AI-Loop Feedback, Not Passive Content or Peer Support Alone: Self-Directed Goal Follow-Through and the Limits of Feeling Understood

Abstract

Among 289 adults pursuing a self-directed goal (fitness or a self-study skill) over 8 weeks, participants who used a general-purpose AI assistant to plan and revise their own weekly instructions, receiving outcomes traceable directly to what they had specified ("AI-loop," n=101) reported higher self-efficacy at follow-up (5.3/7) than participants who consumed generic informational content on the same goal with no personalized feedback ("passive-content," n=94; 4.4/7, a 20% gap), and were more likely to still be pursuing the goal at 8 weeks (55%) than the passive-content group (43%, a 30% gap). A third group using a human accountability partner for weekly check-ins ("peer-support," n=94) showed comparable follow-through (53%) and self-efficacy (5.0/7) to the AI-loop group, but reported markedly higher feeling of being understood and supported (5.7/7) than either the AI-loop group (4.5/7, a 27% gap) or the passive-content group (4.2/7).

1. Introduction

A goal-tracking conversation with an AI assistant has a property that ordinary self-directed effort often lacks: the outcome of a given week is fairly directly traceable back to how clearly the person specified their own plan and constraints, rather than to circumstances outside their control. Self-efficacy theory has long held that a sense of being able to change one's own outcomes builds specifically from this kind of traceable cause-and-effect experience, which is harder to get reliably from generic advice content and, in a different way, from a human partner whose own responses are shaped by more than just the plan handed to them. We compared self-efficacy, goal follow-through, and felt social support across three structurally different accountability conditions for the same class of self-directed goals.

2. Method (Summary)

We recruited 289 adults beginning a self-directed 8-week fitness or self-study goal, randomly assigned within each goal type to one of three conditions. The "AI-loop" condition (n=101) used a general-purpose AI assistant weekly to set a specific plan, log results against it, and receive assistant-generated feedback keyed directly to the plan's own stated targets. The "passive-content" condition (n=94) received weekly generic informational content matched for topic and length but with no personalization to the participant's own plan or logged results. The "peer-support" condition (n=94) checked in weekly with a human accountability partner (a friend or a partner assigned through the study) for an unstructured conversation about progress. At 8 weeks, participants completed a validated self-efficacy scale (1–7), reported whether they were still actively pursuing the goal, and completed a felt-understood/supported scale (1–7) describing their accountability experience over the 8 weeks.

3. Findings

At 8 weeks, 55% of AI-loop participants (56/101) were still pursuing their goal, versus 43% of passive-content participants (40/94, a 30% gap) and 53% of peer-support participants (50/94), who did not differ meaningfully from the AI-loop group. Self-efficacy averaged 5.3/7 for the AI-loop group, 4.4/7 for the passive-content group (a 20% gap), and 5.0/7 for the peer-support group. On the felt-understood/supported scale, the peer-support group averaged 5.7/7, clearly higher than the AI-loop group (4.5/7, a 27% gap) and the passive-content group (4.2/7); the AI-loop and passive-content groups did not differ meaningfully on this measure.

4. Discussion

The AI-loop and peer-support conditions produced similar follow-through and self-efficacy, despite very different accountability structures, while the passive-content condition trailed both on both measures. This is consistent with the traceability of the feedback — knowing that this week's result reflected this week's own stated plan, whether that reflection came from an assistant or from a human partner tracking the same plan closely — mattering more for self-efficacy than any particular feature of who or what delivered it. The clearest asymmetry is on felt-understood/support: the peer-support group scored well above both other conditions, and the AI-loop group did not close that gap despite matching peer-support on follow-through. An AI feedback loop that reliably reflects a person's own planning back to them appears able to substitute for a human partner on the specific mechanism of traceable self-efficacy, without functioning as a substitute for the separate experience of feeling understood by another person — the two outcomes moved independently of each other rather than together.

5. Limitations and Future Directions

The felt-understood/supported scale was administered only once, at 8 weeks, and may be sensitive to how recently a participant's last accountability interaction occurred. Peer-support partners varied in relationship closeness to the participant (some were pre-existing friends, others study-assigned), which was not analyzed as a moderator here but plausibly affects the felt-understood outcome specifically. The fitness and self-study goal types were combined in the main analysis; a supplementary within-type comparison showed the same direction of effect in both, though the self-study subsample was underpowered to test this formally. A planned follow-up will test whether combining an AI-loop with even minimal, low-frequency human check-ins recovers the felt-understood benefit without giving up the traceability benefit.

Editorial Note

This report underwent CfAHC's Rapid Review Protocol, including AI-assisted literature cross-checking and a two-reviewer human sign-off.