Where is AI actually being used in public health? A review found only eight reports
Researchers screened 5,400 records and 155 full texts but found only eight reports describing AI used in public-health practice relevant to high-income countries outside outbreaks. The evidence was mixed and usually did not test effectiveness.
The short version
- What happened
- The review searched four databases for reports from January 2020 to 24 July 2025, screened 5,400 records and assessed 155 full texts.
- Who it affects
- Public-health decisions affect whole communities, so weakly evaluated tools can scale missed needs, confusing advice or unequal access. Practitioners need service outcomes and failure reporting, not only technical demonstrations.
- What remains unproven
- Scoping review
Keep in mind
Scoping review
Eight included reports
High-income contexts
Outbreaks excluded
Search ended 24 July 2025
Mixed source designs
No pooled effectiveness estimate
Most reports lacked robust evaluation
Source analysis · 1 linked source. Primary analysed source date: 29 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The scarcity—not a performance score—is the main finding
A peer-reviewed scoping review found a strikingly thin public record of AI being used in public-health practice. The researchers searched MEDLINE, Embase, Web of Science and Scopus for material published from January 2020 to 24 July 2025. Two reviewers independently screened 5,400 records and 155 full texts. Only eight reports met the definition of an AI tool used in practice in a public-health context relevant to high-income countries, outside outbreak response.
Those eight reports covered three chatbot uses, one machine-learning forecast of health-service demand, two symptom-checking tools and two uses of AI in evidence synthesis. The review deliberately required evidence that a tool had reached practice, a pilot or another public-facing use. A laboratory comparison or proposed use without practical exposure did not qualify. That makes the result useful for procurement: a large research literature does not necessarily translate into inspectable operating evidence.[1]
Mixed examples do not establish an average benefit
The included reports did not support one overall effectiveness estimate. One chatbot was described as helpful by users; another was judged ineffective and capable of undermining trust in the organisation behind it. A large-language-model system used with socially isolated people showed benefits and drawbacks. Most reports did not robustly evaluate effectiveness, cost, unequal effects or unintended harm, and the review included commentaries and editorials as well as primary studies.
Public-health teams therefore cannot treat 'AI' as a single intervention. A demand forecast, symptom checker and conversational service create different error paths and need different comparators. Before adoption, teams should specify the population reached, existing workflow, eligible denominator, accuracy or service outcome, staff checking burden, cost, complaints and equity effects. Publishing negative and inconclusive deployments matters because otherwise the next authority sees only vendor demonstrations and successful pilots.[1]
The distinction that matters
How 5,400 records became eight practice reports
The review separated proposed or laboratory uses from tools that had entered public-health practice or a public-facing pilot.
Records
5,400 screened
Broad database search for AI and public-health practice
Full texts
155 assessed
Reports checked against the practice-focused criteria
Included
8 reports
Mixed uses and little robust effectiveness evidence
The review is deliberately narrow
The search excluded outbreaks through a protocol refinement made before searching, and it focused on contexts relevant to high-income countries. That improves relevance to UK practice but limits global generalisation and omits a large emergency-response literature. Data extraction was completed by one reviewer and checked by another. The study was funded by UK public research programmes, and the authors declared no conflicts of interest.
The assessment would change with prospective, independently reported deployments that publish pre-specified outcomes and costs. For now, the defensible conclusion is modest: public-health agencies may be experimenting with AI, but the accessible evidence describing what was used, for whom and with what result remains too sparse to infer routine benefit.[1]