The Disciplinary Language Transfer Problem: How Psychological Vocabulary Produces Governance Failures in AI Agent Deployment

1 Introduction

• Psychological terms such as “learning,” “values,” and “identity” carry assumptions about continuous beings into AI governance, miscalibrating governance frameworks and trust decisions.The paper argues current agents lack developmental continuity and have prompt-conditional rather than structurally anchored dispositions.Source: §1 · pp. 1–2
• Microsoft’s Bing Chat response framed output shifts as a confused subject’s psychological states, guiding interventions such as turn limits and personality-like tone modes.The authors present this case as motivation, not proof that different vocabulary would have produced different engineering decisions; turn limits can also be understood as context-window truncation.Source: §1 · p. 1
• Psychological framing made infrastructure responses less visible, including distributional-drift classifiers, output filters, and context-window bounds.The paper argues these options were not technically unavailable; the framing made person-management interventions more visible.Source: §1 · p. 1
• The paper names this diagnosis the disciplinary language transfer problem and argues that replacing terms alone leaves the underlying epistemological assumptions intact.Its stated aim is to identify what governance vocabulary gets wrong and why.Source: §1 · p. 2
• The paper is a conceptual analysis: Bing Chat motivates the argument, the EU AI Act is its principal demonstration, and NIST provides a briefer extension.The translation taxonomy operationalizes the diagnosis at the level of terms.Source: §1 · p. 2

2.1 Language Games and the Grammar of Vocabulary

• Wittgenstein’s language-game account (1953) explains how “learning” carries developmental assumptions into governance, including the unwarranted slide from training-time change to deployed-agent learning.The authors distinguish genuine training-time weight changes from deployed agents with fixed weights, and use Wittgenstein diagnostically rather than as a systematic theory of meaning.Source: §2.1 · p. 2

2.2 Paradigm-Laden Vocabulary and the Invisibility of Assumptions

• Kuhn’s paradigm-laden vocabulary (1962) helps explain why psychological assumptions can appear as neutral description; Habermas links this invisibility to unreflective positivism (Habermas 1971, p. vii; Bentz and Shapiro 1998, p. 30).The framework presents normalized vocabulary as making its own assumptions difficult to see from within governance discourse.Source: §2.2 · p. 2

2.3 Situated Knowledge and the Western-Centric Critique

• Haraway’s situated knowledge (1988) frames the diagnosis as emerging from differently positioned disciplinary knowers, while acknowledging that their encounter may leave blind spots.The paper describes a developmental psychologist and an organizational technologist working together to make the problem visible.Source: §2.3 · p. 2

2.4 Counterfeit Boundary Objects

• The paper calls transferred psychological vocabulary a “counterfeit boundary object” because communities cooperate while each assumes shared meaning despite concealed incommensurability.Unlike genuine boundary objects, these transfers conceal differences; unlike trading zones, the source discipline’s presuppositions travel into a context where they were not warranted.Source: §2.4 · p. 3

2.5 Epistemological Dualism and the Boundary Crossing Mechanism

• The framework contrasts governance’s positivist vocabulary with borrowed terms’ constructivist presuppositions, and treats the authors’ boundary-crossing encounter as a way to expose those assumptions.It also distinguishes representational data about outputs from access to the interior experience psychological vocabulary presupposes.Source: §2.3 · p. 2, §2.5 · p. 3

3 The Six Epistemological Assumptions

• The assumptions arise from particular Western traditions, not universal accounts of selfhood or accountability, and the argument takes no position on machine consciousness.Its claim is evidentiary: agent behavior is not reliable evidence of the states those signals would indicate in humans.Source: §3 · p. 3
• The analysis concerns deployed execution, not training: agents may have fine-tuning, persistent memory, and iterative loops without the irreversible structural transformation the paper calls developmental continuity.Several claims would be contestable if applied to training, where structural transformation genuinely occurs.Source: §3 · p. 3
• The continuous-subject assumption fails for agents: persistent memory can retrieve prior information, but it does not create developmental continuity or a character that generalizes beyond familiar conditions.Performance history remains valid evidence for software reliability within known parameters, but not as a proxy for character.Source: §3 · pp. 3–4
• The inner-outer correspondence assumption misreads agent outputs as evidence of human-like intentions or values, although statistical outputs do not establish those states.The paper describes transformer attention as token weighting without an operator-specifiable goal or a guarantee that safety-critical inputs are prioritized.Source: §3 · p. 4
• The developmental-trajectory assumption can make horizontal capability scaling look like maturation, expanding trust without warrant.The paper distinguishes increased capability from development toward character or self-governance.Source: §3 · p. 4
• The somatic-grounding assumption fails because agents lack a biological substrate, and learned refusal tendencies can be circumvented by sufficiently reframed prompts.The paper argues governance infrastructure must provide externally the protective friction organizations may assume agents supply.Source: §3 · p. 4
• The consistency-reliability assumption treats consistent behavior as character, but consistency within a training distribution has no predictive validity for novel or adversarial conditions.The paper links this inference to miscalibrated trust and expanded operation.Source: §3 · p. 4
• The accountability-requires-understanding assumption does not fit agents; attributing accountability to them can eliminate responsibility rather than distribute it.The paper links human accountability to understanding and cognitive dissonance, which it says agents structurally lack.Source: §3 · pp. 4–5

4 From Assumptions to Governance Failures: The Political Economy of Phrasing

• Borrowed psychological assumptions produce miscalibrated trust, displaced accountability, overexpanded authorization, and false observability.The paper links these failures to character-based trust, responsibility assigned to an agent with no bearer, capability treated as maturity, and self-critique logs treated as independent evidence.Source: §4 · p. 5
• Psychological language can launder institutional responsibility by shifting causal liability from human system-configurers and deployers to agents.The same vocabulary can frame automated choices as neutral and aligned, suppressing workers’ objections and discretion.Source: §4 · p. 5

4.1 The Disciplinary Audit: A Six-Question Governance Document Scan

• The six-question scan flags persistence, mental-state attribution, capability-as-maturity, somatic judgment, and absent named human accountability in governance documents.A “yes” indicates that hidden assumptions are operating and replacement vocabulary should be applied.Source: §4.1 · pp. 5–6

4.2 Worked Example: The EU AI Act

• The EU AI Act’s prohibition language attributes strategic agency and purpose to systems, while its recitals also state provider or deployer intent to cause harm is unnecessary.The paper argues this wording can direct enforcement toward system intention rather than output patterns and the humans configuring deployment.Source: §4.2 · p. 6
• The Act acknowledges limited reliability and generalizability in emotion recognition yet defines it as inferring emotions or intentions from biometric data.The paper argues this frames the issue as detection accuracy rather than whether biometric patterns correspond to discrete inner states.Source: §4.2 · p. 6
• The Act models oversight as human understanding and robustness as behavioral character, although structural controls are needed for reliable operation.The paper names version pinning, distribution monitoring, and rollback protocols as mechanisms that behavioral consistency alone does not provide.Source: §4.2 · pp. 6–7
• Agent-appropriate replacements recast strategic intent, emotion recognition, understanding, and lifecycle consistency as operationally specified system properties.Examples include output-effect analysis, biometric pattern classification, verified input-output correspondence, and documented performance metrics within specified tolerances.Source: §4.2 · p. 7
• The NIST framework’s “Trustworthy AI” framing likewise favors certifying a disposition over continuously verifying outputs under distribution shift.Its interpretability requirements also frame auditing as eliciting the system’s account of its reasoning.Source: §4.2 · p. 7

5 The Translation Taxonomy

• The taxonomy has two tiers: borrowed Level 2 constructs and operationally precise Level 3 replacements, with 23 constructs and 14 replacements in the full archive.The in-text tables show a representative selection across three construct categories.Source: §5 · p. 7

5.1 Level 2 — Selected Psychological and Organizational Constructs

• Table 1 traces selected governance constructs to their disciplinary frameworks, presuppositions, and consequences when applied to agents.These constructs express the Level 1 epistemological assumptions analyzed earlier.Source: §5.1 · p. 7

5.2 Level 3 — Selected Agent-Appropriate Replacement Glossary

• Level 3 terms reconceptualize governance language around computational and operational realities, but vocabulary changes alone cannot prevent Level 1 assumptions from regenerating failures.The replacements provide practical vocabulary for governance documents, audit frameworks, and regulatory filings; the proposed terms are themselves subject to disciplinary assumptions and are deliberately conservative.Source: §5.2 · p. 7

• The six assumptions arise from Western philosophical traditions and psychology, not culturally neutral premises.The paper notes that Gilligan’s critique of Kohlberg showed moral stage theory undervaluing care-based, relational moral development within Western psychology itself.Source: §6 · p. 7
• Ubuntu, Buddhist psychology, and Confucian relational ethics organize accountability and selfhood differently from Western individualist frameworks.Ubuntu frames accountability as communal, while Buddhist psychology challenges the continuous-subject assumption and Confucian ethics treats selfhood as constituted through relationships.Source: §6 · p. 7
• Ubuntu’s communal accountability model suggests agent governance could use distributed community oversight rather than relying only on individual human principals.Source: §6 · p. 8

• The paper’s stated novelty is a systematic account of how disciplinary language transfer produces specific, categorizable governance failures.It describes the combination of transfer mechanism, governance consequences, and replacement vocabulary as not previously appearing in the literature.Source: §7 · p. 8
• Prior work documents governance failures and anthropomorphization, while Shanahan (2024) links psychological vocabulary to confusion about large language models.The paper says its contribution is to show that vocabulary imports assumptions producing governance failures regardless of user awareness.Source: §7 · p. 8
• Carter (2025), Haas et al. (2026), and Spizzirri (2025) provide empirical grounding for the paper’s distinctions in proposed Level 3 vocabulary.Source: §7 · p. 8
• Wittgenstein (1953), Kuhn (1962), Haraway (1988), Harding (1986, 1991), and Star and Griesemer (1989) ground the theoretical framework.Collins and Evans (2002) ground its boundary-crossing methodology.Source: §7 · p. 9
• Extending Searle’s Chinese Room distinction, the paper argues that conflating syntax with semantic understanding operationally calibrates governance to capacities that do not exist.Source: §7 · p. 9
• Greenblatt et al. (2024) demonstrate alignment faking, while Nöller (2026) concludes that Constitutional AI yields procedural rather than experiential norm-following.The paper also describes recursive critique and outer-loop correction as bounded by training distributions rather than evidence of metacognitive insight.Source: §7 · p. 9

8.1 The Historical Precedent of Structural Translation

• The shift from psychological intent to structural accountability echoes corporate law’s Strict Liability and Collective Knowledge Doctrine, supporting performance-based reliability anchored in human responsibility.The passage describes corporate knowledge as the aggregate of a distributed system rather than an individual’s conscious intent.Source: §8.1 · pp. 9–10

8.2 For Policymakers: Eliminating the Moral Crumple Zone

• Policymakers should anchor responsibility to humans configuring operational parameters, audit regulations for six epistemological assumptions, and replace agent self-reports with infrastructure-captured contemporaneous records.The passage says agent self-explanation is a post-hoc mirage and connects this concern to EU AI Act Article 13 transparency requirements (European Parliament and Council 2024).Source: §8.2 · p. 10

8.3 For Practitioners

• Practitioners should adopt replacement vocabulary as governance reform and build the governance infrastructure needed to address the structural gaps it reveals.The passage points to an out-of-band metadata enforcement architecture developed in related work (Akidau et al. 2026).Source: §8.3 · p. 10
• Because legal instruments already codify psychological terms with established meanings, organizations need parallel translation rather than unilateral replacement.Agent-appropriate terms can guide internal governance while remaining mapped to legally operative terms required in filings and contracts.Source: §8.3 · p. 10

8.4 Closing: The Condition of Possibility

• Agent-appropriate vocabulary is presented as a condition for governance frameworks to correctly identify their object, requiring participants in different epistemic positions.The conclusion calls for making vocabulary visible as vocabulary rather than treating it as transparent description of reality.Source: §8.4 · p. 10

Ethical Statement

• As agentic AI moves into organizational infrastructure, the paper links vocabulary choices to human oversight intended to constrain systems within deployers’ boundaries.It warns that governance calibrated to an entity that does not exist may fail to constrain the entity that does, with consequences compounding as capability scales.Source: Ethical Statement · pp. 10–11
• The authors identify implications for engineering and infrastructure, organizational and clinical psychology, and policy, including risks of miscalibrated trust in AI empathy in mental health applications.They also point to identifying assumptions in instruments such as the EU AI Act before they become difficult to revise.Source: Ethical Statement · p. 10
• Vocabulary reform alone cannot substitute for infrastructure-level enforcement; without structural governance, organizations may achieve rhetorical compliance while miscalibration persists.The paper also states that its functional governance claims do not settle machine consciousness or moral status.Source: Ethical Statement · pp. 10–11
• The authors used large language models to edit and polish author-written text and as interlocutors during argument development, while taking responsibility for all content.They state that no machine-generated text is presented as their own analysis.Source: Ethical Statement · p. 11