Epistemic Debt in AI and Scientific Software

• Epistemic debt is the cumulative gap in understanding and verification that undermines the trustworthiness of AI outputs and scientific claims.
• It differs from traditional technical debt by emphasizing deficits in cognitive ownership, reproducibility, and interpretability rather than just code quality.
• Empirical studies and formal models reveal that unchecked epistemic debt leads to operational fragility and increased user verification burdens, necessitating proactive, multi-level mitigation.

Epistemic debt denotes accumulated deficits in understanding, verification, evidential grounding, or scientific fidelity that undermine the trustworthiness of knowledge claims produced with software and AI. Recent work operationalizes the concept across several loci rather than through a single canonical definition: in scientific software it appears as scientifically consequential shortcuts that threaten validity, reproducibility, interpretability, and uncertainty management; in GenAI-assisted software engineering it appears as a widening gap between produced code and collective understanding; in AI deployment it appears as user labor forced by unreliable or locally irrelevant outputs; and in formal theory it appears as a stock of unverified reasoning obligations that compounds under substitutive AI use (Melin et al., 15 Jan 2026, Ahmad, 14 Apr 2026, Oyemike et al., 12 Nov 2025, Meng, 13 Jun 2026).

1. Conceptual scope and neighboring constructs

Current usage distinguishes epistemic debt from conventional technical debt primarily by locus and consequence. Traditional technical debt resides in code artifacts and architectural decisions, and is commonly framed around maintainability and future engineering cost. By contrast, the recent literature places epistemic debt in the degradation of knowledge quality, verification capacity, or cognitive ownership. In scientific software, it is the broader category under which scientific debt is treated as a concrete instantiation. In GenAI-assisted development, comprehension debt is described as socio-cognitive and located in team cognition and shared mental models rather than the codebase. In AI-use studies, epistemic debt is a user-side burden triggered when outputs are unreliable, locally irrelevant, or insufficiently evidenced. In formal modeling, epistemic debt is the aggregate counterpart of cognitive debt, namely the stock of unverified reasoning obligations embedded in workflows, documents, models, and decisions (Melin et al., 15 Jan 2026, Ahmad, 14 Apr 2026, Oyemike et al., 12 Nov 2025, Meng, 13 Jun 2026).

Formulation	Locus	Primary consequence
Scientific debt	Scientific code and development artifacts	Threats to validity, reproducibility, interpretability, uncertainty management
Comprehension debt	Team cognition and shared mental models	Inability to maintain, modify, explain, or debug effectively
Epistemic debt in alignment debt	User-side interaction with AI outputs	Verification labor, cross-checking, delayed trust
Cognitive debt	Individual and aggregate reasoning capacity	Systemic fragility under AI substitution

These formulations are not identical, but they converge on a common structure: present-time acceleration is purchased by deferring the work required to justify, reproduce, verify, or interpret future outputs. A plausible implication is that epistemic debt is best understood as a family of debt constructs tied together by a shared failure mode: output production outruns warranted understanding.

2. Epistemic debt in scientific and research software

In scientific and research software, epistemic debt is most explicitly operationalized as scientific debt. One study defines scientific debt as “the accumulation of suboptimal scientific practices, assumptions, inaccuracies, or outdated knowledge in scientific software that potentially compromise the validity, accuracy, and reliability of scientific results,” and analyzes it across 900,358 artifacts from 23 open-source scientific software projects spanning code comments, commit messages, issue tracker sections, and pull request sections. It identifies five recurring forms: translation challenges, assumptions, missing edge cases, computational accuracy, and new scientific findings. These forms are directly mapped to epistemic qualities: validity, reproducibility, interpretability, and uncertainty management. Representative examples include “not correct, but … a standard assumption,” “We assume here that new ice arrives at the surface with the same temperature as the surface,” “This does not work for large molecules that span more than half of the box,” and “for large systems, a float may not have enough precision” (Melin et al., 15 Jan 2026).

The empirical distribution of these admissions is concentrated in collaborative artifacts rather than code comments alone. In the 23-project analysis, pull request sections had the highest overall SATD prevalence at 12.00%, while issue tracker sections concentrated scientific debt most strongly at 1.22%; pull requests contained 0.36% scientific debt, code comments 0.17%, and commit messages 0.06%. The same study reports that a baseline BERT trained only on traditional SATD classes predicts 74.4% of scientific-debt instances as non-debt, with only 25.6% mapped to a traditional category. This supports the claim that epistemically consequential scientific compromises are not reducible to general-purpose SATD taxonomies.

A related multi-method study of research software examines 28,680 self-admitted technical debt comments across nine long-lived open-source projects and uses a taxonomy including Architectural, Build, Code, Defect, Design, Test, Requirements, Documentation, Algorithm, On Hold, and Scientific Debt. It reports that Scientific Debt occurs in all projects and is especially prevalent in CESM at 14.43% of SATD, Elmer at 11.16%, and Firedrake at 9.21%. Within Scientific Debt, assumptions were high in GROMACS at 37.82% and CESM at 34.53%; missing edge cases in Firedrake at 38.24% and Elmer at 33.33%; computational accuracy in Astropy at 31.25% and GROMACS at 30.25%; and translation challenges in MOOSE at 30.56% and Athena at 20.33%. The same study ties these patterns to four broader themes: artifacts as boundary objects, science and organizational goals as debt drivers, people as drivers of success, and complexity as a complicating factor (Ernst et al., 20 Mar 2026).

Taken together, these studies frame epistemic debt in scientific computing as a defect in scientific warrant, not merely in software hygiene. When assumptions are implicit, edge cases are unhandled, numerical precision is unstable, or embedded constants are outdated, the debt is paid as compromised scientific validity and reproducibility rather than only as higher maintenance cost.

3. Socio-cognitive epistemic debt in GenAI-assisted development

A second major line of work relocates epistemic debt from artifacts to cognition. “Comprehension Debt in GenAI-Assisted Software Engineering Projects” defines comprehension debt as “the growing gap between what a development team knows about its codebase and what it actually needs to understand in order to maintain and modify it effectively,” and further as “the cumulative gap between the demands a codebase makes on its developers and [the] collective understanding those developers possess.” In an eight-week undergraduate software engineering course at Karlstad University, based on 621 reflective diaries from 207 students, the study identifies four accumulation patterns: AI-as-black-box code acceptance, context-mismatch debt, dependency-induced atrophy, and verification-bypass. It also identifies a mitigating pattern in which GenAI acts as a comprehension scaffold through explanation-first use, rewriting before commit, and verification against documentation and tests (Ahmad, 14 Apr 2026).

A more interventionist study defines epistemic debt as “the gap between the functional software artifacts a developer can produce with AI assistance and the developer’s cognitive ownership of how those artifacts work,” and states that “Epistemic Debt describes the accumulation of functional software artifacts that the user ‘owns’ legally but does not ‘own’ cognitively.” In a between-subjects experiment with $N = 78$ participants, three conditions were compared: Manual, Unrestricted AI, and Scaffolded AI. Functional utility differed strongly across groups, with one-way ANOVA reporting $F(2, 75) = 48.2$, $p < .001$, $\eta^2 = 0.56$; Unrestricted AI achieved mean utility of 92.4% and Scaffolded AI 89.1%, versus 65.2% for Manual. But on a subsequent AI-blackout maintenance task, corrective competence diverged sharply: success rates were 69.2% for Manual, 61.5% for Scaffolded AI, and 23.1% for Unrestricted AI, with $\chi^2(2) = 13.8$, $p = .001$, $V = 0.42$. The paper characterizes this as a “Collapse of Competence” and describes the unrestricted pattern as producing “Fragile Experts” with high functional utility but critically low corrective competence (Sankaranarayanan, 22 Feb 2026).

Both studies insist that this debt cannot be reduced to static code quality. A system can be technically sound while being cognitively opaque to its developers. The central mechanism is consistent across the qualitative and experimental evidence: GenAI accelerates production, but if it bypasses schema formation, explanation, or verification, the resulting artifact outpaces the team’s capacity to reason about it when modification or debugging is required.

4. User-side epistemic debt, alignment, and fairness

A third formulation treats epistemic debt as a burden borne by AI users rather than developers. “Alignment Debt: The Hidden Work of Making AI Usable” defines alignment debt as “the user-side burden that accumulates when AI systems fail to align with the cultural, linguistic, infrastructural, epistemic, or interactional conditions in which they are used.” Within that four-part taxonomy, epistemic debt is present “when outputs are unreliable, locally irrelevant, or insufficiently evidenced,” and the paper states that “Epistemic debt therefore converts informational uncertainty into user labour.” Operationally, it is coded present if respondents report that AI “Gives wrong answers; Shares misinformation; Doesn’t list its sources or share thought process” (Oyemike et al., 12 Nov 2025).

The empirical study surveys 411 AI users in Kenya and Nigeria, with taxonomy analyses conducted on a measurable subsample of $n = 385$. In that subsample, epistemic debt affects 33.8% ($n = 130$; 95% CI: 29.2–38.6). Item-level indicators highlighted in the study are “doesn’t list its sources” at 19.0%, “gives wrong answers” at 9.2%, and “shares misinformation” at 8.8%. Verification is treated as compensatory labor: overall, 84.6% verify AI outputs using external sources, and among those verifying, the top sources are Google at 64.5%, academic or authoritative databases at 24.1%, and Wikipedia at 13.9%. Users with epistemic debt verify at 91.5%, compared with 80.8% among those without, with $\chi^2 = 6.77$, Holm–Bonferroni corrected $F(2, 75) = 48.2$0, and Cramér’s $F(2, 75) = 48.2$1. Verification intensity also rises with cumulative debt burden, with mean sources increasing from 1.48 for one debt type to 3.50 for four debt types, and Spearman’s $F(2, 75) = 48.2$2, $F(2, 75) = 48.2$3.

This formulation sharpens an important distinction. Cultural and linguistic debt often compels rephrasing; infrastructural debt compels time and bandwidth expenditure; interaction debt compels orchestration and prompt engineering. Epistemic debt, by contrast, compels the work of checking whether an answer is true, evidenced, and locally relevant. The paper’s fairness argument follows directly: model-side accuracy metrics can miss a substantial portion of the burden if they do not measure who must perform this verification labor and under what infrastructural constraints.

5. Formal dynamics: cognitive debt, leverage, and systemic fragility

A formal theory of cognitive debt provides the most explicit dynamic account of epistemic debt accumulation. The model defines cognitive capital $F(2, 75) = 48.2$4 as an agent’s unaided ability to reason, verify, synthesize, and transfer knowledge, and cognitive debt $F(2, 75) = 48.2$5 as the stock of unverified reasoning obligations. It then maps epistemic debt to the aggregate level by defining $F(2, 75) = 48.2$6 and interpreting this stock as organizational epistemic debt. Aggregate leverage is given by $F(2, 75) = 48.2$7, where $F(2, 75) = 48.2$8; higher leverage raises both crisis probability and expected loss (Meng, 13 Jun 2026).

The model’s core state dynamics are

$F(2, 75) = 48.2$9

$p < .001$0

where $p < .001$1 is AI substitution intensity and $p < .001$2 is deliberate practice. Production in the normal state is multiplicative,

$p < .001$3

so cognitive capital functions as collateral for AI adoption: the marginal product of AI is proportional to existing unaided reasoning capacity. Crisis probability is specified as $p < .001$4, increasing and locally convex in leverage $p < .001$5 and increasing in model concentration $p < .001$6.

The paper establishes six propositions. Rational agents incur positive cognitive debt because its costs are deferred, partially external, and masked by short-run productivity gains. During tranquil periods, subjective risk $p < .001$7 falls while true crisis probability $p < .001$8 rises endogenously with leverage, producing a “cognitive Minsky moment.” Expected crisis losses are convex in aggregate leverage. Post-crisis, output pressure can induce a “false-correction loop” in which agents patch AI failures with more AI. The decentralized equilibrium over-adopts substitutive AI relative to the social optimum because of systemic-risk externalities, cognitive public goods, and arms-race externalities. In a heterogeneous-agent economy, initially higher-cognitive-capital agents may adopt AI more intensively and eventually erode their unaided capital below that of initially lower-skilled agents.

This theory makes epistemic debt analytically precise. It is not merely a metaphor for poor understanding; it is a stock variable with issuance, repayment, compounding, leverage, crisis exposure, and macro-level externalities. The paper’s policy analogies follow from that formalization: leverage-indexed AI-use levies, cognitive reserve requirements, verification mandates, stress tests, and concentration policy are presented as direct responses to the dynamics of debt accumulation.

6. Detection, mitigation, and open problems

Empirical work on epistemic debt has already produced concrete detection and management mechanisms. In scientific software, explicit detection is treated as necessary because traditional SATD models miss most scientific debt. A multi-source SATD classifier trained across code comments, commit messages, issue sections, and pull request sections uses a multi-task transformer with artifact-specific classification heads and reports Falcon3-3B-Instruct as the best model, with average accuracy 0.9159 and macro-F1 0.8255; scientific debt itself remains harder, with precision 0.7422, recall 0.5564, and F1 0.6353, partly because it is less than 1% of training data. The same work recommends treating scientific debt as a distinct SATD category, prioritizing issues and pull requests, using artifact-specific heads and domain-aware lexicons, ranking debt by impact on validity, reproducibility, and numerical stability, requiring “Assumptions and Known Limitations” sections in pull requests, adding benchmark and verification tests, quantifying uncertainty, and tracking outdated constants and models (Melin et al., 15 Jan 2026).

For socio-cognitive forms of epistemic debt, the most developed mitigations aim to surface comprehension rather than infer it from code alone. The comprehension-debt study recommends explicit AI literacy, cognitive apprenticeship in verification, scaffold-oriented prompting, rewrite-before-commit as a comprehension gate, structured comprehension review in retrospectives, and active-learning assessments such as oral code walkthroughs and explanation-based grading. The experimental study implements a stronger mechanism: an Explanation Gate that blocks merge or apply actions until an LLM-as-a-Judge rates a learner’s causal explanation at SOLO Level $p < .001$9; the gate operationalizes a Teach-Back protocol and is presented as a way to impose “Metacognitive Friction” precisely where intrinsic reasoning is at stake (Ahmad, 14 Apr 2026, Sankaranarayanan, 22 Feb 2026).

For user-side epistemic debt in deployed AI systems, proposed interventions target both epistemic fit and the feasibility of verification. The alignment-debt framework recommends localizing knowledge access, attaching inline clickable sources to factual claims, exposing confidence bands, explicitly flagging low-confidence answers grounded in sparse data, providing one-click cross-checks to authoritative sources, and reducing the bandwidth cost of verification through low-bandwidth modes, progressive loading, pre-query data-use estimates, offline queueing, and compact on-device models. It also recommends making user burden measurable through verification rates and time, sources consulted per task, and data cost per verified action, and aligning procurement and governance with burden reporting rather than model metrics alone (Oyemike et al., 12 Nov 2025).

The main limitations of the current literature are also explicit. Scientific-debt labeling is subjective; explicit admissions miss unacknowledged issues; integrating multiple datasets introduces label noise and context mismatch; and current scientific-software evidence is concentrated in DOE-supported CASS projects or a small number of long-lived open-source research software projects (Melin et al., 15 Jan 2026, Ernst et al., 20 Mar 2026). Comprehension-debt evidence is drawn from a single-institution student population using reflective diaries, while the experimental epistemic-debt study is restricted to novices and may not generalize to experts because of expertise-reversal effects (Ahmad, 14 Apr 2026, Sankaranarayanan, 22 Feb 2026). Alignment-debt evidence is based on young, highly educated users in two anglophone African countries and uses a profile-based instrument without reported psychometric reliability (Oyemike et al., 12 Nov 2025). The formal cognitive-debt model relies on specific assumptions about misspecified beliefs, debt compounding, and multiplicative production, even though the authors argue that many qualitative results survive broader specifications (Meng, 13 Jun 2026).

Across these strands, a consistent research agenda is already visible. Future work is directed toward project-aware cross-project generalization, more granular subcategories of scientific debt, multimodal artifact fusion, behavioral rather than self-reported measures of comprehension debt, operational indicators such as explanation latency and modification difficulty, validated short-form burden instruments, longitudinal tracking of debt accumulation, and governance mechanisms that tie AI adoption to verification capacity rather than short-run output. This suggests that epistemic debt is becoming a unifying concept for studying how software and AI systems compromise warranted knowledge when production, deployment, or coordination proceeds faster than explanation, verification, and contextual fit.