What AI Added to Goal Setting: Accountability

One result in a March 2026 paper deserves more attention than the headline. An AI career coach helped people report more progress on their goals than no support did. Yet the clearest clue came from a harder comparison, against people who had already completed a serious written reflection exercise. What distinguished the AI experience was social accountability: participants felt that the system would notice whether they followed through.

That question matters for anyone building with conversational AI. A fluent exchange can produce a polished plan, but polish tells us little about what happens after the tab closes. The more useful question is whether conversation changes a person’s relationship with the goal once ordinary life resumes.

Research note: This paper is a March 2026 preprint. It has not undergone peer review.

A stronger comparison than AI versus nothing

Michel Schimpf, Julian Voigt, and Thomas Bohné ran a preregistered three-arm randomised controlled trial with employed adults in the United Kingdom and United States. Participants used a custom mobile app to set three career-related goals for the coming month.

The questionnaire made this a useful active-control design. Both supported groups reflected on broadly similar topics, while the AI condition added a responsive conversation. Participants also rated the AI and questionnaire conditions as similarly structured, so the comparison gives us more information than an AI-versus-empty-screen test could.

The researchers randomised 517 people. The primary analysis used 323 participants who completed the two-week follow-up and passed the attention checks, equal to 62.5% of everyone randomised. Keeping those numbers separate matters because 517 describes the reach of randomisation, while 323 describes the sample behind the main estimates.

What changed

Two weeks later, participants rated their progress across the three goals. The measure asked whether they had made progress, felt on track, and felt they had achieved the goal.

The AI group reported more progress than the no-support control, with a small-to-moderate standardised effect of d = 0.33 and p = .016. The written-reflection group also scored above control, although that comparison did not reach conventional statistical significance (d = 0.24, p = .076).

AI and written reflection did not differ significantly on overall progress (d = 0.08, p = .540). That result blocks the easy sales line. The study supports a short-term AI advantage over no help; it does not show that conversation produced more progress than a well-structured writing exercise.

An exploratory analysis adds useful context. Perceived structured reflection statistically mediated both active conditions over control, while it did not explain the difference between AI and the questionnaire. The authors treat reflection as a plausible shared ingredient. That conclusion needs further testing, but it fits the pattern in the primary results: asking people to examine their situation may already do meaningful work.

Why accountability matters

The AI condition separated more clearly on perceived accountability. Compared with structured reflection, it increased the feeling of being answerable for later progress by d = 0.43. Compared with control, the effect was d = 0.68.

In the preregistered mediation model, accountability carried the AI-over-questionnaire contrast to later progress. The indirect effect was 0.15, with a 95% confidence interval from 0.04 to 0.31. Self-concordance, the extent to which a goal reflects autonomous rather than pressured motives, did not mediate the result and did not differ across conditions.

Mediation does not establish one complete causal chain. It does provide stronger mechanism evidence than a story assembled after the results, because the researchers named accountability in advance and measured it before the follow-up outcome.

The finding also sharpens what “social” can mean in software. Participants knew they were speaking with an AI, yet a responsive exchange still created some sense of an evaluative audience. Follow-up questions, contingent replies, and an invitation to return may make a goal feel witnessed. The paper connects this interpretation to earlier work on computers as social actors and to supportive accountability, where clear expectations and a credible, helpful relationship can support continued engagement.

The trial tested a moment; Annet supports a loop

The intervention concentrated on one goal-setting session, followed by a reminder roughly two weeks later. Annet covers more of the time between those points. A goal can gather evidence through check-ins and through four different activity surfaces, each suited to a different kind of reflection or action.

• Meditation creates a short pause to settle attention, set an intention, and record what shifted.
• Journaling gives experience somewhere to become explicit. Answers to a goal’s open questions can feed back into that goal’s reasoning.
• Planning turns an outcome into visible steps, likely blockers, and timing.
• Mind-mapping lays relationships out spatially, then lets a useful branch continue into a plan.

These activities remain connected to the goals they touch. Recent saved artifacts can become context for later goal reasoning, so reflection does not have to disappear when a session ends. Check-ins add return points, while weekly review creates a slower cadence for looking across the record. People can also invite a Witness or Accountability Partner when human support fits the goal.

This wider loop may cover several gaps left by a single coaching conversation. Someone can need a clearer plan, space to name an obstacle, a way to see tangled commitments, or enough calm to choose the next action. The RCT did not test these activity modes or their combined effect. Their contribution remains a product hypothesis that deserves direct measurement.

How Annet applies this

Annet is short for Accountability Network. The name describes an architecture built around returning: goals keep context, reflection can become usable evidence, AI guidance can respond to what changed, and people can join the loop when the user chooses.

The Cambridge-led trial gives that thesis relevant mechanism evidence, especially for felt accountability during a conversational goal-setting process. It did not test Annet, its activity system, its continuing AI context, or its human-support features.

A research agenda, not borrowed certainty

The weakest use of this paper would be as a validation badge. One preprint, conducted with one intervention over two weeks, cannot establish that a broader product works. A better use is to make the product thesis easier to test.

Annet’s wider surface gives us more mechanisms to study, and more ways to be wrong. Useful questions include whether goal-linked journal answers improve later return behaviour, whether planning contributes beyond reflection, when meditation helps someone re-engage after overload, and whether mind-mapping changes decisions or merely creates a feeling of clarity. Human support raises another comparison: when does accountability from a chosen person add something an AI relationship cannot?

Those questions move evaluation beyond satisfaction scores and attractive plans. A research-led product should measure whether people return, whether the support still feels credible, and which forms of reflection help under which conditions.

What this study cannot tell us

The paper remains a preprint without peer review. Progress was self-reported after about two weeks, so the study cannot tell us whether the effect lasts or appears in objective behaviour. The analytic sample retained 323 of 517 randomised participants. Although later dropout did not differ by condition, the questionnaire’s enforced 20-minute minimum caused substantial early attrition and made the two active experiences less comparable.

The sample included English-speaking employed adults aged 18 to 50 in the United Kingdom and United States. The study used one AI configuration, did not compare AI with a human coach, and asked for career-related goals even though the AI group often surfaced goals outside that domain. The questionnaire matched broad reflective topics rather than isolating one conversational ingredient.

The next useful move is therefore specific: measure both the quality of the plan and whether a person feels expected back. Then watch what happens when they return.

References

• Schimpf, M., Voigt, J., & Bohné, T. (2026). AI-Assisted Goal Setting Improves Goal Progress Through Social Accountability. arXiv preprint arXiv:2603.17887.
• Schimpf, M., Voigt, J., & Bohné, T. (2026). Preregistration for the three-arm randomised controlled trial.
• Harkin, B., Webb, T. L., Chang, B. P. I., et al. (2016). Does monitoring goal progress promote goal attainment? A meta-analysis of the experimental evidence. Psychological Bulletin, 142(2), 198-229.
• Mohr, D. C., Cuijpers, P., & Lehman, K. (2011). Supportive accountability: A model for providing human support to enhance adherence to eHealth interventions. Journal of Medical Internet Research, 13(1), e30.