Owning what one did not make: psychological ownership, agency, and the authorship legitimation deficit in human–AI co-creation

Synthesis: Hong and colleagues argue that human–AI co-creation produces a problem that is neither lost productivity nor lost feeling. A person can be named author of a work they did not produce, and of a work they do not experience as their own. Their construct, the authorship legitimation deficit, is the mismatch between a particular authorship designation and the experiential or contribution-based grounds that role requires under existing criteria. They call the contingent alignment of making, owning, and being answerable the labor nexus of authorship, and decompose it into separate routes to psychological ownership, the sense of agency, and assigned answerability. The contribution is a role-relative classification: one person with one experience can hold a supported editorial role and an unsupported sole-creator claim at once.

This page summarizes a conceptual article, published under the journal's "Conceptual Analysis" type. It collects no data and tests no model of its own: the propositions P1–P6 and conditions B1–B6 below are the authors' proposals, and every figure or empirical result mentioned is attributed to the study that produced it rather than reported here as a finding.

What the analysis argues

1. Three relations can align in being an author, and they are not the same relation. Psychological ownership is the felt state of a target as "mine," grown through control, intimate knowing, and investment of self. The sense of agency is the experience of authoring one's actions and, through them, their effects. Responsibility covers attributability, answerability, and accountability, which are distinct questions. Naming their possible alignment the labor nexus flags a contingent configuration, not a second construct.
2. P1 and P2: delegation changes the routes, not necessarily the person. Holding task and target quality comparable, greater delegation of externally specified production steps is predicted to reduce initial-production effort (P1). Meaningful engagement is predicted to support ownership (P2a), shaping control to support task-specific agency (P2b), and knowledge plus release authority to support observers' judgments of reasons owed and, under harm and avoidability conditions, blame (P2c). P2d states only that these routes can align, without required covariance.
3. P3 is the diagnosis, not an effect. A deficit exists where an applicable ground falls short of its prior criterion for the specified role. The authors insist that this is analytic: it is not another causal consequence of weakened ownership, and it cannot be obtained by subtracting a designation from an ownership score.
4. P4 and P6 are the relational hypotheses. A greater role–ground mismatch is predicted to produce more refusal or qualification of the offered designation among people who understand and accept the criteria and have low-cost alternatives (P4). Recurrent, documented mismatches are predicted to precede community recognition of alternative grounds such as direction, selection, substantive editing, and verification (P6).
5. P5a and P5b belong to neighboring literatures, not to the deficit. Experienced agency is predicted to track felt answerability while release duties are held fixed (P5a). For the same approving person, replacing a human delegate with an AI delegate is predicted to reduce creative credit for good outputs more than blame for harmful ones (P5b). The authors state that a deficit diagnosis is not required for either.

Why the field needs a role-relative account

Ownership theory explains mineness, agency research explains experienced influence, and responsibility research separates grounds of answerability. None compares a specified role against the grounds that role requires.

Draxler and colleagues documented a "ghostwriter effect": users of personalized text generation reported little ownership while continuing to declare themselves authors. That dissociation is not itself a deficit. If an editorial role is supported by what the person actually did, and carries no requirement of strong felt ownership, then low ownership undermines nothing.

Agency has the same structure. The experimental literature concentrates on control over bodily movement and immediate sensory effects. Creative work also involves distal decisions about what a work should express, which candidate to select, and how to revise. Agency over those decisions can differ from agency over detailed production, and intentional binding — the perceived compression of time between action and outcome — was validated on discrete action–effect episodes, not on extended texts.

Responsibility splits further. A person can feel little ownership of an output and still owe reasons for releasing it, because they selected, checked, approved, or published it. Observers can hold someone answerable for approval while recognizing limited creative contribution. Neither response is a moral verdict, and neither settles an integrity question on its own.

Diagnosing a deficit without turning it into a feeling

The authors' worked example is a personal-essay workshop whose brief, agreed before composition, distinguishes an editorial contributor from the author of a personally endorsed self-expressive essay. Both roles require substantive direction, selection, and revision. Only the second additionally asserts that the writer experiences the finished essay as their own. A writer who shaped the text, feels clear agency over the changes, and reports that the essay does not feel like theirs has adequate contribution-based warrant, intact agency, and one experiential shortfall. The same record supports editorial credit.

Three features follow. First, the construct sits at the interface of individual experience and social-normative structure. An experiential expectation counts only when it forms part of the role claim and is accepted in advance. Second, the deficit is multidimensional, and its facets need not move together: strong ownership cannot repair insufficient contribution-based warrant, and adequate warrant supplies no experiential deficit. Third, diagnosing a deficit under a community's standards is not endorsing those standards. A recognized role can rest on unjust criteria, and an unsupported claim can arise without deception.

Three outcomes, and the conditions that mute them

Authorial disavowal is the observable refusal or qualification of an offered designation. Its mechanism is a collision between a claim and the grounds the person accepts for it, and a narrower label can resolve the mismatch without withholding the work. Low felt ownership, concealment of AI use, and continued unqualified claiming are three different observations, and none counts as hidden disavowal. Draxler's ghostwriter effect limits any inference from non-ownership to refusal, and social penalties for AI use explain concealment without the proposed mechanism.

Felt answerability and observer judgment concern the user's experienced obligation and other people's credit and blame. The mechanism is mind perception: a user who reads the system as the source of content may feel less obligation to explain the choices embodied in it. Ciardo and colleagues found that attributing intentional agency to a robot reduced participants' own experienced agency in a joint task, which the authors treat as motivation for a hypothesis, not an established creative-work effect.

Reorganization of legitimation is the community-level outcome. When a designation stops communicating who directed, shaped, or checked a work, explicit contribution roles make claims assessable. An editing history can show which passages a person substantially revised and who verified and approved release, so an editorial claim can be judged without assuming sole creation. Records make contribution legible; they cannot supply a contribution that did not occur, nor guarantee that an audience values it.

Six conditions bound these predictions. Visible effort makes low-execution claims harder to warrant (B1). Market dependence suppresses observable disavowal when qualification threatens income (B2). A strong singular-origination norm weakens the warrant for delegated roles (B3). Pre-existing attribution infrastructure conditions later recognition (B4). Workflow configuration precedes the effort and experience predictions (B5). Perceived system agency strengthens the agency-to-answerability association (B6).

What is contributed, and what remains untested

The authors contribute three things. The first is the construct itself: a domain-relative mismatch between a designated role and the grounds that role requires, defined so it can be diagnosed without measuring a feeling. The second is the decomposition into separately assessable routes, so a profile rather than a score carries the analysis. The third is the role-relative classification, which separates a supported editorial role from an unsupported sole-creator claim when contribution and experience are unchanged.

What remains untested is nearly everything empirical, and the authors say so. Existing findings establish task-specific relations among output evaluation, ownership, creative self-efficacy, and agency, but they do not jointly validate the deficit or its consequences. Doshi and Hauser found that access to generative ideas raised evaluated creativity in short-story writing, with larger gains for less creative writers, while reducing collective diversity — a result about evaluated output and group variety, not about ownership. McGuire and colleagues compared editing an AI-generated poem with iterative co-creation and measured creativity and creative self-efficacy; the authors note it does not establish recovery of psychological ownership. Bellaiche and colleagues varied purported human versus AI origin while holding artworks constant, and judgments moved with the source label, appraisal criterion, and perceived effort — audience evaluations, not creators' reports.

The account's explanatory reach rests on P4 and P6 surviving tests against separate component effects and stigma-based alternatives. If relational comparisons add no predictive value, the classification still clarifies which claims are supported.

What this means for practice

• Designers and studio leads. Put consequential shaping in the workflow, not just the prompt. Selection, revision, and verification can support ownership and agency; accepting a first output offers few grounds for either.
• Editors and clients. Ask what a person did, not how much. An editing history naming substantive revisions and the person who approved release supports a narrower role without assuming sole creation.
• Instructors. Separate the role from the feeling. A student who directed and revised AI-assisted work may hold a supported claim and still say the work is not theirs; that is a finding about experience, not misconduct.
• Publishers and platforms. Disclosure that AI was used says nothing about the human contribution. Naming the contribution is what makes a claim assessable.

Limitations

• The paper reports no data, no participants, and no trial; the workshop and architect–contractor cases are constructed illustrations, not observations. The manuscript discloses Kimi K2.5 for English polishing only, with no role in argumentation.
• The standards that determine supported authorship are not universal. Conceptions of ownership, agency, and the self vary across cultures, and originality norms and effort visibility vary across fields, so diagnosis must specify the role, evaluator, and criteria.
• A deficit may be transitional. Communities can narrow a designation, recognize different contributions, or revise workflows, while others preserve unequal credit or intensify disclosure penalties.
• A deficit does not by itself show wrongdoing, and it does not remove a duty to check and stand behind a released work.

Citation

Hong, S., Li, Q., Zhou, Z., Dong, Y., Hsiao, Y.-C., & Cao, Q. (2026). Owning what one did not make: psychological ownership, agency, and the authorship legitimation deficit in human–AI co-creation. Frontiers in Psychology, 17, 1928186.