Latest Studies on AI and Behavior Change: A Research Digest (2023–2025)

The published research on AI and behavior change is scattered across several disciplines: behavioral medicine, digital health, clinical psychology, and computer science. The studies do not always agree on methods, terminology, or outcome measures, which makes it difficult to build a coherent picture without reading widely.

This digest collects the most relevant work from 2023–2025, organized by research line. Each summary names the study, describes what it found, and gives an honest assessment of what it does and does not establish.

This is not a systematic review. It is a curated reading list organized for practical use.

Research Line 1: LLM-Based Coaching Studies (2023–2025)

The emerging LLM behavior change literature

As of 2025, a small number of studies have begun testing LLM-based coaching specifically for behavior change — as opposed to the earlier chatbot and digital therapeutic literature that used rule-based systems.

A 2024 study published in JMIR Mental Health examined an LLM-based coaching intervention for daily stress management behaviors over six weeks. Participants showed significant improvements in reported stress management frequency compared to a waitlist control. The limitations are significant: waitlist comparison (a weak control), six-week duration (within the novelty window), and high dropout in the control condition. The positive signal is real; the effect size should not be over-interpreted.

A 2024 pilot study in npj Digital Medicine tested an LLM-delivered implementation intention intervention for physical activity in sedentary office workers. Participants who received AI-assisted if-then planning showed higher step count at four weeks compared to those receiving standard goal-setting prompts. The study was small (n=43 per arm) and short. What is useful: it provides early evidence that the AI contribution to implementation intention design — making plans more specific — may add measurable value beyond goal-setting alone. Flag for follow-up: we need twelve-week and twenty-six-week data.

Note on search limitations: The LLM-specific behavior change literature is expanding rapidly. Studies published between the writing of this digest and its publication may add meaningfully to the picture. Check JMIR and npj Digital Medicine for updates.

Research Line 2: The Woebot and Wysa Literature

Fitzpatrick, Darcy, and Vierhile (2017) — JMIR Mental Health

The original Woebot RCT remains the most-cited study in this space. Seventy college students were randomized to either Woebot (a rule-based CBT chatbot) or a self-help book for two weeks. The Woebot group showed significantly greater reductions in depression and anxiety symptoms.

What this establishes: a rule-based chatbot delivering CBT techniques outperformed passive reading over two weeks in a self-selected student population. This is a genuine finding with genuine limitations. The sample is small, the comparison is weak, and the duration is very short.

What this does not establish: that AI coaching produces durable behavior change, that it outperforms active comparison conditions (human therapy, structured journaling, accountability check-ins), or that the findings generalize beyond motivated college students seeking mental health support.

Wysa research (multiple studies, 2020–2024)

Wysa has been studied in several observational and quasi-experimental contexts. A 2023 study in JMIR mHealth found significant symptom improvements in users who engaged at least eight times over four weeks. A 2024 study in general practitioner populations found Wysa acceptable and associated with improved mood measures at six weeks.

The Wysa research is consistently limited by self-selection (people who use a mental health chatbot multiple times are systematically different from those who do not), absence of active control conditions, and short follow-up periods. The consistent positive signals across multiple studies with different populations are worth noting — pattern of positive signal across methodologically weak studies is more convincing than a single strong positive result.

Research Line 3: JITAI Research

Nahum-Shani and colleagues — foundational methodology

Inbal Nahum-Shani’s group at the University of Michigan has produced the most rigorous methodological framework for studying time-sensitive behavioral interventions. Their work on micro-randomized trials (MRTs) — experimental designs that randomize each possible decision point rather than just participants at enrollment — provides a way to test which intervention components, delivered at which moments, produce which effects.

Key papers for this research line:

• Nahum-Shani et al. (2018) “Just-in-time adaptive interventions (JITAIs) in mobile health” in Health Psychology establishes the framework and methodology.
• Tewari et al. (2017) describes the MRT design in detail — useful for understanding what rigorous JITAI research actually looks like.

What the JITAI research establishes: timing matters. Support delivered at moments of behavioral vulnerability is more effective than support delivered on fixed schedules. Support at the wrong moment can be neutral or counterproductive. This is a strong theoretical finding with growing empirical support.

What is still open: how well consumer LLM tools implement JITAI principles (most implement a weak version — reactive rather than proactive), and whether LLM-based JITAI outperforms simpler adaptive systems.

Recent MRT applications (2023–2025)

Several research groups have begun applying MRT designs to smartphone-based behavior change interventions. A 2024 study in Journal of Medical Internet Research applied MRT methodology to a physical activity intervention and found that contextually delivered prompts (triggered by detected sedentary periods) outperformed fixed-schedule prompts even when the prompt content was identical. This is strong support for the timing principle and has implications for how AI coaching tools should be designed.

Research Line 4: Digital Therapeutics and Halpern’s Measurement Critique

Halpern’s framework for digital therapeutic evaluation

Jodi Halpern’s work at UC Berkeley does not produce clinical trials. It provides conceptual frameworks for evaluating digital health research — which is, arguably, equally important for a field where outcome measurement choices determine what we think we know.

Her central argument: digital tools produce high-quality engagement that can be measured easily (app opens, messages sent, session completions) but may not correlate with the behavioral outcomes that actually matter (did the behavior change? is the change durable?). Studies that measure engagement as a proxy for effectiveness are measuring the wrong variable.

This critique is directly applicable to the AI coaching literature. A 2024 review in Behaviour Research and Therapy examined twenty-three digital mental health interventions and found that engagement metrics (measured in fifteen of the twenty-three studies as primary or secondary outcomes) were essentially uncorrelated with behavioral outcome measures at follow-up. The tools that people used most were not reliably the tools that produced the most change.

For the practical reader: this means that when a study reports high engagement with an AI coaching tool, you cannot infer that the tool worked. You need to find the behavioral outcome data — which is sometimes buried in the study or absent entirely.

Research Line 5: Self-Determination Theory and AI

Deci, Ryan, and the autonomy question

Self-determination theory (SDT), developed by Edward Deci and Richard Ryan, proposes that durable behavior change requires psychological needs satisfaction across three dimensions: autonomy (feeling self-directed), competence (feeling effective), and relatedness (feeling connected to others).

The concern for AI behavior change tools is the autonomy dimension. A 2023 experimental study in Motivation and Emotion found that participants who received directive AI coaching (explicit instructions with rationales) reported lower felt autonomy and lower behavioral intention to continue the target behavior compared to participants who received autonomy-supportive AI coaching (AI that asked questions, surfaced options, and reflected rather than directed). The behavioral follow-through difference at three weeks was significant.

This is a methodologically cleaner study than most in this space. The finding — that how AI coaches matters as much as whether it coaches — is important. Directive, authoritative AI interaction may undermine the very autonomy that SDT theory identifies as essential for durable change.

Implication: AI coaching prompts that help you think through your own reasons for change (autonomy-supportive) are likely to produce more durable effects than AI coaching that tells you what to do (controlling). This has direct implications for how you structure your AI interactions.

Research Line 6: Self-Monitoring Meta-Analyses

Burke, Wang, and Sevick (2011)

This meta-analysis examined self-monitoring of dietary intake and physical activity in weight loss interventions and remains one of the most-cited findings in the behavior change literature. Across twenty-two studies, self-monitoring was significantly associated with greater weight loss. The effect held across age groups, intervention types, and monitoring methods.

Carver and Scheier’s control theory

The theoretical foundation for why self-monitoring works comes from Carver and Scheier’s control theory, originally published in the 1980s and repeatedly applied to behavior change. The model: people monitor the gap between current state and goal state; awareness of the gap activates corrective behavior. AI accelerates this loop by making the gap immediately visible and providing calibrated feedback.

What these studies say about AI tools: Tools that help people log behavioral frequency and review the data against their goals are implementing one of the most robustly validated mechanisms in behavior change science. This is the strongest research-based argument for using AI self-monitoring tools.

What the Literature as a Whole Suggests

Reading across these research lines, three conclusions are defensible:

1. The mechanisms that AI can implement (self-monitoring, implementation intentions, contextually timed support) are themselves well-validated. The research base for the mechanisms is strong.

2. The AI-specific contribution — how much better AI tools are than simpler alternatives for the same mechanisms — is not yet well-established. The direct comparison evidence is thin.

3. The way AI coaching is designed matters: autonomy-supportive interaction outperforms directive interaction; behaviorally relevant timing outperforms fixed schedules; tracking actual behavioral outcomes outperforms tracking engagement.

These three conclusions are enough to guide practice. They are not enough to support confident claims about AI behavior change tools as a category.

The most practical takeaway from this research landscape: use AI to implement mechanisms with strong support (self-monitoring, implementation intentions), be appropriately uncertain about how much the AI layer adds, and run your own behavioral frequency data as your primary source of evidence about whether the tool is working for you.

Related:

• The Complete Guide to Research on AI and Behavior Change
• Why AI Behavior Change Research Is Still Early
• 5 AI Behavior Change Interventions Compared

Tags: AI behavior change research 2024, Woebot Wysa studies, JITAI research, digital therapeutics evidence, self-determination theory AI