A 2025 Science Advances study reported that low-dose metformin's glucose-reducing effect in mice depended on Ras-related protein 1 (Rap1) in the brain. The experiments did not test whether the pathway has the same importance in people receiving routine oral treatment. They instead add one candidate mechanism to a drug whose biological action has remained the subject of continuing research.

The paper matters less as a verdict on metformin than as an example of how evidence accumulates. A measured outcome, a proposed mechanism, and an applicable patient population are separate claims, and each requires a different test.

The mouse study identifies one pathway

The researchers compared control mice with animals that lacked Rap1 in the forebrain. The knockout mice had a weaker glucose response to low-dose metformin but remained responsive to other glucose-reducing agents tested in the study. Experiments that changed Rap1 activity and measured neurons in the ventromedial hypothalamus gave the proposed pathway biological support.

The design also limits what the result can establish. It used gene-altered mice and high-fat diets, and some experiments administered metformin directly to the brain. Those procedures cannot determine whether the same pathway drives long-term oral treatment in people.

An explanation and a mechanism are not equivalent

A drug-mechanism experiment asks whether changing a biological pathway changes an observed effect. An explanation tool for artificial intelligence may instead display a heat map, feature ranking, or other account of why a model produced an output. The comparison is useful for separating evidence layers, but the two forms of explanation use different methods and cannot be validated as if they were interchangeable.

A 2023 systematic review screened 882 papers on explainable AI in health care and included six; it found varied measures of fidelity, interpretability, plausibility, user satisfaction, trust, and task performance. The authors found no comprehensive, agreed framework for explanations or standardized method for evaluating whether they work for different users. A readable display therefore does not by itself establish that it faithfully represents the model's decision process.

Clinician trust can rise, fall, or stay unchanged

A 2024 systematic review screened 778 articles and included 10 studies that measured clinicians' trust: five reported an increase with explainable AI, three found no effect, and two found that explanations could increase or decrease trust. The reviewers rated every included paper at moderate or moderate-to-high risk of bias. Complex or contradictory explanations were among the factors that could reduce trust.

The review also separates trust from correctness. Greater trust in a wrong recommendation and lower trust in a correct recommendation can both reduce clinical accuracy. A trust survey cannot establish that a system improves patient care, and an explanation should be tested for whether it calibrates reliance on the model.

Validation has to follow the intended use

The FUTURE-AI consortium built its 2025 consensus guideline through work involving 117 experts from 50 countries and set 30 practices covering design, development, validation, regulation, deployment, and monitoring. Its six principles are fairness, universality, traceability, usability, robustness, and explainability. The practices span technical, clinical, socioethical, and legal questions rather than treating model accuracy as a complete evidence package.

For a medical AI claim, the evidence chain starts with a defined task, population, setting, and model version. Independent testing needs data that reflect sites, equipment, and records outside the development set. Prospective evaluation can then compare care with and without the system using predefined measures such as missed cases, false alerts, workload, and patient outcomes.

Deployment adds another layer. Developers and health organizations need records of version changes, data drift, incidents, cybersecurity controls, and human overrides. Monitoring those events tests whether performance persists after the model leaves its original dataset.

Global principles do not create a universal approval

The World Health Organization's 2023 publication grouped its considerations into documentation and transparency, life-cycle risk management, intended use and external validation, data quality, privacy and data protection, and stakeholder collaboration. The publication addresses governments and regulatory authorities across different settings. It supplies common questions, not an authorization for a particular product.

Whether software is regulated as a medical device depends on its intended use and the law in each market. Approval, permitted claims, and availability must therefore be checked by country. This report does not assess the status or deployment record of any named medical AI product.

What the metformin example does and does not prove

The mouse study adds a candidate biological explanation, but it does not settle metformin's mechanism in people. The AI reviews show that a plausible-looking explanation may affect clinician trust, while the small and inconsistent evidence base leaves its effect uncertain. Neither finding validates a medical AI system.

A medical AI case becomes stronger when performance replicates outside the development dataset, benefit is measured prospectively in the intended workflow, people can intervene, and post-deployment changes are tracked. The explanation remains one part of that record rather than a substitute for it.