TL;DR
While Large Language Models (LLMs) can generate structurally flawless "SMART" goals, delegating the goal-writing process to AI severely undermines our motivation to achieve them. This paper demonstrates a sharp "quality-motivation dissociation": AI-authored goals lead to a massive drop in psychological ownership and a 26% decrease in actual behavioral follow-through, especially among individuals who already struggle with self-efficacy.
What Problem Does This Paper Address?
AI productivity tools and coaches frequently generate plans, surface tasks, and articulate goals on behalf of users. Because LLMs excel at turning vague intentions into highly specific, measurable, and time-bound (SMART) objectives, it seems intuitive that they would improve user outcomes.
However, classical goal-setting theory assumes that goals are self-authored. When an AI writes a personal goal, it mediates the narrative of who the user wants to become. The authors investigate a critical gap in Human-Computer Interaction (HCI): what happens to human motivation, commitment, and real-world behavior when we delegate the cognitive and emotional labor of goal formulation to an AI?
Methods
The authors conducted a preregistered, between-subjects experiment with 470 U.S. adults. The study proceeded as follows:
- Baseline & Reflection: All participants completed a trait self-efficacy scale and wrote a personal reflection on their past year.
- Random Assignment:
- Self-authored condition: Participants wrote three of their own SMART goals based on their reflection.
- LLM-authored condition: An LLM (OpenAI's GPT-5.2) generated three SMART goals based on the participant's reflection. To standardize physical effort across groups, these participants retyped the AI-generated goals verbatim.
- Evaluation: Subjective motivational metrics (psychological ownership, commitment, importance, clarity) were collected immediately. An LLM judge (Gemini 3.1 Pro) objectively evaluated the structural quality of all goals against a SMART rubric.
- Follow-up: Two weeks later, a follow-up survey ($N=332$) assessed free-recall accuracy and whether participants had taken concrete actions toward their goals.
Datasets
Not discussed in the paper. The study relies entirely on a newly collected behavioral dataset from human subjects via Prolific, rather than pre-existing machine learning benchmarks.
Key Insights and Ideas
- The Quality-Motivation Dissociation: There is a fundamental mismatch between how good a goal looks on paper and how motivated a person is to pursue it. AI makes goals structurally perfect but psychologically empty.
- Bypassing the Ownership Routes: Psychological ownership relies on three routes: investment of self, intimate knowing, and control. Writing a goal naturally satisfies all three. Delegating the task to an AI bypasses them entirely, causing users to treat the goal as something "received" rather than "internalized."
- The Self-Efficacy Paradox: Users low in trait self-efficacy—the very people most likely to seek out AI assistance for productivity—suffered the steepest erosion in psychological ownership when using AI-generated goals. AI strips away agency from the users who need it most.
Key Results
- Structural Quality: LLM-generated goals were dramatically better structured, scoring significantly higher on the objective SMART rubric ($d = -2.26$).
- Motivational Deficit: Despite their structural flaws, self-authored goals led to much higher psychological ownership ($d = 1.38$), commitment ($d = 1.19$), and perceived importance ($d = 1.13$) at baseline. These differences persisted at the two-week follow-up.
- Behavioral Follow-through: At the two-week mark, 72.8% of participants in the self-authored condition had acted on two or more of their goals, compared to just 46.6% in the LLM condition—representing a massive behavioral gap.
- Memory Retention: Self-authored participants recalled their goals significantly more accurately, proving that the process of writing a goal supports cognitive encoding better than surface-level text richness.
Limitations and Future Work
- Full Generation vs. Co-Authorship: The study tested a full-delegation model where the AI wrote the complete initial draft. The authors note that future work should explore "co-authorship" or partial scaffolding designs (e.g., AI editing a human draft or walking a user through dimensions one by one) to see if they can preserve ownership while capturing AI's structural benefits.
- Measurement Window: The behavioral follow-through was tracked over a relatively short two-week window and relied on self-reported actions. Longer longitudinal studies are needed to see if users eventually internalize AI goals over time.
- Personal Fit: Goal quality was evaluated purely via the formal SMART rubric; alternative measures like alignment with personal values or long-term fit were not explored.
Summary updated