutional and market incentives, meaning that the default trajectory of deployment without deliberate literacy redesign produces a gap that does not self-correct. The pace mismatch between curriculum cycles and product cycles is one mechanism producing this default trajectory, but it is not the only one, and we agree it should not bear the full weight of the permanence claim. The deeper structural argument is that deploying organizations incur the debt but do not pay it (the incurred-by/paid-by asymmetry established on p.3), so the market incentives that would naturally close the gap — namely, that the party creating the debt bears its cost — are absent. Technology-embedded literacy delivery is technically feasible, as p.7 acknowledges, but its adoption requires treating user literacy as a first-class design objective, which current product incentives do not reward. This is a contingent but structurally reinforced condition, not a law of nature. We will revise p.5 to replace 'permanent condition' with language such as 'structurally self-reinforcing under current incentive structures' and will add a paragraph clarifying that the claim of structural permanence is conditional on the absence of intervention, not absolute. We will also strengthen the p.7 passage to make clear that the feasibility of technology-embedded literacy is precisely what makes a revision: no

Circularity Check

0 steps flagged

No circularity: conceptual framing paper with no fitted parameters, equations, or self-citation chains that reduce to inputs.

full rationale

This is a conceptual/position paper that introduces the term 'agentic literacy debt' and argues for its structural significance. There are no equations, no fitted parameters, no quantitative predictions, and no mathematical derivation chain to inspect. The paper's argument builds on external references (Ladson-Billings [3], Petrozzino [4], Floridi [19,20], OWASP [10], FTC/FBI data [8], etc.) that are independent of the author. The only self-citation is to a companion paper [21], which is referenced once as forthcoming and is explicitly stated to cover 'the full specification, including proficiency levels and design imperatives' — i.e., it is not invoked to support the central claim of the present paper. The conceptual extension from 'education debt' (Ladson-Billings) and 'ethical debt' (Petrozzino) to 'agentic literacy debt' is acknowledged and attributed, not presented as a derivation. The skeptic's concern about the tension between 'permanent condition' (p.5) and 'technology can help close it' (p.7) is an internal consistency or correctness issue, not a circularity issue — the paper does not define a quantity in terms of itself, fit a parameter and rename it as a prediction, or invoke a self-citation as a load-bearing mathematical fact. No circularity is present.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 1 invented entities

The paper is a conceptual essay with no free parameters or fitted values. It introduces one invented conceptual entity (agentic literacy debt) and relies on four axioms, two of which are ad hoc to the paper (the permanence of the pace mismatch and the compounding mechanism). The axioms are stated explicitly in the text but are supported by analogy rather than by systematic evidence.

axioms (4)
  • domain assumption Existing AI literacy frameworks assume the user evaluates AI outputs and decides whether to act (citing Long and Magerko 2020 and successors).
    This is the foundational premise of the paper's argument, invoked in the opening paragraph (p. 1) and elaborated on p. 2 ('Existing AI literacy rests on three assumptions that agentic AI invalidates').
  • ad hoc to paper The pace of institutional curriculum reform (5-7 years) is permanently mismatched with the pace of agentic AI product cycles (months).
    Invoked in 'Why the Gap Is Structural, Not Temporary' (p. 5) to support the claim that the debt is permanent rather than a temporary lag. No evidence is provided that the mismatch is permanent rather than contingent on current institutional structures.
  • ad hoc to paper The three compounding channels (normalization, ecosystem complexity, institutional path dependence) produce a widening deficit rather than an additive gap.
    Invoked on pp. 2-3 to distinguish agentic literacy debt from a simple lag. The compounding claim is asserted through analogy to technical debt rather than demonstrated with a model or data.
  • domain assumption Evidence from adjacent domains (clinical AI trust, general fraud, digital divide) is illustrative of the trajectory of agentic AI literacy debt.
    Invoked on p. 4 where the paper states 'the sections that follow draw on evidence from adjacent domains to illustrate the trajectory rather than to document the endpoint.'
invented entities (1)
  • Agentic literacy debt no independent evidence
    purpose: Names the accumulating societal deficit from deploying autonomous AI agents without literacy infrastructure.
    The concept is introduced as a framing device. The paper provides illustrative evidence but no falsifiable test or measurement that would independently validate the concept as a distinct phenomenon from general AI literacy gaps.

pith-pipeline@v1.1.0-glm · 9133 in / 2925 out tokens · 311258 ms · 2026-07-05T07:26:36.962334+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages

  1. [1]

    walletdrain

    Page 1 of 9 Agentic Literacy Debt: A Structural Problem the AI Literacy Field Has Not Yet Named Rohith Nama Abstract Autonomous AI agents now plan, decide, and act on behalf of users across healthcare, financial services, and workplace contexts, often without step-by-step human approval. Existing AI literacy frameworks were built for a world in which huma...

    work page 2020

  2. [2]

    achievement gaps

    reframed educational disparities in the United States by arguing that annual “achievement gaps” are merely snapshots of a cumulative “education debt” with historical, economic, sociopolitical, and moral dimensions. Her insight, that focusing on annual gaps obscures the structural, compounding nature of the problem, applies directly to AI literacy. Petrozz...

    work page 2024

  3. [3]

    sufficient level of AI literacy

    construct four-dimensional agentic profiles for proportional governance. Both describe governance challenges but not the literacy infrastructure required for populations to exercise the governance roles they prescribe. Part of the challenge is also architectural. Every production agentic system generates detailed action logs, but these are designed for de...

    work page 2025

  4. [4]

    CrowdStrike Blog (2026)

    CrowdStrike: What security teams need to know about OpenClaw, the AI super agent. CrowdStrike Blog (2026). https://www.crowdstrike.com/en-us/blog/what-security-teams-need-to-know-about-openclaw-ai-super-agent/

    work page 2026

  5. [5]

    arXiv:2509.10540 (2025)

    Reddy, K., et al.: EchoLeak: the first real-world zero-click prompt injection exploit in a production LLM system. arXiv:2509.10540 (2025)

    work page arXiv 2025

  6. [6]

    Ladson-Billings, G.: From the achievement gap to the education debt: understanding achievement in U.S. schools. Educ. Res. 35(7), 3–12 (2006). https://doi.org/10.3102/0013189X035007003

    work page doi:10.3102/0013189x035007003 2006

  7. [7]

    https://doi.org/10.1007/s43681-020-00030-3

    Petrozzino, C.: Who pays for ethical debt in AI? AI Ethics 1, 205–208 (2021). https://doi.org/10.1007/s43681-020-00030-3

    work page doi:10.1007/s43681-020-00030-3 2021

  8. [8]

    Grand View Research (2025)

    Grand View Research: Agentic AI in healthcare market size, share & trends analysis report, 2025–2030. Grand View Research (2025). https://www.grandviewresearch.com/industry-analysis/agentic-ai-healthcare-market-report

    work page 2025

  9. [9]

    JMIR AI 3, e53207 (2024)

    Rosenbacke, R., Melhus, Å., McKee, M., Stuckler, D.: How explainable artificial intelligence can increase or decrease clinicians’ trust in AI applications in health care: systematic review. JMIR AI 3, e53207 (2024). https://doi.org/10.2196/53207

    work page doi:10.2196/53207 2024

  10. [10]

    NEJM AI (2024)

    MIT Media Lab: People overtrust AI-generated medical advice despite low accuracy. NEJM AI (2024). https://www.media.mit.edu/publications/NEJM-AI-people-overtrust-ai-generated-medical-advice-despite-low-accuracy/

    work page 2024

  11. [11]

    (2025); Federal Bureau of Investigation: Internet Crime Report

    FTC, Washington, D.C. (2025); Federal Bureau of Investigation: Internet Crime Report

    work page 2025

  12. [12]

    IC3, Washington, D.C. (2025)

    work page 2025

  13. [13]

    Deloitte Center for Financial Services (2024)

    Lalchand, S., Srinivas, V., Maggiore, B., Henderson, J.: Generative AI is expected to magnify the risk of deepfakes and other fraud in banking. Deloitte Center for Financial Services (2024). https://www.deloitte.com/us/en/insights/industry/financial-services/deepfake-banking-fraud-risk-on-the-rise.html

    work page 2024

  14. [14]

    Open Web Application Security Project (2025) Page 9 of 9

    work page 2025

  15. [15]

    https://www.gartner.com/en/newsroom/press-releases/2025-09-22-gartner-survey-reveals-generative-artificial-intelligence-attacks-are-on-the-rise

    work page 2025

  16. [16]

    https://www.itu.int/itu-d/reports/statistics/facts-figures-2024/

    ITU, Geneva (2024). https://www.itu.int/itu-d/reports/statistics/facts-figures-2024/

    work page 2024

  17. [17]

    Lintner, T.: A systematic review of AI literacy scales. npj Sci. Learn. 9, 50 (2024). https://doi.org/10.1038/s41539-024-00264-4

    work page doi:10.1038/s41539-024-00264-4 2024

  18. [18]

    arXiv:2506.12469

    Feng, K.J.K., McDonald, D.W., Zhang, A.X.: Levels of autonomy for AI agents. arXiv:2506.12469. Knight First Amendment Institute, Columbia University (2025)

    work page arXiv 2025

  19. [19]

    arXiv:2504.21848 (2025)

    Kasirzadeh, A., Gabriel, I.: Characterizing AI agents for alignment and governance. arXiv:2504.21848 (2025)

    work page arXiv 2025

  20. [20]

    Harvard University Press, Cambridge (2015); Selbst, A.D., et al.: Fairness and abstraction in sociotechnical systems

    Pasquale, F.: The Black Box Society: The Secret Algorithms That Control Money and Information. Harvard University Press, Cambridge (2015); Selbst, A.D., et al.: Fairness and abstraction in sociotechnical systems. In: Proceedings of the ACM Conference on Fairness, Accountability, and Transparency, pp. 59–68. ACM, New York (2019). https://doi.org/10.1145/32...

    work page doi:10.1145/3287560.3287598 2015

  21. [21]

    Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), Art

    work page 2024

  22. [22]

    Floridi, L., Cowls, J.: A unified framework of five principles for AI in society. Harv. Data Sci. Rev. 1(1) (2019). https://hdsr.mitpress.mit.edu/pub/l0jsh9d1

    work page 2019

  23. [23]

    Floridi, L.: AI as agency without intelligence: on ChatGPT, large language models, and other generative models. Philos. Technol. 36, 15 (2023). https://doi.org/10.1007/s13347-023-00621-y

    work page doi:10.1007/s13347-023-00621-y 2023

  24. [24]

    arXiv:XXXX.XXXXX (2026)

    Nama, R.: From Evaluator to Principal: The Agentic AI Literacy Framework (AALF) for Delegated Autonomy. arXiv:XXXX.XXXXX (2026)

    work page 2026