Are We Automating the Joy Out of Work? Designing AI to Augment Work, Not Meaning

Abstract: Prior work has mapped which workplace tasks are exposed to AI, but less is known about whether workers perceive these tasks as meaningful or as busywork. We examined: (1) which dimensions of meaningful work do workers associate with tasks exposed to AI; and (2) how do the traits of existing AI systems compare to the traits workers want. We surveyed workers and developers on a representative sample of 171 tasks and use LMs to scale ratings to 10,131 computer-assisted tasks across all U.S. occupations. Worryingly, we find that tasks that workers associate with a sense of agency or happiness may be disproportionately exposed to AI. We also document design gaps: developers report emphasizing politeness, strictness, and imagination in system design; by contrast, workers prefer systems that are straightforward, tolerant, and practical. To address these gaps, we call for AI whose design explicitly focuses on meaningful work and worker needs, proposing a five-part research agenda.

• The paper finds that tasks high in agency and creativity are most exposed to AI, challenging assumptions that automation only targets routine labor.
• It employs a four-phase design combining stratified sampling, surveys, and LM-based annotation to capture nuanced worker and developer perspectives.
• The findings emphasize that misaligned AI trait design—favoring politeness over practicality—risks eroding meaningful work experiences across sectors.

Automating Agency and Joy: AI Exposure, Misalignment, and Design Implications

Study Motivation and Design

The study interrogates how current and near-term AI exposure intersects with perceived meaningfulness in daily work and identifies the trait-level misalignment between what workers want from AI augmentation and what developers intend to implement. The research combines stratified task sampling from O*NET, survey-based annotation by both workers and developers, and large-scale extension via LM-assisted ratings, achieving coverage across all major U.S. occupational sectors.

Figure 1: Four-phase study design—task selection, multidimensional worker ratings, psychological trait elicitation, and large-scale LM annotation.

The sampling pipeline filtered tasks to those primarily completed digitally and performed with high frequency, then further stratified by occupational sector and AI exposure indices. Worker and developer cohorts were recruited through rigorous screening and matched to their domain, with survey items covering dimensions from perceived value and bullsh*t to well-being and flourishing. Developers and workers rated desired AI traits (e.g., straightforwardness, tolerance, practicality) for augmentation scenarios. LM-based annotation procedures were validated against human intra-class correlation, achieving moderate-high agreement.

Figure 2: Sector distribution comparison across Prolific, LM-annotated, and BLS datasets confirms high fidelity of LM-annotated coverage.

AI Exposure and Meaningful Work

The core empirical result is a robust, contradictory finding: tasks with high worker-perceived agency, happiness, and creativity are disproportionately likely to be exposed to AI augmentation, while tasks anchored in social connection, in-person interaction, and emotional awareness remain less exposed. Mixed-effects regression over the meaningfulness dimensions yielded significant gaps (FDR $<0.05$, $\Delta \geq 0.1$ Likert points), challenging prevailing assumptions that automation targets only routine or low-value labor.

Figure 3: Differential AI exposure gaps for meaningfulness—the most exposed tasks are high in novelty, autonomy, and positive affect; the least exposed are high in social/emotional connection.

Sectoral decomposition shows creative and scientific domains (Arts, Community Service, Education, Life/Physical/Social Sciences) consistently associate exposed tasks with high novelty, autonomy, and positive affect.

Figure 4: Creative/socially-oriented sectors rank highest in novelty, positivity, autonomy for exposed tasks; routine/manual sectors are lowest.

By contrast, less exposed tasks (education, social service, healthcare) are heavily weighted toward relationship-building, emotional attunement, and face-to-face engagement.

Figure 5: Human-facing sectors consider not-exposed tasks as socially/emotionally salient; routine/technical sectors rate these dimensions as negligible.

Worker–Developer Trait Misalignment in AI Design

Trait-level analysis—instantiated via explicit Likert comparisons and cross-sectoral clustering—reveals strong and systematic misalignment between worker preferences and developer design intentions. Workers consistently prefer AI augmentation traits such as straightforwardness, tolerance, and practicality, especially for high-agency or production tasks. Developers favor politeness, strictness, and imagination, particularly in high-structure or high-stakes domains.

Figure 6: Top misaligned traits (straightforward vs polite, tolerant vs strict, practical vs imaginative) and top aligned traits (definitive vs open to challenge, personalized, and comprehensive), with sector/task exemplars.

Quantitatively, the trait of straightforward vs polite yields the largest misalignment, with negative worker–developer gaps exceeding 1.6 points in production and engineering settings. Sector-specific divergence is apparent: in technical and oversight domains, politeness is experienced as unnecessary friction rather than useful support (Figure 7). Developers, often influenced by generalized interface conventions and regulatory considerations, default to politeness, strictness, and imaginative interaction while workers valorize concise, actionable, and flexible output.

Figure 7: Strong misalignment for straightforward vs polite traits, especially in production and engineering sectors.

Trait alignment is observed only on personalization, insightfulness, and openness-to-challenge, where both cohorts prefer deep comprehension and adaptability over generic or rigid systems.

Methods for Large-Scale Annotation and Validation

The study leverages LM-as-expert annotation for scaling to 10K tasks, employing chain-of-thought prompting and persona conditioning validated by ICC and MAD metrics. Analysis of major discrepancies shows LM annotations typically diverge from humans only when functional/procedural interpretations conflict with social or emotional readings of tasks—a minority of cases. Replication with human-only data confirms that core exposure, trait gap, and sector cluster findings are robust and not artifacts of model bias.

Figure 8: Distribution comparison of meaningful work dimensions between LM and Prolific users by sector reveals similar central tendencies, but richer human variance.

Theoretical and Practical Implications

The study advances the field by providing task-level granularity on the intersection of AI exposure and experienced meaning, supported by multidimensional annotation and cross-modal scaling. Findings undermine the notion that augmentation automatically liberates workers for higher-value labor: generative systems risk absorbing the most creative and agency-rich activities, relegating humans to reactive editing or quality-control roles. Trait misalignment signals persistent gaps in human-centered design and amplifies resistance or friction in practice.

The practical implications are considerable:

• Interaction Design: Systems should offer explicit control for workers to preserve agency and authorship in high-exposure tasks and minimize friction introduced by redundant politeness or arbitrary strictness.
• Sector-Specific Defaults: Technical and oversight domains require concise, straightforward, practical augmentation; social domains benefit from warmth and personalization, not rigid automation.
• Governance and Participation: Current power asymmetries allow management and technical teams to unilaterally determine exposure and trait profiles; more participatory mechanisms (worker voice, sectoral consultation) are needed to avoid “mass demoralization.”
• Measurement and Deployment: Deployment should capture not only productivity metrics but also longitudinal effects on worker agency, creative engagement, and relational work, informed by direct in-product survey and log analysis.

Theoretical extensions include new models for the distinction between augmentation and automation—operationalized via meaningfulness dimensions—and sector/task modulated trait design, moving toward a formal framework for value alignment at the human-AI interface.

Future Directions

Speculative avenues include algorithmic customization of augmentation levels and trait settings adapted to individual, team, and sector preferences, real-time monitoring of agency and creativity loss, and dynamic interface controls for “intent-preserving” AI. Longitudinal studies on demoralization, skill retention, and creative output in augmented workflows will be critical. Incorporation of union/worker representation in AI exposure governance could counteract management/technical biases and lock in meaningful work as a central objective, not an afterthought.

Figure 9: Developer cohort profile across technical expertise, AI usage, and sectoral function—a necessary precursor to trait gap analysis.

Conclusion

The study rigorously demonstrates that current AI exposure disproportionately targets high-agency and joyful work, with persistent misalignment between worker preferences and developer design intentions for AI augmentation traits. These patterns are sector- and task-dependent, contradict conventional narratives and expose structural risks for loss of meaning and demoralization in the workplace. The findings underscore the necessity for worker-centered design, sector-aware trait defaults, participatory governance, and multifaceted measurement in the deployment of AI augmentation tools (2603.14963).