← Back to the overview
Evidence

The evidence, in full

Each study below links to its source. We give the numbers, say what kind of study it was, and list its limits.

The honest line: these are findings from published research, not Sikivu's own results. We are pre-launch and have measured no outcomes of our own yet. When we do, every number will carry its confidence level and its source.

The ingredients

The therapies and ways of delivering them that Sikivu is built on, with what the research does and does not show.

Grounded in CBT, ACT and behavioural activationUmbrella review

An umbrella review of 39 meta-analyses found that internet-delivered therapy reduces symptoms across a wide range of conditions, with effects from small to large.

A 2024 umbrella review pooled 39 meta-analyses of internet-delivered CBT, acceptance and commitment therapy (ACT) and behavioural activation in adults. Against control conditions, pooled effects ranged from SMD 0.10 (stress in employees) to SMD 1.20 (depressive symptoms in older adults).

Limits

  • The reviewers note that many of the included meta-analyses had methodological weaknesses.
  • This is evidence about the therapies, not about Sikivu. We have no outcome data of our own yet.
Käll, Biliunaite & Andersson. Digital Health, 2024 ↗Source last re-read 2026-10-09
Works outside the labStudies without a control group

In routine care, guided internet CBT was acceptable to patients, and on average 2.9% of people got worse.

A 2020 review pooled 19 studies (30 groups) of guided internet CBT for depression and anxiety delivered in routine clinical settings. Uptake, adherence and satisfaction indicated moderate to high acceptability, and average deterioration across studies was 2.9%.

Limits

  • The studies had no control group, so improvement cannot be separated from natural recovery. The authors advise choosing interventions proven in randomised trials.
  • These were "guided" programmes in clinical settings, not an AI app.
  • A 2025 correction fixed the abstract’s pooled depression effect size (1.78 to 1.18). This page does not rely on that figure.
Etzelmueller et al. JMIR, 2020 ↗Source last re-read 2026-10-09
Guidance helps, most when symptoms are worseMeta-analysis

Both guided and unguided internet CBT beat control conditions. Guidance added a small extra benefit that was gone by 6 and 12 months.

An individual patient data analysis of randomised trials (36 trials and 8,107 participants analysed) found that guided internet CBT beat unguided at the end of treatment, with PHQ-9 scores 0.8 points lower (95% CI 0.2 to 1.4). There was no difference at 6 or 12 months. Baseline severity was the strongest modifier: the advantage of guidance was small for mild symptoms (PHQ-9 5 to 9) and grew for moderate, moderately severe and severe depression.

Limits

  • The long-term comparison rests on 8 studies.
  • Confidence intervals are wide for people with mild symptoms.
  • The authors report weak evidence of publication bias.
  • Guidance here means human support, not an AI coach.
Karyotaki et al. JAMA Psychiatry, 2021 ↗Source last re-read 2026-10-09
Tested in AfricaCluster randomised trial

In a Zimbabwe trial of therapy from trained lay health workers, 2.3% of people reported suicidal ideation at 6 months, versus 12.3% with enhanced usual care.

In 24 primary care clinics in Harare (573 participants), six sessions of individual problem-solving therapy from trained lay health workers, with an optional peer group, led to lower symptoms at 6 months. The primary outcome, the SSQ-14 symptom score, was 3.81 versus 8.90. Suicidal ideation, a secondary outcome, was 2.3% (6 people) versus 12.3% (32 people). Sikivu Circles are inspired by this lay-worker and peer-group model. They are not the same intervention.

Limits

  • Suicidal ideation was a secondary outcome, not the primary one.
  • Zimbabwe is in Southern Africa, not East Africa.
  • The comparison was enhanced usual care, not usual care.
  • The trial tested in-person therapy from lay health workers, not an app.
Chibanda et al. JAMA, 2016 ↗Source last re-read 2026-10-09
Structured, measured careMeta-analysis

Across 29 randomised trials, care built on measurement and collaboration raised response and remission rates for depression.

A 2021 meta-analysis of 29 randomised trials (15,255 participants) found that enhanced evidence-based care raised response (RR 1.30, 95% CI 1.13 to 1.50) and remission (RR 1.35, 1.11 to 1.64) compared with usual care. Of the 29 trials, 21 tested collaborative care and 27 ran in the US or Europe. The measurement-based care and algorithm-guided subgroup alone showed no significant difference in response (RR 1.19, 0.97 to 1.47) but did for remission (RR 1.28, 1.06 to 1.54). We track PHQ-9 and GAD-7 so that we can learn whether Sikivu helps. This evidence does not show that measurement alone will raise Sikivu’s outcomes.

Limits

  • Heterogeneity was high (I² above 80% for both pooled results).
  • Most trials were collaborative care in US and European clinics, not digital self-help.
Xiao et al. Translational Psychiatry, 2021 ↗Source last re-read 2026-10-09

The AI conversation

The closest published trial of a generative AI conversation for mental health, and the limits of what it tells us.

An AI conversation, testedRandomised trial

A US trial of a generative AI chatbot found larger symptom reductions than a waitlist, with clinicians reviewing every reply.

In a 2025 randomised trial, 210 US adults with clinically significant depression, generalised anxiety or high risk of an eating disorder used the Therabot chatbot for four weeks (106 people) or waited (104). Depression scores fell by 6.13 points versus 2.63 on the waitlist at 4 weeks. Effect sizes (Cohen’s d, 4 and 8 weeks) were 0.845 to 0.903 for depression and 0.794 to 0.840 for generalised anxiety.

Limits

  • The comparison was a waitlist, so the trial cannot separate the chatbot from the effect of getting any attention or support.
  • Trained clinicians reviewed every reply after it was sent. They contacted participants 15 times about safety concerns and 13 times to correct inappropriate replies.
  • People with active suicidality were excluded.
  • All participants were US adults recruited online. The authors note possible selection bias toward younger, more technology-minded people and call for larger trials.
  • The intervention lasted four weeks, with follow-up to eight.
Heinz et al. NEJM AI, 2025 ↗Source last re-read 2026-10-09
What the evidence does not cover

We did not find a randomised trial of a generative AI mental health companion in East Africa.

On 9 October 2026 we searched ClinicalTrials.gov and Europe PMC for trials of a chatbot, large language model or generative AI tool for mental health in Kenya, Uganda, Tanzania, Rwanda and Ethiopia. We found no randomised trial. We found a 50-person proof-of-concept study in Uganda, a co-design study in Kenya and a cross-sectional survey in Tanzania, all unreviewed preprints. The AI trial above ran in the United States.

  • We did not search the Pan African Clinical Trials Registry or ISRCTN, so an absence in our searches is not proof that no such trial exists. If you know of one, please tell us. Contact us
Searched 2026-10-09

Other trials and reviews we read

Further trials and reviews we read, including ones in Africa and ones with weaker or mixed results.

Lay providers in KenyaRandomised trial

In Kenya, a four-session group programme led by trained lay providers reduced adolescent depression and anxiety more than a study-skills control.

In four Nairobi and Kiambu schools, 413 adolescents aged 13 to 18 with elevated symptoms had four weekly group sessions led by lay providers aged 18 to 26 who had 10 hours of training. Compared with a study-skills group, depression improved by d 0.35 and anxiety by d 0.37 at 4 weeks, and by d 0.45 and d 0.44 at 7 months.

Limits

  • Only 54% of participants completed the 7-month assessment.
  • The control group also improved.
  • The programme developers are affiliated with the funder and ran the trial.
  • The depression measure omits the suicide item.
Osborn et al. JAMA Psychiatry, 2021 ↗Source last re-read 2026-10-09
Guided digital supportRandomised trial

In Lebanon, a WHO digital behavioural-activation programme with weekly non-specialist contact reduced depression among Syrian refugees.

In a trial of 569 displaced Syrians with depression, the Step-by-Step programme, delivered by app or web with up to 15 minutes of weekly phone or message contact from a non-specialist, reduced PHQ-9 depression scores more than enhanced care as usual (g 0.48, 95% CI 0.26 to 0.70 after treatment, and g 0.61 at three months).

Limits

  • Only 53.8% returned the post-test questionnaire and 32.2% of the intervention group completed all sessions.
  • The article’s own attrition figures are inconsistent, and a correction was published in 2023.
  • The trial ran in Lebanon with a displaced population, not in East Africa.
Cuijpers et al. PLoS Medicine, 2022 ↗Source last re-read 2026-10-09
A chatbot trial in AfricaRandomised trial

In Malawi, a scripted CBT chatbot reduced depression scores among health workers. The anxiety result was not statistically significant.

In an eight-week trial among health workers in Blantyre and Lilongwe, the Vitalk chatbot reduced PHQ-9 depression scores more than a passive resource page (difference-in-differences −0.68, 95% CI −1.15 to −0.21). The anxiety difference (−0.44, 95% CI −0.88 to 0.01) included zero.

Limits

  • Retention was 29% in the chatbot group and 40% in the control group, and the analysis used complete cases.
  • The Malawi version had no natural language understanding, so it was not a generative AI companion.
  • Participants were health workers, not the general public, and the control group was waitlisted.
  • The authors’ causal wording goes further than the design supports.
Kleinau et al. PLoS One, 2024 ↗Source last re-read 2026-10-09
An early conversational CBT trialRandomised trial

An early trial of a rule-based CBT chatbot found lower depression scores after two weeks than an information-only control.

In a 2017 trial of 70 US college students with symptoms of depression and anxiety, two weeks of the Woebot chatbot lowered PHQ-9 scores more than an information ebook (d 0.44). There was no between-group difference in anxiety (d 0.14).

Limits

  • The trial was unblinded, small and had no follow-up.
  • The control was information only, so the trial cannot separate conversational delivery from CBT content.
  • One author founded the company that made the chatbot.
Fitzpatrick, Darcy & Vierhile. JMIR Mental Health, 2017 ↗Source last re-read 2026-10-09
Chatbot trials, pooledMeta-analysis

A 2025 meta-analysis of 31 trials found small to moderate effects of AI chatbots on distress in people aged 15 to 39, with generative systems still inconclusive.

The review pooled randomised trials of AI chatbots in people aged 15 to 39 (31 trials, 29,637 participants). Mental distress fell by SMD −0.35 (95% CI −0.46 to −0.24), depressive symptoms by −0.43 and anxiety by −0.37. Retrieval-based systems had consistent effects. For generative systems the authors judged the overall effectiveness inconclusive.

Limits

  • We read the published abstract, not the full paper.
  • The age range is 15 to 39. It does not cover older adults.
  • The authors call for safety protocols and evaluation frameworks for generative systems.
Feng et al. JMIR, 2025 ↗Source last re-read 2026-10-09 (abstract only)
Generative AI chatbots, pooledMeta-analysis

A 2025 meta-analysis of 14 trials of generative AI chatbots found a pooled effect of 0.30, with a wide margin of uncertainty.

The review pooled 14 randomised trials (6,314 participants) of generative AI chatbots and found a pooled effect of 0.30 (95% CI 0.004 to 0.59; 95% prediction interval −0.85 to 1.67).

Limits

  • The interval runs from almost no effect to a large one, and the prediction interval includes both harm and benefit.
  • The protocol was registered after the review began.
  • With 14 trials, the finding that social-oriented chatbots outperformed task-oriented ones is exploratory.
  • We read the abstract, not the full paper.
Zhang et al. JMIR, 2025 ↗Source last re-read 2026-10-09 (abstract only)

Risks we are designing against

Research on how AI chatbots can fail people in distress. We cite it as a reason for our safeguards, not as evidence that Sikivu works.

Crisis handling in chatbotsReported research

Scale AI reported that in about 35% of 718 simulated crisis conversations, chatbots recognised distress but did not point the user to help.

Nineteen licensed clinicians and crisis counsellors wrote 718 realistic chats simulating people in crisis, and Scale AI tested 25 frontier models against them. TIME reported the results on 9 October 2026.

Limits

  • This is company research shared with TIME, not a peer-reviewed study.
  • The conversations were simulated, and no per-model breakdown was published in the article.
TIME, 9 October 2026 (research by Scale AI) ↗Source last re-read 2026-10-09
Ethics in therapy-style chatsReported research

Brown University researchers identified 15 ethical risks in how prompted language models handle therapy-style conversations.

Seven CBT-trained peer counsellors held chats with CBT-prompted language models, and three licensed psychologists reviewed a subset. The researchers grouped 15 risks into five categories: lack of contextual adaptation, poor therapeutic collaboration, deceptive empathy, unfair discrimination, and lack of safety and crisis management.

Limits

  • The chats were simulated from real counselling sessions, and the news release does not say how many chats were analysed.
  • The release does not state whether the work has been peer reviewed.
  • The study tested prompted general-purpose models, not specific consumer products.
Brown University news release, 21 October 2025 ↗Source last re-read 2026-10-09
Chatbots and psychosisMedical record review

A Vanderbilt record review found 28 patients whose clinical notes met criteria for psychosis linked to AI chatbot use.

A research letter reviewed 578,058 records (215,712 patients) at one academic medical centre and identified 28 patients who met criteria for “AI psychosis”. Seventeen of the 28 (60.7%) were in a first psychotic episode.

Limits

  • The study cannot show that chatbot use caused the psychosis.
  • It covers a single site, with ascertainment bias and a rating system that has not been clinically validated.
  • The paper describes its prevalence denominator inconsistently, so we do not quote a rate.
Bergson et al. JAMA Psychiatry, 2026 ↗Source last re-read 2026-10-09