Delegating Leisure: AI, Bots, and the Limits of Automating Creative Writing, Play, and Video Consumption for Paid Work

A balanced sociotechnical thesis based on a structured integrative review and comparative documentary analysis
Initial research completed: 29 October 2026
Peer-review revision completed: 30 October 2026

Abstract#

Artificial intelligence increasingly promises to perform activities normally associated with leisure: generating stories, playing games, solving puzzles, and condensing videos into summaries. A stylized productivity-centred interpretation treats these capabilities as ways to release time for paid work. This thesis evaluates that interpretation through a desk-based structured integrative review, comparative documentary analysis, and conceptual analysis of creative writing, gameplay and puzzle solving, and video consumption. It distinguishes output delegation (obtaining an artefact, reward, answer, or summary) from process delegation (transferring the activity through which enjoyment, mastery, immersion, or restoration would ordinarily occur), and from time reallocation (deciding what happens to any time saved). Leisure is operationalized as relatively voluntary activity valued partly or wholly for experiential reasons, while allowing outputs and social recognition to matter simultaneously. Four evaluative dimensions are applied across the cases: functional, experiential, agentic, and temporal value.

The conceptual analysis shows that many defensible uses remove non-leisure components embedded within leisure practices rather than automate valued experience itself. It also identifies important counterexamples: idle games and tool-assisted speedruns can make delegated execution constitutive of play; shared control can enable participation; curatorial creativity can be valuable; and spoiler research complicates universal claims about narrative suspense. Technical reliability remains uneven across writing, games, puzzles, and video. However, no included study directly compares delegation levels in natural leisure settings and measures subsequent experience, so the preference for partial, reversible delegation is presented as a disconfirmable design expectation rather than a demonstrated finding.

Time saved at task level is not automatically discretionary time or additional paid work. The labour synthesis ranks evidence by identification and directness: workplace experiments and surveys are indirect for leisure delegation, while leisure-quality and household-technology research supplies closer but still non-AI mechanisms. No study identifies the full leisure-delegation-to-paid-hours pathway. The thesis therefore recommends provisional, purpose-sensitive delegation and explicit community rules, while treating labour-governance proposals as normative positions. Its central conclusion is limited but clear: output delegation, experiential participation, and time reallocation require separate evidence, and additional paid work is not a presumptive welfare improvement.

Keywords: artificial intelligence; leisure; automation; creative writing; gameplay; puzzle solving; video summarization; time use; self-determination; labour

1. Introduction#

1.1 The productivity promise applied to leisure#

Automation is usually justified as a way of reducing necessary labour. A newer possibility is more conceptually awkward: using AI and bots to perform activities that people supposedly choose because they want to do them. A language model can propose a plot or produce a complete story. A game agent can explore an environment, collect resources, and learn skills. A puzzle solver can return an answer without its user attempting the puzzle. A video system can produce a synopsis without its user watching the source. If these activities are treated as time-consuming outputs, delegating them appears to create time for paid work.

This thesis analyses that interpretation as a stylized conceptual proposal, not as a claim attributed uniformly to named vendors, researchers, or users. The question is whether the proposal is logically and empirically sufficient, not how many people endorse it.

That interpretation contains a hidden assumption. It treats leisure as a set of deliverables rather than a set of experiences. In ordinary work, the output is often separable from the process: a completed report may retain its value regardless of whether a person or a machine drafted it. Leisure frequently differs. The difficulty of a puzzle, the uncertainty of a game, the formation of a writer’s voice, and the unfolding of a film may be constitutive of the activity’s value. Removing the process can remove the reason for wanting the output in the first place.

Yet a purely romantic defence of unassisted leisure would also be inadequate. Not every part of a hobby is intrinsically valued. Writers can experience research, formatting, or a blank page as obstacles rather than pleasures. Players can resent repetitive grinding while valuing exploration or social play. Viewers can need to triage a large collection of informational videos without wishing to watch each in full. Automation can also make leisure accessible to people for whom standard interfaces, timing demands, or modes of expression create exclusion. The relevant question is therefore not whether AI should automate leisure in general, but which components can be delegated without destroying the values that matter in a particular context.

A second assumption concerns time. Even when automation genuinely saves time, the saving does not come with instructions for its use. It may become paid work, additional leisure, care, rest, education, or delegation overhead. The mechanism differs by domain: employers can directly absorb time saved within a job, whereas time released from a hobby is more immediately shaped by household obligations, leisure preferences, income needs, and opportunities to sell additional labour. Workplace spillover remains possible but requires a boundary-crossing mechanism. “Freeing time for paid work” therefore describes a contingent outcome, not an inherent technical effect.

1.2 Research aim and questions#

This thesis aims to explain when AI-mediated delegation of leisure is valuable, when it becomes self-defeating, and whether it can plausibly expand time available for paid work. It addresses four questions:

• RQ1: What can current AI systems and bots plausibly delegate in creative writing, gameplay and puzzle solving, and video consumption, and what important limitations remain?
• RQ2: How do different forms and degrees of delegation affect functional output, subjective experience, agency, and the social rules of leisure?
• RQ3: Under what conditions can time saved through delegation become additional paid work, and what evidence challenges the assumption that it will?
• RQ4: What design and governance principles could support beneficial delegation while protecting leisure, accessibility, fairness, and autonomy?

1.3 Central argument#

The thesis advances three connected claims:

1. Delegation is component-specific. AI can often produce an output or remove a burden without reproducing the experience associated with the unautomated activity. Assessments must therefore identify the component delegated and the purpose of the activity.
2. Automation has no single leisure value. Its effects vary with user goals, technical degree, social rules, and activity design. Partial, reversible delegation is expected to preserve agency in some settings, but that expectation has not been directly tested against full delegation in natural leisure use. Idle play, TAS, and accessibility counter any universal anti-automation position.
3. Saved-time mechanisms are domain-specific. A task-level efficiency gain is not equivalent to discretionary time. Workplace savings and leisure savings follow different allocation mechanisms, and neither automatically becomes paid work. Additional labour may be voluntary, economically compelled, or offset by leisure, care, overhead, and rebound.

The resulting position is neither technological celebration nor refusal. It is a defence of selective delegation under meaningful human control, combined with scepticism toward claims that experiential leisure can simply be consumed by proxy.

1.4 Scope and contribution#

The three cases were selected to vary the mode of engagement rather than because they exhaust leisure. Creative writing foregrounds expressive production; gameplay and puzzles foreground rule-bound interaction and performance; video foregrounds temporal reception, interpretation, and affect. Video is not treated as wholly passive: viewers interpret, anticipate, feel, and participate socially. Including reception beside production and interaction tests whether the framework travels across different forms of participation. Informational video and grinding are retained as boundary cases where the delegated component may not itself be leisure.

Creative writing includes fiction and related expressive writing, not the full publishing industry. Gameplay includes digital agents, repetitive labour, idle and tool-assisted play, competitive automation, and puzzles. Video includes summarization and question answering, with recommendation systems acknowledged but not comprehensively reviewed. Paid work is analysed as one possible use of saved time, not as a complete labour-market forecast. Chores and life administration are used only as a conceptual contrast: where an output is wanted and the process has little experiential value, straightforward substitution is less controversial, so a fourth full case would add less analytical leverage.

Delegation also predates contemporary AI. Walkthroughs, Sudoku solvers, ghostwriters, boosting services, macros, and tool-assisted speedruns already separate outputs from participation. What is potentially new is the combination of generality, lower marginal cost, conversational accessibility, personalization, and scale across several domains. The thesis therefore evaluates a change in delegation’s availability and scope, not the invention of delegation itself.

The contribution is a cross-case framework linking technical capability to experiential and institutional analysis. It separates analytical distinctions from empirical expectations and identifies the direct evidence required to test them.

2. Literature Review and Theoretical Foundations#

2.1 Leisure as activity, experience, and discretionary time#

Leisure can refer to time remaining after obligations, to activities undertaken voluntarily, or to a subjective condition of freedom and enjoyment. These meanings overlap but are not identical. Time away from employment may be filled with care or other obligations; an activity classified as recreation can feel compulsory; and a person can experience playful autonomy within a setting not conventionally labelled leisure. This thesis consequently treats leisure as a sociotechnical achievement rather than merely an empty interval on a schedule.

For the purposes of this thesis, leisure is treated primarily as voluntary activity valued partly or wholly for experiential rather than instrumental reasons. This is the operational definition used in the comparative analysis. Discretionary time describes a temporal condition, non-work describes a social location, and hobbies describe conventional activity categories; none is sufficient on its own. Restoration is one possible experiential benefit rather than a defining requirement. The phrase “partly or wholly” is important: an activity does not cease to count as leisure because it also produces income, reputation, learning, or useful outputs. Voluntariness is likewise a matter of degree where economic pressure, platform design, or social expectation constrains choice.

Classical theories of play help explain why process matters. Huizinga (1955) characterized play as voluntary and distinct from ordinary life. Csikszentmihalyi’s (1990) account of flow links deep engagement to conditions such as clear goals, feedback, and a balance between challenge and skill. These accounts do not imply that every leisure moment is intensely immersive. They do imply that efficiency is not always the appropriate measure: an activity may be valuable because it absorbs attention, permits exploration, or creates a temporary order different from productive routine.

Suits’s (1978) definition of a game as the voluntary attempt to overcome unnecessary obstacles makes the point especially sharply. In a game, the most efficient means may be disallowed because the constraint is part of the activity. A solver that bypasses a puzzle’s reasoning process may achieve the formal answer while dissolving the puzzle as a leisure practice. The same logic does not apply uniformly to all outputs, but it warns against importing an unmodified productivity model into voluntary difficulty.

2.2 Intrinsic motivation, autonomy, competence, and relatedness#

Self-Determination Theory (SDT) distinguishes intrinsic from extrinsic motivation and identifies autonomy, competence, and relatedness as central psychological needs (Deci & Ryan, 2000). Applied to games, Ryan, Rigby, and Przybylski (2006) found that experiences of autonomy and competence were important predictors of enjoyment and future play. The framework is relevant beyond games:

• Autonomy concerns whether the user experiences the activity as chosen and self-endorsed, rather than controlled by a system or external demand.
• Competence concerns effective action, learning, and mastery. Full delegation can reduce opportunities to practise, while well-designed assistance can make competence attainable.
• Relatedness concerns connection with others. Shared stories, games, and viewing can support social bonds even when the immediate activity is solitary; automation may either facilitate participation or produce outputs detached from shared practice.

SDT does not establish that manual production is always superior. A writer who delegates an unwanted mechanical task may experience greater autonomy. A disabled player using assistive control may gain both autonomy and competence. Conversely, a user who merely approves machine outputs may retain formal control while losing the engaged authorship or mastery that made the activity satisfying. The distinction between nominal and meaningful control is therefore essential.

2.3 Automation, skill, and oversight#

Bainbridge’s (1983) “ironies of automation” remains relevant to generative systems. Automating routine components can leave humans with the unusual, difficult cases, while reduced practice can weaken the skills needed for intervention. An operator may be asked to monitor a system even though sustained vigilance is a poor substitute for active engagement.

Delegated leisure can produce a smaller-scale version of this problem. A writer who always accepts generated prose may have less practice detecting cliché or structural weakness. A player who delegates ordinary decisions may be unprepared for a novel challenge. A viewer who relies on summaries may lack the contextual knowledge needed to notice an omission. Oversight is not costless: prompting, checking, editing, and recovering from errors can consume part of the apparent saving. The relevant unit of analysis must therefore be the whole workflow, not just machine execution time.

This does not imply that humans must remain involved in every operation. It implies that designs claiming “human control” should specify what control consists of, whether the user can exercise it competently, and whether the remaining oversight task is realistic.

2.4 Delegated enjoyment and interpassivity#

The possibility of delegating enjoyment is not unique to contemporary AI. Fizek (2018) analyses idle games through interpassivity: the game plays or progresses on the player’s behalf, and pleasure can arise from that delegation itself. Idle games complicate the assumption that agency must always mean continuous input. A player may enjoy configuring a system, anticipating progress, checking results, or temporarily refusing the demand to act.

This insight prevents two opposite errors. The first is to assume that delegated play is never real play. The second is to assume that because delegation can be pleasurable in idle games, any activity can be delegated without loss. In an idle game, delegation is part of the designed rule system and the player’s chosen relation to it. In a skill-based puzzle or competitive game, the same external automation can defeat the challenge or violate fair-play expectations. Delegation must be interpreted within the activity’s social and formal structure.

2.5 Productivity and the disputed leisure dividend#

The expectation that technology will reduce work and expand leisure has a long history, most famously in Keynes’s (1930/2010) speculation about the economic possibilities for future generations. Yet productivity gains can be converted into higher consumption, organizational output, profits, or reduced working time. Technology changes what is possible; it does not by itself determine distribution.

Contemporary AI evidence reinforces this distinction. Noy and Zhang’s (2023) experiment demonstrates substantial task-level savings in professional writing, but it does not show that equivalent savings would occur in leisure writing or that saved hours would remain discretionary. Humlum and Vestergaard’s (2026) Danish study found modest average savings and predominant reallocation to other work tasks. Jiang et al. (2025) associate higher occupational AI exposure with longer hours and reduced leisure, while Suh and Oh’s (2026) Korean survey reports savings absorbed substantially as on-the-job leisure. These studies concern workplace mechanisms and cannot by themselves explain time released from hobbies.

2.6 Household technology and leisure quality#

Two adjacent literatures provide a closer conceptual bridge to non-work time. Cowan’s (1983) historical analysis of household technology shows that supposedly labour-saving devices can reorganize tasks, standards, and responsibilities rather than simply release time. The analogy is limited—housework is not leisure and twentieth-century appliances are not generative AI—but it warns against treating technical capacity as a direct time dividend.

Labour-supply research also shows that the value of leisure itself matters. Aguiar, Bils, Charles, and Hurst (2021) combined time-diary data with leisure-demand estimates and calculated that innovations in gaming and recreational computing lowered young men’s market work hours by approximately 2% over their study period. This does not estimate the effect of delegating play. It demonstrates a different but crucial mechanism: when leisure becomes more attractive, it can absorb time rather than release it for paid work. A complete leisure-for-work argument must therefore model both time saving and changes in leisure quality.

3. Conceptual Framework#

3.1 Three distinctions#

The framework separates three processes that productivity narratives often combine.

Output delegation means obtaining the result associated with an activity: a story, a game reward, a puzzle answer, or a video synopsis. This can be useful when the result has value independent of the process.

Process delegation means allowing AI to perform the activity through which enjoyment, learning, immersion, expression, or restoration would normally arise. It may be partial, as when a model offers alternatives, or extensive, as when a bot plays an entire session.

Time reallocation means the subsequent use of any net time saved. It is analytically downstream of delegation and depends on preferences, obligations, workplace norms, and power.

A simple causal chain—delegate activity, save time, perform more paid work—is therefore not assumed. Each link requires separate evidence.

3.2 Technical degrees of delegation and an orthogonal interpretive axis#

Experiential replacement is an orthogonal interpretive axis, not a fourth degree of automation. At any technical degree, a user, community, marketer, or productivity narrative may make a weak claim that an output supplies information or an artefact, or a strong claim that it stands for having undergone the activity itself. The latter is labelled a supposed experiential-replacement claim because its meaning is contested rather than technologically established. Separating this axis allows, for example, full delegation accompanied only by an output claim, or partial assistance accompanied by an inflated claim of authorship.

3.3 Mixed motivations and layered value#

The distinction between instrumental and experiential goals is analytical, not a claim that real activities have only one motive. Writing can provide expressive satisfaction, a publishable output, income, and recognition; streamed gameplay can combine play, performance, sociability, and earnings; educational viewing can be enjoyable as well as useful. Speedrunning and puzzle solving may likewise join mastery and discovery to competition or social prestige. Output, process, and social signalling can therefore be valuable simultaneously.

Each case is consequently assigned a motivational profile rather than a single purpose. The analysis asks how much weight the user and community give to the artefact or result, participation in the process, and social meanings such as recognition, belonging, or status. These weights may change across components of the same activity: a writer may delegate formatting but not drafting, while a streamer may automate repetitive progression to preserve live interaction. Nor are the categories mutually exclusive—an output can carry experiential meaning, and practising a process can have instrumental benefits.

Delegation is assessed against that profile. Removing a low-value component can protect several other values, while automating one valued component may be worthwhile to the user even when it sacrifices another. The framework therefore evaluates trade-offs and conflicts within an activity rather than treating “mixed motivation” as an exception to an otherwise binary model.

3.4 Four dimensions of value#

Each case is evaluated on four dimensions:

1. Functional value: Does the system produce a useful, reliable output or achieve the desired external result?
2. Experiential value: Are enjoyment, restoration, immersion, discovery, or other sought experiences preserved or improved?
3. Agentic value: Does the user retain meaningful authorship, autonomy, competence, relatedness, and the ability to revise or refuse automation?
4. Temporal value: Is there a net saving after setup and verification, who controls it, and does its use improve welfare?

These dimensions can conflict. High functional value can coexist with low experiential value. Accessibility-oriented automation may sacrifice a narrow conception of independent performance while substantially improving autonomy and participation. A bot may save time but be illegitimate because it harms other players. No single scalar “automation benefit” is adequate.

Agentic value is also infrastructure-dependent. Local and cloud deployment alter data custody, model continuity, configurability, and dependence on a provider without constituting separate degrees of delegation. A locally run model can offer offline access and deliberate control over prompts, context, sampling settings, and model versions, but requires hardware, setup, maintenance, and attention to licensing. A managed service can provide greater convenience or capability while making access and behaviour partly dependent on provider decisions. Neither architecture determines psychological ownership or legal authorship; those require separate assessment.

3.5 Analytical criteria, empirical expectations, and normative standard#

The framework distinguishes conceptual criteria from claims that evidence could reject:

• C1 — Constitutive-process criterion (analytical). Where value is defined by performing a process, receiving its output is not the same form of participation. This follows from the activity’s motivational profile; it is not presented as an empirical discovery.
• C2 — Component-boundary criterion (analytical). A leisure practice can contain non-leisure components. Removing formatting, search, or unwanted grinding may preserve leisure precisely because the delegated component was not itself valued as leisure. Informational video used solely to obtain a required fact is likewise a boundary case, not evidence that experiential viewing has been automated.
• H1 — Friction-and-control expectation (empirical). For components users initially experience as unwanted friction, partial and reversible delegation is expected to preserve more experiential and agentic value than opaque full delegation. This expectation would be disconfirmed if comparable users reported equal or greater enjoyment, autonomy, competence, and reflective endorsement under full delegation after accounting for output quality, accessibility, and oversight. The reviewed corpus contains no direct leisure study making that comparison, so H1 remains a design expectation rather than a tested finding.
• H2 — Reallocation expectation (empirical). Any effect of leisure-time savings on paid work depends on labour-supply opportunities, preferences, income security, care obligations, and household and workplace institutions. It would be disconfirmed for a population if exogenous leisure delegation produced a stable, net increase in paid hours across those conditions. Existing studies do not identify that effect.
• H3 — Social-expectation rebound (empirical). Where highly productive delegation becomes visible and status-conferring, community baselines for acceptable output volume or polish may rise, increasing perceived pressure to delegate. H3 would be disconfirmed in settings where such adoption leaves expectations and perceived pressure unchanged or reduces them, for example because shared resources lower everyone’s production burden. The reviewed corpus contains no direct test; social rebound is a conditional mechanism, not an inherent effect of AI adoption.

Self-defeating delegation is defined normatively as delegation that frustrates goals the user would reflectively endorse with adequate information about process loss, oversight, and social consequences. Immediate preference is relevant but not infallible: affective-forecasting research shows that people can misjudge the intensity and duration of future emotional responses (Wilson & Gilbert, 2005). This standard does not permit an analyst to override any preference they dislike; it requires evidence about the user’s endorsed goals, expectations, and subsequent experience.

3.6 How the framework’s categories fit together#

The checklist in Section 9.8 operationalizes these stages rather than adding another taxonomy. Configurations are outcomes of the analysis, not properties inferred from the technical degree alone.

4. Methodology#

4.1 Research design#

The study uses a structured integrative review combined with comparative documentary analysis and conceptual analysis. An integrative design is appropriate because the research question spans computer science, human–computer interaction, game studies, creativity research, copyright guidance, household technology, and labour economics. The aim is to assess the internal logic and evidence requirements of a stylized leisure-for-paid-work proposal, not to estimate an effect size or measure the prevalence of that proposal in public discourse.

No interviews, surveys, diary study, or participant tasks were conducted. Existing user studies are treated as secondary evidence. AI-roleplayed interviews were excluded because simulated responses cannot establish human experiences and would introduce fictitious empirical material.

4.2 Iterative research process#

Research and revision proceeded through nine cycles:

1. Scoping and concept formation. The initial productivity framing was decomposed into output delegation, process delegation, and time reallocation. Search concepts were developed for each case and for leisure theory.
2. Case-specific evidence collection. Searches targeted creative co-writing, creativity and diversity, game agents, idle play, puzzle benchmarks, fair-play rules, video summarization, multimodal understanding, and accessibility.
3. Counterexample search. Evidence was deliberately sought against the provisional argument, especially idle games, shared control, selective writer delegation, and query-focused summaries. This prevented a simplistic equation of automation with experiential loss.
4. Cross-case and labour synthesis. Workplace experiments, surveys, and time-diary studies were compared to test whether task savings could reasonably be treated as paid-work capacity.
5. Adversarial revision. Claims were checked for task context, numerical accuracy, counterarguments, and overgeneralization. Disputed or unnecessary numerical estimates were removed.
6. Reference and consistency audit. Core metadata and links were checked, working papers and preprints were labelled, official policies were separated from scholarly evidence, and in-text citations were reconciled with the reference list.
7. Conceptual clarification after critical review. The operational definition of leisure was fixed, mixed motivations and economic compulsion were integrated, communal leisure and recommendation systems were acknowledged, and repeated formulations were condensed.
8. Major revision after peer review. Analytical criteria were separated from empirical expectations, the labour evidence was re-ranked by design and directness, closer non-work literatures and counterexamples were added, and the review audit trail and AI-assistance disclosure were expanded.
9. Infrastructural and communal boundary-case revision. Deployment context was added as a condition of agency, automation construction was distinguished from automation use, and social rebound was formulated as the conditional and disconfirmable H3 expectation rather than an observed community effect.

The iterative record is summarized in Appendix A.

4.3 Source selection#

Initial searches were conducted on 29 October 2026; targeted citation verification and review-driven supplementary searches were conducted on 30 October 2026. Locations included scholarly web search, arXiv, ACL Anthology, CVF Open Access, publisher and institutional repositories, NBER, backward citation chasing, and official policy and community-rule websites. Four initial saved scholarly result sets contained 37 records: 12 for creative writing, 12 for games and bots, 3 for video summarization, and 10 for time and labour. These counts describe the saved sets, not all records encountered through subsequent exact-title and official-document searches.

Query groups combined delegation terms with domain terms—for example, “generative AI” with “creative writing,” “storywriters,” or “creative diversity”; “LLM agent” or “bot” with “gameplay,” “idle games,” “puzzle,” or “shared control”; and “video summarization” with “personalized,” “long video,” “hallucination,” or “omission.” Labour searches initially combined “generative AI” with “time savings,” “time allocation,” “work hours,” and “leisure.” Supplementary searches targeted “leisure luxuries” and labour supply, household technology, spoilers and enjoyment, affective forecasting, and tool-assisted speedrunning. Exact-title searches were used to verify primary records and disputed metadata.

A single reviewer conducted source selection and individual-source extraction, with AI-assisted search and drafting support as disclosed below; no independent duplicate screening was undertaken. Inclusion required direct relevance to at least one research question, a traceable primary record, and a clear role as empirical evidence, theoretical interpretation, technical documentation, or institutional rule. The corpus prioritized peer-reviewed studies, seminal theoretical work, primary technical papers, and official guidance. Working papers and preprints were included only for emerging questions and labelled by status. Sources were excluded or downweighted for unsupported marketing claims, duplication without added analysis, conflation of professional and leisure tasks, or numerical claims that could not be traced to their original context.

This audit trail makes the strategy inspectable, but it is not a PRISMA-style systematic-review log. Iterative targeted searches were not deduplicated into a single screening database, so no comprehensive screening total or inter-rater reliability statistic is claimed. The review remains purposive and vulnerable to selection and confirmation bias despite deliberate counterexample searches and external critical review.

4.4 Analysis and synthesis#

Evidence was extracted into case-specific matrices recording the activity, delegation level, purpose, outcome, limitations, and source type. Sources were weighted on two separate axes: strength of design for the claim actually studied and directness for leisure delegation. A randomized workplace experiment can therefore be strong evidence for professional task productivity while remaining indirect for hobby experience; an official community rule can be strong evidence for a governance category while saying nothing about prevalence.

The operational definition in Section 2.1 governed the comparison: an activity was analysed as leisure when it was relatively voluntary and valued at least partly experientially, not simply because it occurred outside paid work. Each activity was assigned a mixed motivational profile covering output, process, and social significance. The four value dimensions were then applied within each case. Cross-case synthesis used convergence and divergence, while ordinal comparisons in Section 8.1 record relative judgements within this corpus rather than measured population effects.

Comparative documentary analysis was used for platform rules, accessibility guidance, and copyright guidance because these documents define legitimate conduct or institutional categories rather than report user effects. Their role in the argument is therefore different from that of an experiment.

4.5 Evidentiary conventions and limitations#

The thesis distinguishes reported findings from analytical inference. For example, a creativity experiment can support a claim about story evaluations under its conditions; it cannot by itself establish how a particular hobbyist will experience AI over years. Technical benchmark performance does not demonstrate experiential substitution. A platform prohibition establishes a rule, not the prevalence or moral status of every possible use.

The central evidentiary limitation is that no included study directly compares partial, reversible delegation with full delegation in a natural leisure setting and measures subsequent experience and paid-work time. Experiential conclusions are therefore conceptual inferences supported by adjacent user studies, not demonstrated effects of leisure automation. Workplace time-use findings concern a different mechanism and are used only as limited analogies. Spoiler studies, idle-game research, and community rules complicate the framework but do not fill that direct-evidence gap.

Further limitations include rapid model change, heterogeneous benchmarks, publication lag, uneven geographic coverage, and limited longitudinal research. The review is restricted to accessible English-language sources and does not measure the distribution of user practices. Fan-writing disputes, local-model users, automation-building hobbyists, botting and gold-farming communities, boosting services, and household delegations were not systematically sampled; the included community documents are illustrative rather than representative. These limitations require conditional conclusions and prevent strong claims about welfare, prevalence, or long-run labour supply.

4.6 AI assistance and research ethics#

AI tools assisted query expansion, scholarly and web-source discovery, metadata comparison, drafting, and consistency checks. They did not serve as evidentiary sources, and AI-generated summaries were not substituted for the primary records used to support central claims. Citation metadata and key numerical claims were checked against accessible primary or publisher records; inaccessible or provisional materials are identified rather than treated as settled. No interviews, surveys, diary study, or participant tasks were conducted, and AI-roleplayed interviews were excluded because simulated responses cannot establish human experience.

Because no new human participants were recruited, the study collected no personal participant data. Ethical analysis instead concerns transparent representation of evidence, accessibility, fair play, authorship, and labour. Particular care was taken not to equate disability-related assistance with cheating or to assume that greater paid-work time is inherently welfare-enhancing. The single-reviewer and AI-assisted workflow increases rather than removes the need for independent checking before formal submission.

5. Case Study I: Creative Writing#

5.1 What is being automated?#

“AI creative writing” covers several materially different practices. A model can supply prompts, names, plot possibilities, stylistic examples, or feedback; continue a passage; rewrite text; generate a complete story; or manage mechanical work surrounding composition. These practices do not have equivalent implications for leisure. A writer may welcome an alternative ending while rejecting automated drafting, or use generated prose as raw material while retaining extensive editorial control.

The empirical literature accordingly supports a differentiated account. Biermann, Ma, and Yoon’s (2022) study of 7 hobbyist and 13 professional writers found that writers imagined AI not simply as a tool but as a more companion-like collaborator, while raising concerns about autonomy, control, ownership, values, and writing strategies. Guo et al.’s (2025) study of 18 experienced AI-using creative writers similarly found selective integration: writers delegated points of friction but evaluated model contributions against their own voice, vision, quality standards, and genre expectations. These are small, situated studies rather than population estimates, but they directly challenge the idea that writers either accept or reject automation wholesale.

A useful distinction is between generative latitude and expressive commitment. Brainstorming expands a space of possibilities without committing the writer to any one of them. Drafting commits the work to particular language, pacing, and emphasis. The latter may save more time, but it also moves more of the expressive process into the model. The relevant design question is not simply how much text AI can produce, but where the writer wants commitment to occur.

5.2 Functional gains and individual creativity#

There is credible evidence that generative AI can improve evaluated outputs in some writing tasks. Doshi and Hauser (2024) conducted an online experiment with 293 participants writing short stories. Access to AI-generated ideas improved evaluations of individual stories, with larger benefits for participants assessed as less creative at baseline. The study is important because it concerns creative writing rather than only workplace composition. It suggests that assistance can reduce the intimidation of invention and provide material that users might not otherwise produce.

Porter and Machery’s (2024) poetry experiments offer a complementary finding about reception: non-expert participants did not reliably distinguish AI-generated poems from human-written ones and often rated the AI-generated poems favourably. The result should not be inflated into a general proof of artistic equivalence. It involved particular models, poems, and non-expert judgments. It nevertheless shows that origin-based assumptions about quality can diverge from blind reception.

Workplace research supplies a benchmark for speed, not a direct measure of leisure benefit. In Noy and Zhang’s (2023) experiment with 453 college-educated professionals, access to ChatGPT reduced average completion time for mid-level professional writing tasks by approximately 40% and increased output quality by 18%, as assessed within the experiment. These tasks were not hobbies, and the study did not measure the satisfaction of writing. The result demonstrates that functional efficiency and quality can improve together; it does not demonstrate that a person who loves writing should want to write less.

5.3 Collective diversity and the problem of convergence#

Functional improvement at the individual level can coexist with a collective cost. Doshi and Hauser (2024) found that AI-assisted stories were more similar to one another. Anderson, Shah, and Kreminski (2024) likewise reported group-level homogenization in creative ideation among ChatGPT users; users produced more detailed ideas but felt less responsible for them. These findings are not identical in method or task, but they converge on an important distinction: a model can help each user exceed an individual baseline while drawing many users toward similar possibilities.

For leisure writing, homogenization is not merely an aesthetic concern. A hobby can be a way of developing idiosyncratic voice and exploring unusual perspectives. If generated suggestions repeatedly occupy the same stylistic and narrative centre, convenience may narrow the writer’s exploratory range. This risk is likely to depend on use. Requesting many contrasting possibilities, supplying distinctive constraints, using local or personal material, and treating outputs as provisional could preserve diversity more effectively than accepting the first fluent continuation.

The evidence therefore rejects both a blanket claim that AI “kills creativity” and a blanket claim that higher-rated outputs prove unqualified creative enhancement. Individual quality, individual experience, and collective diversity are separate outcomes.

5.4 Authorship, ownership, and experiential value#

Creative writing can combine expressive discovery, the satisfaction of sustained effort, publication, income, and recognition. The significance of full delegation depends on this motivational profile: it can accelerate an artefact-oriented project while removing much of the discovery available to a writer who values composition itself. The assessment therefore concerns which values are gained or lost, not whether the generated text is inherently good or bad.

This is not an argument that only solitary, unaided composition counts as authorship. Writers have always used prompts, editors, collaborators, conventions, and tools. The relevant issues are the degree and location of contribution, the user’s understanding of that contribution, and whether the arrangement supports the writer’s goals. A person may value curating and directing generated material as a new creative form. That practice should be assessed on its own terms rather than treated either as fraudulent or as identical to conventional drafting.

Meaningful control involves more than pressing a regenerate button. It includes the ability to set purposes, reject suggestions, revise at the desired granularity, understand the system’s role, and preserve a stable relation between intention and output. Guo et al.’s (2025) account of writers checking contributions against voice and vision illustrates such evaluative control. Biermann et al.’s (2022) findings show why control and ownership cannot be assumed merely because a human remains nominally “in the loop.”

5.5 Copyright and provenance#

Legal authorship is a related but distinct issue. The U.S. Copyright Office (2025) states that the use of AI as an assistive tool does not prevent copyright protection, but protection requires sufficient human authorship in the resulting work. Prompts alone generally do not provide sufficient control over expressive elements to make the prompter the author of the generated expression; independently authored material, creative selection and arrangement, or creative modifications may nevertheless be protectable. This guidance concerns U.S. copyrightability, not artistic merit, personal fulfillment, or every jurisdiction.

For leisure communities, provenance matters even where no legal dispute exists. Writing groups, competitions, and platforms may require disclosure or restrict generated text. Such rules protect expectations about what participants have contributed and what a comparison between entries means. Provenance should therefore be treated as part of social legibility: recipients should not be led to attribute a process to a person when that process was largely delegated.

5.6 Local deployment, cloud services, and psychological ownership#

The reviewed writing studies should not be generalized into a single, infrastructure-free account of AI. A writer using a managed cloud service and one running a smaller model locally may both engage in partial delegation while facing different constraints. Local inference can keep drafts and prompts on the writer’s own hardware, support offline work, and permit deliberate control of model version, system prompt, sampling settings, and context. These affordances may make the model feel like a stable, personally configured instrument; tuning it can also become part of the creative process.

That possibility does not establish that local output is automatically more authentic or psychologically owned. Psychological ownership remains an empirical question, while legal authorship depends on human contribution rather than the location of computation. Local systems also require suitable hardware, installation, model selection, maintenance, and sometimes different capability expectations. Model weights, code, and generated content can be subject to distinct terms, so “open-weight” should not be treated as synonymous with unrestricted use. Cloud services can offer useful revision tools, accessibility, and greater capability while introducing dependence on provider retention, updates, and access rules. The relevant comparison is therefore among specific configurations of data custody, continuity, configurability, capability, and labour—not a binary opposition between corporate opacity and local authenticity.

5.7 Accessibility and beneficial delegation#

Automation can expand creative participation. Language support can help a person write in an additional language; speech and generative interfaces can reduce motor demands; structural suggestions can help users organize ideas; and feedback can make an isolating hobby more approachable. These uses can support autonomy and competence rather than diminish them.

The accessibility case also exposes a flaw in treating unassisted production as the sole authentic baseline. All writing depends on tools and conventions. The relevant ethical question is whether a configuration enables the person to pursue their own expressive purposes, not whether it satisfies an arbitrary standard of technological purity. At the same time, accessibility should not be used to excuse opaque systems that overwrite a user’s voice or make revision difficult.

5.8 Case assessment#

Within this corpus, assistance and selective partial delegation have the most direct support for preserving writers’ purposes. No included study tests long-term leisure satisfaction across delegation levels, so the preference for selective delegation remains a provisional design implication.

6. Case Study II: Gameplay and Puzzle Solving#

6.1 A diverse automation landscape#

Game automation ranges from macros and scripted bots to generalist agents and external puzzle solvers. Voyager demonstrated an LLM-powered Minecraft agent capable of autonomous exploration and skill acquisition within its environment (Wang et al., 2023). Google DeepMind’s SIMA project demonstrated a generalist agent following natural-language instructions across several 3D game environments, while its reported results remained below general human-level performance (SIMA Team et al., 2024). SIMA 2 subsequently reported substantial improvement and a narrower gap with human performance across its evaluation portfolio, but its results still concerned defined environments and tasks rather than reliable play in arbitrary games (SIMA Team et al., 2025). These systems show meaningful technical progress, but they do not establish that any game can be played dependably on a user’s behalf.

Puzzle performance is similarly uneven. Loredo Lopez, McDonald, and Emami (2025) evaluated models on 358 New York Times Connections puzzles and identified a substantial human–model performance gap. Earlier work on 438 Connections games likewise found that the best evaluated model at that time fully solved only 18% (Samadarshi et al., 2024). Such results are snapshots of particular systems and prompts, not permanent ceilings. They show why capability claims must be task-specific even before experiential questions are considered.

For analysis, four purposes of game automation must be separated:

1. removing repetitive or unwanted in-game labour;
2. obtaining rewards, status, or progression;
3. enabling access or assisting control;
4. bypassing the challenge that constitutes play.

The same technical mechanism can serve different purposes and have different ethical implications.

6.2 Why solving is not always playing#

Suits’s (1978) unnecessary-obstacle account clarifies the puzzle case. If a person asks for the answer to a crossword, the functional problem may be solved, but the leisure activity of solving has not occurred. Even a hint changes the challenge, although it may preserve enough reasoning to support enjoyment and learning. Assistance can therefore be placed on a continuum from orienting feedback to complete substitution.

Games also provide experiences of competence and autonomy associated with enjoyment (Ryan et al., 2006). A bot that performs difficult actions can improve an account’s measurable progress while reducing opportunities to develop skill. Yet play, output, and social value often overlap. A streamer may automate repetitive progression to preserve live performance and interaction, whereas a speedrunner’s audience may value the record precisely because it certifies a demanding process. The outcome therefore depends on the whole motivational profile, not a simple opposition between wanting a reward and wanting to play.

Shared knowledge of how a reward was obtained contributes to its social meaning. An achievement earned through a bot may have the same database representation as one earned through play, but not the same standing in a community that values performance. Social signalling can thus make both the visible output and the process behind it valuable simultaneously.

6.3 Grinding and the case for partial delegation#

Not all gameplay is challenging or intrinsically rewarding. Repetitive resource collection, daily tasks, and mandatory waiting can resemble labour. Players may automate these components to reach exploration, narrative, competition, or social interaction they value more. Here H1 applies as a testable expectation: delegation may protect leisure by removing demands that obstruct it, provided monitoring and rebound do not erase the benefit.

Background and idle games show that the boundary is already unstable. Keogh and Richardson (2018) describe how background games become entangled with everyday routines and demands. Cutting, Gundry, and Cairns’s (2019) study of Neko Atsume identified engagement through time spent playing, checking frequency, and sociability around the game, rather than through continuous direct input alone. Waiting and progress while away are not simply absences of play; they organize anticipation and return.

This means “automate the grind” is not a universally benign instruction. A designer may deliberately use pacing, scarcity, or waiting to structure experience. Automation can also encourage more frequent checking rather than release attention. The practical question is whether delegation reduces an unwanted demand or intensifies a compulsion to monitor progress.

6.4 Idle games, tool-assisted play, and automation construction as counterexamples#

Idle games challenge the claim that play requires continuous manual performance. Fizek (2018) argues that delegated enjoyment can be part of the idle-game aesthetic: progress occurs without constant action, while the player configures, anticipates, interprets, or checks outcomes.

Tool-assisted speedruns (TAS) provide a different, explicitly governed counterexample. TASVideos’ submission rules define distinct publication classes, require technical reproducibility and attribution, and regulate which tools and modifications are permitted (TASVideos, n.d.). Automated or frame-precise execution can therefore become the material of optimization and creative performance rather than an illegitimate substitute for play. These documents establish a community rule structure, not how often players participate or how much they enjoy it.

Automation construction adds a further boundary case. Designing, scripting, debugging, and validating a macro or bot can be the challenging activity the hobbyist values. An HTML/JavaScript tool, a Python validation script, or a modular automation system can turn repetitive grinding into an engineering puzzle whose obstacles are voluntarily adopted. The valued process then shifts from executing game actions to constructing and understanding the means of executing them. This is transformative automation, not the disappearance of leisure.

This case also disciplines temporal analysis. Building and maintaining automation may take longer than performing the original task, yet remain worthwhile because engineering is the desired leisure. Alternatively, development effort may be repaid through repeated execution savings. Assessment must therefore count development, debugging, documentation, and maintenance separately from later execution time. Authorization and effects on others still matter: elegant engineering does not make an unauthorized competitive bot fair.

The implication is not that all game activities are equally delegable. Rather, the legitimacy of delegation depends on the game’s design, the participant’s chosen relation to it, and the community’s categories. An external bot in a skill-based contest, an internal idle mechanic, and a declared TAS may all reduce direct input while occupying very different social positions.

Idle games also weaken the assumption that automation necessarily converts leisure into work time. A game designed to progress while absent can coexist with work, rest, or other activities, but notifications and optimization can continue to claim attention. Delegation may alter the rhythm of leisure rather than simply remove its duration.

6.5 Competitive integrity and platform rules#

Automation can impose costs on others. Blizzard’s End User License Agreement prohibits unauthorized bots that automate control of its games, and Chess.com’s Fair Play Policy requires players’ moves in applicable games to be their own, prohibiting engine or bot assistance during play while distinguishing permitted play against computer opponents. These documents are context-specific institutional rules, not universal moral laws. They nevertheless demonstrate why functional success is insufficient: a bot may achieve its owner’s goal while undermining fair competition, matchmaking, and trust.

Disclosure alone does not make every use acceptable. In a rated contest, other participants cannot necessarily consent to engine-assisted opposition. In cooperative or economic game environments, unauthorized automation can also affect scarcity and other players’ opportunities. Conversely, clearly labelled bot opponents, private sandbox environments, accessibility tools, and designer-authorized automation occupy different ethical positions.

Governance should therefore distinguish at least three questions: Is the automation authorized? Does it affect other participants? Is the resulting achievement represented as equivalent to unassisted performance? A single anti-bot label obscures these distinctions.

6.6 Accessibility and shared control#

Assistive automation can enable participation for players who cannot use conventional input schemes or sustain required actions. Shared-control systems can combine a player’s intent with automated execution or navigation. Ahmetovic et al.’s (2026) shared-control work for people with upper-limb impairments illustrates both the promise and the design difficulty: assistance must interpret intent and coordinate control without leaving the player feeling displaced. Because this is a recent preprint, it is used here as a technical design example rather than settled population-level evidence.

Accessibility complicates competitive and experiential judgments. An assistive configuration may automate a motor component while preserving strategic choice and challenge. Equating all automation with cheating would confuse the purpose of the rule with its letter. At the same time, accessibility does not eliminate the need to consider fairness in specific competitive contexts. The appropriate response is granular assessment and configurable accommodation, not blanket exclusion.

6.7 Case assessment#

The case motivates selective delegation and accessibility-sensitive shared control as testable design expectations. Idle play, TAS, and automation construction show why those expectations must remain subordinate to the activity’s rules and community-defined forms of participation.

7. Case Study III: Video Consumption on the User’s Behalf#

7.1 Summarization as a family of tasks#

Video summarization is not one operation. Systems may select keyframes or shots, produce a short video skim, generate a textual synopsis, answer a query, or personalize selection according to user interests. Apostolidis et al. (2021) review deep-neural approaches and the field’s evaluation challenges. Peronikolis and Panagiotakis (2024) place user preferences at the centre of personalized summarization, while Sharghi, Laurel, and Gong (2017) frame query-focused summarization around relevance to a user’s information need.

This variety matters because different tasks substitute for different amounts of viewing. A timestamped outline can help a viewer navigate the original. A query-focused summary can answer whether a tutorial contains a required step. A generic synopsis can support triage. None necessarily represents every visual, auditory, narrative, or emotional feature of the source.

The phrase “watching on behalf of the user” is therefore misleading unless qualified. A system processes signals and produces an output; it does not literally transfer a viewing experience. The strongest defensible claim is that it can perform information extraction, selection, or attention triage.

7.2 Technical reliability and evaluation#

Long-video understanding remains technically demanding. Video-MME contains 900 videos totaling 254 hours and 2,700 question–answer pairs across durations and modalities, providing a broad benchmark for multimodal video analysis (Fu et al., 2025). Its construction illustrates the range of temporal and audiovisual reasoning required, rather than proving that any single model can replace attentive viewing.

Reliability involves more than selecting plausible highlights. Rawal et al.’s (2025) ARGUS benchmark evaluates both hallucination and omission in free-form video descriptions. This dual focus is important: a fluent description can include unsupported claims, while an accurate but incomplete one can leave out exactly the event a viewer needs. Multiple-choice accuracy or generic quality ratings do not fully capture either failure.

Evaluation also depends on purpose. An extractive summary judged representative by an aggregate metric may omit a minority perspective or a user-specific detail. Personalization can improve relevance, but it requires assumptions about interests and can narrow exposure. A summary should therefore be treated as a fallible representation, not a transparent window onto the source.

7.3 Informational, entertainment, and restorative viewing#

Video serves multiple purposes. A viewer may need instructions, news, evidence, entertainment, companionship, aesthetic experience, or background relaxation. Delegation has different implications in each case:

• For information extraction, a summary may be an effective substitute when its claims are accurate and the required detail is preserved.
• For triage, a synopsis or quality estimate can help decide whether full viewing is worthwhile; the summary complements rather than replaces viewing.
• For entertainment and immersion, the unfolding of events, performances, editing, music, and suspense are often constitutive. Reading a plot summary is a different activity.
• For restoration or companionship, a concise answer may remove the very duration and low-demand attention that the viewer sought.

Restoration is relevant here as one possible experiential benefit, not as a necessary feature that every leisure activity must display. These purposes also frequently coexist. An educational video may offer enjoyment and community membership as well as instruction; an entertainment video may supply cultural knowledge or material for conversation. Summarization should therefore be evaluated against the viewer’s particular mixture of informational, experiential, and social goals rather than against a fixed classification of the video’s genre.

Spoiler research provides an adjacent experiential counterpoint. Leavitt and Christenfeld (2011) reported that participants rated spoiled short stories as more enjoyable in their experiments, while later work identifies boundary conditions and differences between forecast and experienced responses (Yan & Tsang, 2016). These studies do not test AI summaries or establish that viewers prefer synopses to video. They do show that suspense and prior plot knowledge do not have one universal relationship with enjoyment, weakening any simple claim that compression necessarily impairs narrative experience.

7.4 The verification paradox#

Summaries are easiest to trust when the user already knows the source well enough to check them, but that is precisely when summarization saves least. For unfamiliar material, verification may require watching relevant segments or consulting additional sources. This creates a verification paradox analogous to Bainbridge’s oversight problem: the user is asked to supervise a compressed representation without necessarily possessing the context needed to detect omissions.

Practical designs can reduce this problem. Timestamped claims, linked transcript evidence, confidence-sensitive language, and easy transitions from summary to source make verification possible without restarting the entire video. Summaries should distinguish observed content from inference and avoid presenting uncertain detail as fact. For consequential uses, full reliance on an unverified synopsis is inappropriate.

Not every entertainment viewing requires factual verification. The paradox matters most when the summary is used to make decisions, form beliefs, or claim knowledge of the source. In those settings, the net temporal value must include any required checking.

7.5 Personalization, autonomy, and attention#

Personalized summaries can reduce overload and improve accessibility, particularly for users who need shorter sessions, alternative modalities, or content matched to specific goals. They can also shift agency from the viewer to a system that predicts what matters. If the selection criteria are opaque, the viewer may not know what has been excluded or how interests have been inferred.

Summarization is also downstream of a more pervasive form of delegation: recommendation, ranking, and feed curation often determine which videos become available for leisure in the first place. These systems are outside the thesis’s detailed technical scope, but their exclusion does not imply lesser importance. They can reduce search effort while shaping exposure, tastes, social conversation, and the amount of time users ultimately spend watching. A complete account of automated video leisure would therefore need to examine curation and summarization together.

A balanced design should allow users to specify queries, adjust length and detail, inspect evidence, and move fluidly between summary and source. The goal is not to preserve every minute of viewing at all costs. It is to ensure that compression serves the viewer’s purposes rather than converting all media into an obligation to extract information efficiently.

7.6 Case assessment#

Video delegation has its clearest functional justification in navigation and bounded information extraction. Its experiential effects remain under-evidenced and purpose-dependent. Systems should make summaries inspectable and preserve a low-friction route back to the original.

8. Cross-Case Analysis: From Delegated Tasks to Paid Work#

8.1 What the cases share#

Across the cases, beneficial delegation depends on alignment with the activity’s motivational profile and social setting. A central finding is also negative: many of the clearest benefits concern non-leisure components embedded within leisure practices, or instrumental media uses that were not leisure under the operational definition.

The following ratings are ordinal judgements within the reviewed corpus, not measured effect sizes. “Directness” refers specifically to evidence about leisure delegation.

The comparison also shows that more automation is not always less agency. Agency depends on the relation between the user’s purpose and the system’s action. A disabled player can gain agency through shared control; a writer can gain agency by delegating an unwanted mechanical task; a viewer can gain agency through navigable summaries. Conversely, full delegation can reduce agency even when it formally follows a user instruction, particularly where the system’s selections are opaque or difficult to revise.

8.2 Four configurations of delegation and experience#

The cases can be organized into four configurations:

1. Complementary automation: AI removes friction while the user retains the valued process. Examples include feedback on a draft, a puzzle hint, or video chaptering. This configuration has the strongest claim to improving leisure.
2. Transformative automation: delegation changes the activity into a different but potentially legitimate practice. Curating generated stories, configuring an idle game, engineering a bot, or comparing video summaries can be creative or playful in themselves. The activity should be described accurately rather than treated as unchanged.
3. Substitutive automation: AI supplies an output where the user’s goal is primarily instrumental. A required factual answer or a routine game resource may be obtained without concern for process. This can create genuine time savings.
4. Self-defeating automation: the system removes the source of value while the user continues to describe the output as the original leisure experience. Fully delegating a puzzle one enjoys solving, or replacing a beloved film with its plot summary, exemplifies this configuration.

These categories are relational, not fixed properties of tools. The same system can be complementary for one user and self-defeating for another. “Self-defeating” is judged against the reflective-endorsement standard in Section 3.5, not merely an observer’s preference for unassisted leisure. Substitutive cases often sit at the leisure boundary: they can release time without demonstrating that an experiential component has been successfully delegated.

8.3 Net time is not machine execution time#

A credible temporal analysis must include the whole workflow:

Net minutes released = activity minutes no longer required − total delegation overhead (design, setup, development, prompting, monitoring, verification, revision, and error recovery).

This expression counts time only; it is not a welfare equation. It also avoids treating supervision, verification, and recovery as mutually exclusive measured categories: empirical studies should define their boundaries before summing them. Rapid generation can overstate savings when editing, configuration, or source checking is substantial, while other bounded uses may release considerable time with little overhead. When building automation is itself the desired leisure, development hours are not welfare-deadweight losses, but they still count against the net minutes available for paid work.

Minutes are not fungible. An hour of fragmented monitoring is not equivalent to an uninterrupted hour of leisure or paid work, and removing a valued leisure hour can reduce welfare even when the clock shows a saving. A welfare appraisal therefore needs a separate, value-weighted account of enjoyment, restoration, attention, timing, income, and reflective endorsement. The thesis does not assign numerical weights because the corpus provides no defensible basis for them. Temporal value consequently includes both net minutes and the quality and control of the time involved.

8.4 Evidence ranked by directness and identification#

No included study estimates the full chain from leisure delegation to net leisure-time savings and then to paid hours. The evidence must therefore be ranked by both the quality of identification for its own research question and its directness for this thesis’s mechanism.

Closer non-work evidence. Aguiar et al. (2021) connect leisure quality to labour supply using time diaries and a leisure-demand model. Their estimate that gaming and recreational-computing innovations reduced young men’s work hours by about 2% is closer to the leisure–labour margin than workplace productivity studies, but it concerns improved leisure, not delegated leisure or exogenous time savings. Cowan’s (1983) household-technology analysis similarly supports a mechanism of task and standard reorganization, but as historical interpretation rather than an estimate of AI effects.

Indirect workplace evidence. Within their own settings, Noy and Zhang’s randomized experiment offers the strongest causal identification for professional task performance; Humlum and Vestergaard’s linked administrative design offers stronger evidence on recorded earnings and hours than self-report alone; Jiang et al.’s occupation-level exposure analysis is observational; and Suh and Oh’s self-reported preprint has the weakest causal identification. None directly identifies the leisure-delegation mechanism.

The correct synthesis is therefore not that these findings are equally weighted or simply contradictory. They answer different questions with different designs. They establish that task savings are possible and that workplace allocation varies, while leaving the proposed leisure-to-work pathway unidentified. In particular, employers cannot directly raise targets on an hour released from a hobby in the same way they can absorb an hour saved within employment. Workplace intensification is a possible spillover, not a default mechanism for freed leisure time.

8.5 Pathways specific to leisure-time savings#

These pathways can coexist and distribute unevenly. A freelancer may be able to sell an extra uninterrupted hour, while a salaried employee receives no income for doing so. A person with care obligations may have fragmented rather than usable time. Leisure quality also matters: Aguiar et al.’s (2021) results show that more attractive leisure can reduce labour supply, the opposite of the proposal’s desired direction. Direct tests should therefore measure usable time, leisure quality, household constraints, and actual paid hours together.

8.6 Is automating leisure for paid work coherent?#

The proposal is coherent under limited conditions: the delegated component is unwanted or instrumentally valued; the system is reliable; net usable time remains; the person can sell additional labour; and doing so reflects either a preferred goal or a financially constrained decision whose welfare effects are evaluated separately. These are demanding empirical conditions, not consequences of technical capability.

The strong experiential-replacement version is conceptually confused when it claims that obtaining an output necessarily supplies the participation valued in the original activity. The weaker version—removing a non-leisure component from a leisure practice—is often coherent but should be described accurately. Chores provide the limiting contrast: if only a clean output matters and no valued process or social rule is lost, delegation needs no special leisure theory. Leisure becomes analytically distinctive where experience, performance, and social recognition are themselves at stake.

There is also a rebound problem. If AI makes outputs cheaper, norms may shift toward expecting more stories, progression, media, and professional output. The user can become busier managing abundance rather than liberated from activity.

A distinct social rebound occurs if highly productive participants change the baseline against which everyone else’s contribution is judged. In amateur fiction or tabletop roleplaying communities, for example, heavily delegated production of lore, adventures, or visual assets could make volume and polish newly salient. Participants might then feel pressure to automate in order to retain recognition, meet perceived obligations, or keep pace with collaborative preparation. Leisure would remain formally optional while the practical costs of belonging or standing within the community increased.

H3 makes that mechanism conditional and testable. Heavy output need not become a norm; communities can value short or improvised contributions, share generated resources to reduce everyone’s workload, or reject volume as a status criterion. Whether social rebound occurs, how it interacts with unequal hardware and skill access, and whether additional work improves welfare remain open empirical questions rather than conclusions established by the reviewed evidence.

9. Design and Governance Recommendations#

These recommendations are provisional design implications, not interventions tested by this review. They follow from the framework’s normative commitments to reflective endorsement, autonomy, accessibility, fair participation, and transparent evidence. Where they extend beyond leisure-specific evidence—especially summary verification and labour governance—that broader basis is stated explicitly.

9.1 Design for purpose-sensitive delegation#

Systems should ask what the user wants to preserve, not only what they want completed. A writing assistant could distinguish brainstorming from ghostwriting; a game assistant could distinguish repetitive execution from strategic decisions; a video system could distinguish navigation from replacement viewing. Default settings should favour assistance where the central process is likely to be intrinsically valued.

Purpose sensitivity can be implemented through explicit modes, configurable boundaries, and legible action histories. It should not depend on a system inferring sensitive preferences without consent. Users should be able to change the boundary as their goals change.

9.2 Make delegation selective, reversible, and inspectable#

Beneficial delegation requires meaningful control. Here that is operationalized through five observable capabilities: set the delegation boundary; inspect actions and provenance; revise or reject outputs; reverse or stop system action; and resume the activity without avoidable loss of skill or access. “Human in the loop” is insufficient if the person can only accept or regenerate an opaque output.

For writing, this means preserving drafts and identifying generated contributions. For games, it means bounded permissions and clear separation between assistive and autonomous actions. For video, it means timestamps, source links, and an easy transition from a claim to its audiovisual evidence. Evaluation can combine these behavioural checks with perceived autonomy and competence, psychological ownership, enjoyment, oversight workload, and the user’s reflective endorsement after experience.

Deployment choices belong to the same assessment. Evaluators should record where prompts and drafts are processed, whether model versions and configurations can be retained or exported, applicable licence terms, and the labour required to maintain the arrangement. Local deployment can improve data custody and continuity for some users, while a managed service can reduce technical burdens for others; neither should serve as a proxy for meaningful control or psychological ownership.

9.3 Protect intrinsically valued process#

Leisure interfaces should not treat every delay, difficulty, or duration as inefficiency. Challenge, anticipation, and unhurried attention can be features. Designers should provide options to disable optimization prompts and avoid presenting skipped activity as an achievement equivalent to participation.

A useful design test is: if the user receives the output without performing the process, would they still have obtained what they wanted? If the answer is no, full delegation is not an appropriate default.

9.4 Treat accessibility as a central design requirement#

Accessibility should be built into delegation architectures from the outset. Configurable assistance, multiple input and output modalities, adjustable pacing, and shared control can expand participation. Evaluation should involve disabled users and measure perceived control and enjoyment, not only task completion.

Accessibility accommodations should not be collapsed into general anti-automation rules. Where competitive fairness is relevant, rules should identify which functions affect contest integrity and which enable equivalent participation.

9.5 Preserve provenance and social legibility#

Users should be able to represent the nature of their contribution accurately. Writing and puzzle communities can require disclosure appropriate to the activity, while avoiding burdensome reporting of trivial assistance. Game platforms should distinguish designer-authorized automation, external bots, accessibility support, and labelled computer opponents.

The purpose of provenance is not to stigmatize new practices. It is to make comparisons, expectations, and consent meaningful. Curatorial AI creativity and delegated play can be legitimate when participants understand what is being shared or evaluated.

Communities should also assess whether output escalation is making participation less voluntary. Provisional safeguards include separating recognition from volume, continuing to welcome short or low-polish contributions, preserving roles that do not require extensive preparation, and making any comparison categories explicit. These are untested governance options, not established remedies; disclosure requirements should also account for accessibility, privacy, and the burden of documenting trivial assistance.

9.6 Design summaries for verification#

Video and other summarization systems should:

• connect claims to timestamps or transcript passages;
• distinguish direct observation from inference;
• expose query and personalization criteria;
• support adjustable detail and alternative selections;
• communicate uncertainty without presenting every fluent statement as equally reliable;
• facilitate source viewing when the cost of error is high.

Evaluation should measure omission as well as hallucination and should test fitness for defined user goals, not only generic representativeness.

9.7 A normative labour-governance position#

The following is a normative policy position, not a leisure-specific empirical finding. It holds that productivity gains should be open to distribution as pay, reduced hours, or greater schedule control, and that economic compulsion should not be misread as revealed welfare improvement.

The paid-work question cannot be solved by interface design alone. Employers and policymakers should monitor total workload, after-hours availability, work intensity, and discretionary time, rather than assuming that adoption benefits workers because task speed improves. Where productivity gains are substantial, employees should have meaningful opportunities to share them through pay, reduced hours, or greater schedule control.

Organizational assessments should ask whether AI introduces new tasks and monitoring, whether output expectations change, and who decides how savings are used. Protecting rest and leisure is compatible with allowing individuals to choose additional paid work, but evaluations should distinguish that choice from work compelled by hidden escalation or economic necessity. Extra hours sold under financial pressure should not be interpreted as revealed proof of improved welfare.

9.8 A decision checklist for users and evaluators#

Before delegating a leisure component, six questions provide a practical test:

1. Purpose: What combination of output, process, social recognition, and shared participation do I value?
2. Boundary: Which specific component is unwanted, and which must remain mine for the activity to retain its value?
3. Control: Can I inspect, revise, stop, or reverse the system’s actions, and who controls my data, model version, and continued access?
4. Others: Does delegation affect fairness, consent, accessibility, or community expectations, and could my output raise pressure on others to automate?
5. Net time: After setup, verification, and monitoring, what time and attention are actually saved?
6. Allocation: Who will control the saved time, and is additional paid work genuinely my preferred use?

This checklist operationalizes the thesis without assuming that one answer fits every activity.

10. Discussion#

10.1 Answering the research questions#

10.2 Theoretical implications#

The analysis extends SDT-based accounts by showing that the same automation can either support or frustrate psychological needs. The effect depends on which component is delegated. Removing a motor barrier can support competence; removing all decisions can undermine it. Automating an unwanted task can increase autonomy; opaque personalization can decrease it. Delegation is thus not external to motivation but part of how the activity is structured.

The analysis also qualifies the constitutive-obstacle account of games. Delegation can dissolve an obstacle, but some games deliberately delegate progress and create play around configuration, waiting, and return. Suits’s emphasis on voluntary constraints remains useful, provided that the relevant rule system includes idle mechanics and the player’s attitude toward them.

Bainbridge’s oversight irony appears in a new form. The user is not merely supervising a safety-critical system but evaluating outputs against experiential and contextual criteria: voice, fairness, relevance, or completeness. These judgments can be difficult precisely when the user has delegated the practice that develops them. Maintaining skill may therefore require periods of unassisted participation even where automation remains available.

10.3 A sociotechnical rather than purely philosophical conclusion#

It would be possible to defend manual leisure as inherently authentic and reject delegation on philosophical grounds. The evidence does not support that simplicity. Tools have always mediated leisure, and automation can remove exclusion and drudgery. The problem is not artificiality but misalignment between system action, user purpose, and social context.

Equally, it would be possible to argue that users reveal their preferences whenever they automate, making external evaluation unnecessary. This position overlooks uncertain expectations, persuasive design, workplace pressure, and harms to other participants. A choice to automate can be genuine without being fully informed or welfare-enhancing. Sociotechnical analysis must hold together individual preference, technical behaviour, and institutional constraint.

10.4 “Experiential replacement” as an interpretive claim#

Experiential replacement is not a technical capability that a benchmark could establish, nor a mode of system operation alongside assistance and full delegation. It is a claim about meaning: an output is treated as evidence that its recipient has undergone a valued activity. Users, fan communities, platform metrics, marketers, and productivity narratives may all make or contest such claims for different reasons.

Analysing the term does not presuppose that every claim of replacement is false or dishonest. A person may say they have “watched” a video as shorthand for knowing its contents, or may genuinely prefer curation and anticipation to conventional participation. The analytical questions are what the claim means in context, who advances it, and what social or economic work it performs. This keeps the debate at the level of user interpretation and social convention rather than requiring metaphysical claims about machine experience.

10.5 Social and communal leisure#

The case studies necessarily foreground individual experience, but much leisure is constituted through reciprocal interaction. Collaborative storytelling, multiplayer roleplay, fan communities, watch parties, and social viewing can make contribution, recognition, and shared attention more important than either a private experience or a standalone output. Automation may support these practices through translation, accessibility, coordination, or preparation. It may also alter reciprocity when participants delegate the contributions on which a shared practice depends.

The same intervention can therefore have different communal effects. A generated premise may supply material for collaborative storytelling, while fully delegated participation may remove a member’s expected contribution; a summary may help someone join a watch-party discussion without reproducing the shared viewing itself.

Social rebound adds a collective temporal mechanism: one participant may save preparation time while inadvertently helping to raise the community’s expected standard of preparation. The resulting pressure cannot be inferred from output volume alone. It requires evidence about norms, recognition, perceived obligation, and the experiences of participants who use little or no AI. A fourth research direction, alongside individual experience, technical reliability, and time allocation, should consequently examine how delegation redistributes participation, recognition, and belonging within leisure communities.

10.6 Future research#

Future studies should examine longitudinal leisure delegation rather than one-off output quality. Useful designs include:

• diary studies distinguishing active enjoyment, delegated anticipation, monitoring, and subsequent time use;
• comparisons of assistance, partial delegation, and full delegation within the same hobby;
• matched local and cloud configurations measuring psychological ownership, perceived control, capability, data custody, and maintenance labour rather than inferring them from deployment location;
• studies of automation builders that treat design and debugging as possible leisure while measuring their full time cost;
• measures of voice, perceived authorship, competence, and diversity over repeated use;
• community studies testing H3 through output norms, recognition, perceived obligation, and the experiences of low- and non-AI participants;
• participatory research with disabled users on shared control and fair-play accommodation;
• video-summary studies measuring both decision accuracy and changes in viewing experience;
• research connecting recommendation and ranking systems with summarization, including effects on exposure, autonomy, and total viewing time;
• linked task-time and whole-day time-use measures to identify rebound and oversight costs;
• labour research connecting specific delegation practices to contracts, bargaining power, income adequacy, economic necessity, and discretionary time.

Research should report model versions and dates because capability snapshots age quickly. It should also avoid assuming that more paid work is the desirable endpoint; welfare, rest, autonomy, and distribution require separate measurement.

11. Conclusion#

AI and bots can produce stories, progress through game environments, return puzzle answers, and condense videos. These capabilities have practical value, particularly for bounded information needs and for removing components users experience as friction. They do not establish that the experience associated with an activity has been transferred.

That conclusion is primarily conceptual. Where participation defines the value sought, output receipt and process participation are different goods. The empirical part of the thesis is more limited: adjacent studies show useful writing assistance, uneven technical reliability, diverse community rules, and legitimate practices built around delegated execution. Idle games, TAS, automation construction, curatorial creativity, shared control, and spoiler research prevent a categorical claim that automation must diminish leisure. No reviewed study directly demonstrates that partial, reversible delegation preserves leisure experience better than full delegation over time.

The leisure-to-paid-work pathway is likewise unestablished. Net usable minutes must survive delegation overhead, and their use depends on labour opportunities, preferences, care, income security, living costs, and institutional power. Workplace studies show task absorption and varying time allocations but do not identify what people do with time released from hobbies. Closer leisure-quality research shows that attractive leisure can itself draw time away from paid work.

The defensible proposal is therefore modest: delegate components whose loss the user reflectively endorses, preserve values that remain important, and assess the quality as well as the quantity of time released. Whether additional paid work improves welfare depends on preference, necessity, and distribution. The thesis’s closing normative position is explicit rather than value-neutral: rest, autonomy, and leisure are legitimate ends, so additional paid work should not be treated as the presumptively best use of every saved hour.

Appendix A: Iterative Research and Revision Record#

Appendix B: Evidence Interpretation Guide#

References#

Aguiar, M., Bils, M., Charles, K. K., & Hurst, E. (2021). Leisure luxuries and the labor supply of young men. Journal of Political Economy, 129(2), 337–382. https://doi.org/10.1086/711916

Ahmetovic, D., Manzoni, M., Corti, F., & Mascetti, S. (2026). Video game accessibility through shared control for people with upper-limb impairments. arXiv:2601.11218 (preprint). https://arxiv.org/abs/2601.11218

Anderson, B. R., Shah, J. H., & Kreminski, M. (2024). Homogenization effects of large language models on human creative ideation. Proceedings of the 16th ACM Conference on Creativity & Cognition. https://doi.org/10.1145/3635636.3656204

Apostolidis, E., Adamantidou, E., Metsai, A. I., Mezaris, V., & Patras, I. (2021). Video summarization using deep neural networks: A survey. Proceedings of the IEEE, 109(11), 1838–1863. https://doi.org/10.1109/JPROC.2021.3117472

Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098%2883%2990046-8

Bick, A., Blandin, A., & Deming, D. J. (2026). The rapid adoption of generative AI. Management Science, 72(10), 8440–8456. https://doi.org/10.1287/mnsc.2025.02523

Biermann, O. C., Ma, N. F., & Yoon, D. (2022). From tool to companion: Storywriters want AI writers to respect their personal values and writing strategies. Designing Interactive Systems Conference 2022, 1209–1227. https://doi.org/10.1145/3532106.3533506

Blizzard Entertainment. (n.d.). End User License Agreement. Retrieved 29 October 2026, from https://www.blizzard.com/en-us/legal/fba4d6f2-5f46-4d24-9f79-1f9b401e2b3d/blizzard-end-user-license-agreement

Chess.com. (n.d.). Fair play policy. Retrieved 29 October 2026, from https://www.chess.com/legal/fair-play

Cowan, R. S. (1983). More work for mother: The ironies of household technology from the open hearth to the microwave. Basic Books.

Csikszentmihalyi, M. (1990). Flow: The psychology of optimal experience. Harper & Row.

Cutting, J. T., Gundry, D. E., & Cairns, P. A. (2019). Busy doing nothing? What do players do in idle games? International Journal of Human-Computer Studies, 122, 133–144. https://doi.org/10.1016/j.ijhcs.2018.09.006

Deci, E. L., & Ryan, R. M. (2000). The “what” and “why” of goal pursuits: Human needs and the self-determination of behavior. Psychological Inquiry, 11(4), 227–268. https://doi.org/10.1207/S15327965PLI1104_01

Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. https://doi.org/10.1126/sciadv.adn5290

Fizek, S. (2018). Interpassivity and the joy of delegated play in idle games. Transactions of the Digital Games Research Association, 3(3), 137–163. https://doi.org/10.26503/todigra.v3i3.81

Fu, C., Dai, Y., Luo, Y., Li, L., Ren, S., Zhang, R., Wang, Z., Zhou, C., Shen, Y., Zhang, M., Chen, P., Li, Y., Lin, S., Zhao, S., Li, K., Xu, T., Zheng, X., Chen, E., He, R., & Sun, X. (2025). Video-MME: The first-ever comprehensive evaluation benchmark of multi-modal LLMs in video analysis. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 24108–24118. https://openaccess.thecvf.com/content/CVPR2025/html/Fu_Video-MME_The_First-Ever_Comprehensive_Evaluation_Benchmark_of_Multi-modal_LLMs_in_CVPR_2025_paper.html

Guo, A., Sathyanarayanan, S., Wang, L., Heer, J., & Zhang, A. X. (2025). From pen to prompt: How creative writers integrate AI into their writing practice. Proceedings of the 2025 ACM Conference on Creativity & Cognition, 527–545. https://doi.org/10.1145/3698061.3726910

Huizinga, J. (1955). Homo ludens: A study of the play-element in culture. Beacon Press. (Original work published 1938)

Humlum, A., & Vestergaard, E. (2026). Still waters, rapid currents: Early labor market transformation under generative AI. NBER Working Paper No. 33777 (issued May 2025; revised March 2026). https://doi.org/10.3386/w33777

Jiang, W., Park, J., Xiao, R. (J.), & Zhang, S. (2025). AI and the extended workday: Productivity, contracting efficiency, and distribution of rents. NBER Working Paper No. 33536. https://doi.org/10.3386/w33536

Keogh, B., & Richardson, I. (2018). Waiting to play: The labour of background games. European Journal of Cultural Studies, 21(1), 13–25. https://doi.org/10.1177/1367549417705603

Keynes, J. M. (2010). Economic possibilities for our grandchildren. In Essays in persuasion. Palgrave Macmillan. (Original work published 1930). https://doi.org/10.1007/978-1-349-59072-8_25

Leavitt, J. D., & Christenfeld, N. J. S. (2011). Story spoilers don’t spoil stories. Psychological Science, 22(9), 1152–1154. https://doi.org/10.1177/0956797611417007

Loredo Lopez, A. Y., McDonald, T., & Emami, A. (2025). NYT-Connections: A deceptively simple text classification task that stumps System-1 thinkers. Proceedings of the 31st International Conference on Computational Linguistics, 1952–1963. https://aclanthology.org/2025.coling-main.134/

Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. https://doi.org/10.1126/science.adh2586

Peronikolis, M., & Panagiotakis, C. (2024). Personalized video summarization: A comprehensive survey of methods and datasets. Applied Sciences, 14(11), 4400. https://doi.org/10.3390/app14114400

Porter, B., & Machery, E. (2024). AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably. Scientific Reports, 14, 26133. https://doi.org/10.1038/s41598-024-76900-1

Rawal, R., Shirkavand, R., Huang, H., Somepalli, G., & Goldstein, T. (2025). ARGUS: Hallucination and omission evaluation in Video-LLMs. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 20280–20290. https://doi.org/10.1109/ICCV51701.2025.01886

Ryan, R. M., Rigby, C. S., & Przybylski, A. (2006). The motivational pull of video games: A self-determination theory approach. Motivation and Emotion, 30, 344–360. https://doi.org/10.1007/s11031-006-9051-8

Samadarshi, P., Mustafa, M., Kulkarni, A., Rothkopf, R., Chakrabarty, T., & Muresan, S. (2024). Connecting the dots: Evaluating abstract reasoning capabilities of LLMs using the New York Times Connections word game. arXiv:2406.11012 (preprint). https://arxiv.org/abs/2406.11012

Sharghi, A., Laurel, J. S., & Gong, B. (2017). Query-focused video summarization: Dataset, evaluation, and a memory network based approach. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 4788–4797. https://arxiv.org/abs/1707.04960

SIMA Team et al. (2024). Scaling instructable agents across many simulated worlds. arXiv:2404.10179 (technical report). https://arxiv.org/abs/2404.10179

SIMA Team et al. (2025). SIMA 2: A generalist embodied agent for virtual worlds. arXiv:2512.04797 (technical report). https://arxiv.org/abs/2512.04797

Suh, D., & Oh, S. (2026). Generative AI and the reallocation of time: Productivity, leisure, and fulfilling work. arXiv:2602.12695 (preprint). https://arxiv.org/abs/2602.12695

Suits, B. (1978). The grasshopper: Games, life and utopia. University of Toronto Press.

TASVideos. (n.d.). Movie rules. Retrieved 30 October 2026, from https://tasvideos.org/MovieRules

U.S. Copyright Office. (2025). Copyright and artificial intelligence, Part 2: Copyrightability. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf

Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., & Anandkumar, A. (2023). Voyager: An open-ended embodied agent with large language models. arXiv:2305.16291 (technical report). https://arxiv.org/abs/2305.16291

Wilson, T. D., & Gilbert, D. T. (2005). Affective forecasting: Knowing what to want. Current Directions in Psychological Science, 14(3), 131–134. https://doi.org/10.1111/j.0963-7214.2005.00355.x

Yan, D., & Tsang, A. S. L. (2016). The misforecasted spoiler effect: Underlying mechanism and boundary conditions. Journal of Consumer Psychology, 26(1), 81–90. https://doi.org/10.1016/j.jcps.2015.05.003