AI-assisted work makes pay feel less 'earned'—participants reported lower effort and ownership when AI generated or selected output—but a randomized lab study found no clear effect on how those small payments were invested or spent.

Main Finding

AI assistance (from augmented suggestions to full substitution) reliably reduces people’s subjective effort, psychological ownership, perceived self-contribution, and the felt “earnedness” of income. However, in this one-shot lab experiment with small stakes (10 RMB endowment), those perceptual changes did not produce statistically significant differences in incentive-compatible investment behavior, hedonic vs. utilitarian spending choice, or measured risk preference; preregistered mediation through effort and ownership was not supported. The authors interpret this as clear evidence that AI changes earning perceptions, and as bounding — not disproving — behavioral effects at this stake size.

Key Points

• Manipulation: three conditions (Manual, Augmented, Substituted) across three rounds of a real-effort copywriting task. Augmented = participant selects and can edit among 3 LLM-generated headlines; Substituted = LLM auto-generates and auto-submits; Manual = participant types headline.
• Earnings: task completion produced a fixed 10 RMB endowment for everyone; participants then made a real-stakes investment decision over that 10 RMB and completed consumption/risk measures.
• Strong perceptual effects: monotone, large reductions (Manual → Augmented → Substituted) in subjective effort, psychological ownership, perceived self-contribution, AI attribution increase, and felt earnedness decrease. Scales: effort (adapted NASA-TLX, α = .825), ownership (adapted, α = .924).
• Null behavioral results: no significant differences by condition in (a) fraction invested in a 50% chance ×2.5 gamble, (b) hedonic vs. utilitarian consumption choice, or (c) self-rated or staircase risk-preference measures. Bootstrap mediation analyses (effort and ownership as parallel mediators) produced confidence intervals that included zero.
• Estimates were directionally consistent with the mental-accounting (AI → more windfall-like → more risk/hedonic spending) prediction but were imprecise and below the study’s minimum detectable effect; the data therefore rule out medium-to-large effects at this stake size but cannot exclude small effects.
• Mental-accounting categorization (a four-way earned↔windfall item) shifted toward “windfall” under AI assistance, but the item’s anchors partially restated the manipulation, so this shift is not an independent behavioral proof.
• Design and preregistration: between-subjects, randomized within-session; Helmert contrasts used (D1: Manual vs. AI-pooled; D2: Augmented vs. Substituted); preregistered hypotheses and analysis plan; main analytic sample n = 179 (convenience sample from a Chinese university participant pool; ~75% female). Average final payout ≈ 12 RMB (range 0–25 RMB). LLM used for headline generation: DeepSeek-V3 via API.

Data & Methods

• Sample: main confirmatory N = 179 (61 Manual, 59 Augmented, 59 Substituted). Additional pilot batch (n = 23) collected prior to preregistration; pooled robustness checks reported.
• Procedure: online scheduled group sessions (oTree + Tencent Meeting). Three copywriting rounds with condition-specific feedback badges; then mediator scales, filler, incentive-compatible investment gamble, binary hedonic/utilitarian choice, mental-accounting items, GPS and staircase risk tasks, demographics, manipulation checks.
• Incentive-compatible gamble: participants chose what percent of the 10 RMB to invest (0–100% slider) in a one-shot lottery: 50% chance to multiply invested amount by 2.5, otherwise lose invested portion; uninvested portion kept. Real payment via Alipay.
• Measures:
  • Subjective effort: 3-item adapted NASA-TLX (α = .825).
  • Psychological ownership: 4-item adapted scale (α = .924).
  • Perceived self-contribution: numeric 0–100%.
  • AI attribution: 1–5 Likert.
  • Mental-accounting: 4-category earned↔windfall classification + 1–7 continuous windfall-feel rating.
  • Risk preference: GPS qualitative item and a staircase-based CE risk task.
  • Consumption: binary hedonic vs. utilitarian choice.
• Analysis: Helmert contrasts for primary hypotheses, ANOVA/Cohen’s d for effect sizes, bias-corrected bootstrap mediation (5,000 resamples). Minimum detectable effects reported in paper; behavioral estimates underpowered relative to small effects.
• Robustness/limitations noted by authors: quality-bonus announcement (up to 50 RMB) may have altered subjective expectations; completion fee (10 RMB) disclosed prior to task (could dampen “earnedness” even in Manual); session clustering checked; demographic imbalance mostly female and university students (external generalizability limited).

Implications for AI Economics

• Perceptions matter: AI assistance attenuates the cues (felt effort, authorship, ownership) that underlie earnedness. If these perceptual shifts generalize outside the lab, mental-accounting and identity-linked behaviors tied to “earned” income could be affected by AI adoption.
• Do perceptions translate to economic choices? This experiment finds no reliable one-shot behavioral effects at small stakes, implying that policy or managerial claims about spending, saving, or risk-taking shifts due to AI should not be inferred solely from self-reports. Field-scale earnings, repeated compensation, larger stakes, or different institutional framing may be necessary to generate measurable behavioral changes.
• Design of compensation and incentives: firms and platforms might need to consider how AI-mediated work is framed and rewarded. If workers (or consumers) feel less ownership/earnedness, that could affect effort, satisfaction, willingness to invest proceeds, or how bonuses are used. Explicit attribution, editability, or co-authorship signals might restore ownership cues.
• Labor-market and welfare implications: any macroeconomic effects of AI on household consumption, savings, or risk-taking require that perception changes scale up to sustained, economically meaningful income flows. The present bounds suggest medium-to-large effects are unlikely at small stake and one-shot settings; estimating real-world impacts requires field experiments or longitudinal data.
• Policy and platform recommendations:
  • Transparency and labeling: indicate worker contribution or co-authorship status to preserve ownership/earnedness cues where desirable.
  • Bonus and quality-pay design: make the tie between worker effort and reward salient (e.g., performance-contingent bonuses, certifications of contribution) to counteract dilution of earnedness.
  • Monitor possible “taint” effects: if AI causes earnings to be seen as illegitimate rather than merely unearned, spending patterns could shift toward utilitarian (opposite of classic windfall predictions).
• Research agenda:
  • Test with larger, recurring payments and in real workplaces (field experiments, payroll-level treatments).
  • Vary framing: emphasize human contribution, enable edits, or display co-authorship metadata to test mitigation strategies.
  • Measure longer-run outcomes: savings rates, durable good purchases, investment behavior, identity and job satisfaction, turnover.
  • Explore heterogeneity: effects by task type (creative vs. routine), worker role, cultural context, and baseline AI familiarity.
  • Decompose valence: distinguish “unearned” from “illegitimate/tainted” money to predict direction of consumption effects.
• Methodological caution: do not infer behavioral change from self-report alone. Perceptual measures are necessary but insufficient; incentive-compatible behavioral measures and adequate power at realistic stakes are required to draw economic inferences.

Summary takeaway: AI assistance reliably reduces felt effort and ownership — altering the psychological signals that, in theory, drive mental accounting — but this experiment found no robust downstream spending or risk-taking effects at small stakes in a one-shot lab setting. Translating perceptual shifts into economic policy conclusions requires field-scale, longitudinal, and higher-stakes evidence.