A two-week randomized trial tested Resonance, an AI journaling tool that connected people’s past memories with possible future activities. Here is what the researchers measured, what the results mean, and where the evidence stops.
In a two-week randomized controlled study, 55 participants completed the required study activities. The research team compared 28 people using Resonance with 27 people in a control condition. Resonance generated action-oriented ideas from each participant’s own logged memories and asked them to imagine carrying out a suggestion.
The paper reports that 71 people started and 55 completed the study. Participants were asked to journal daily, so the results describe people who completed a demanding two-week protocol, not everyone who might download an app and use it casually.
The trial was published in the proceedings of the Augmented Humans International Conference 2025. Read the published paper from ACM or its open preprint. The MIT Media Lab project page describes the AI-powered journal.
For the 28 participants in the Resonance group, the mean PHQ-8 score went from 6.11 before the study to 4.96 after it. The 27-person control group went from 6.11 to 6.14. The researchers’ mixed-effects model found a statistically significant decrease in the Resonance group and no significant change in the control group.
Daily positive affect also increased more when participants said they were more likely to act on a suggestion. The paper reports a small, immediate change around the interaction; it does not establish a lasting mood effect.
PHQ-8 scores are symptom-screening scores, not a diagnosis. On commonly used severity bands, 0–4 is none to minimal and 5–9 is mild; a score of 10 or more is a screening threshold that may indicate current depression and calls for clinical interpretation. Those bands describe scores, not people’s diagnoses. See the U.S. National Center for Health Statistics explanation of PHQ-8 score bands.
A larger and more varied sample would help show whether the result holds across different people and settings. Two weeks cannot show whether any changes lasted after people stopped using the tool. And because the trial tested Resonance, its results cannot be assumed to apply to Reflect Nightly or another app.
The trial provides an early signal, with limited evidence about how well it generalizes.
The study did not show whether any changes lasted after participants stopped using the tool.
The trial did not evaluate Reflect Nightly or its features.
The group mean started in the mild symptom band and ended just below it. Averages do not reveal every participant’s experience, and crossing a symptom-band boundary does not establish recovery or treatment. The study is best read as an early signal that deserves replication, not proof that AI journaling treats depression.
A 2026 individual-patient-data meta-analysis estimated that an individual PHQ-8 score generally needs to change by about 5.5 points to be confident, at the 95% level, that the change is beyond measurement error. That is a benchmark for interpreting change in an individual’s score. It is not a cutoff for a group-average change in a randomized study.
So the trial’s 1.15-point mean decrease in the Resonance group should not be dismissed by comparing it directly with the 5.5-point individual threshold. The trial used a mixed-effects analysis across participants and reported a statistically significant group change, but the small sample and short duration still leave uncertainty about size, durability, and practical importance. Read the 2026 BMJ meta-analysis on detectable change in PHQ scores for the measurement details.
“A real signal is a reason to keep studying a question, not a treatment claim.”
What one small trial can tell us — and what it cannot.Resonance used past memories to suggest future activities and invited participants to imagine doing them. Its suggestions were rated as more effective when they felt personal, novel, and connected to logged memories. That is a specific design pattern, not a general finding that all AI-generated reflection helps.
Reflect Nightly is designed around reviewing the day and noticing grounded wins or next steps from what someone actually shared. The MIT study did not evaluate our app, and we do not claim its results for Reflect Nightly. Our product is a wellness app, not therapy or treatment, and it is not a substitute for professional mental-health care.
If you are struggling, consider reaching out to a qualified professional or a local support service.
No. It found a statistically significant short-term change in one small trial of one specific tool. The result needs further study and does not establish a treatment effect.
The study is relevant to how we think about connecting reflection with a next step. It did not test Reflect Nightly, and we make no medical claims based on it.
The paper is published in the Augmented Humans 2025 proceedings, with an open preprint on arXiv.
Reflect Nightly offers a low-pressure way to talk through the day and notice what you want to carry into tomorrow.
Get Reflect Nightly on Google Play