AI Can Read Evidence. Can It Read the Room? 

Artificial intelligence can read a hundred research papers before I finish my morning coffee. Impressive. But can it tell me whether an intervention tested in London will work in Bamenda? 

It may recognise the intervention, population, and reported effect, then produce a beautifully confident recommendation. What it may miss is the understaffed facility, unreliable internet, unaffordable clinic visits or community health worker who understands why patients are reluctant to return. AI can read the evidence, but reading the room is harder. 

This question shapes my work as young African evidence professional. I have used AI to screen studies, assist with data extraction, map decision-making systems, and assess whether the transferability of evidence between settings. The experience has made me optimistic but not dazzled. AI is fast and useful; it can also be confidently wrong. That distinction matters because a technically impressive system can still produce poor decisions when its training data, evaluation design or operating assumptions do not match local reality in practice. 

When evidence travels without its luggage 

Healthcare evidence constantly moves across contexts. When a programme succeeds in one setting, decision-makers ask whether it should be introduced elsewhere. AI can find and summarise relevant studies, but that is only the beginning. 

Evidence carries assumptions about its original system: funding, personnel, infrastructure, technology, policy support and public trust. When these remain invisible, an intervention can appear universally transferable even when it is not. 

This challenge sits at the centre of our Transfer Model, which assesses evidence using five contextual predictors: relevance, implementation complexity, cost, potential impact and expected effect size. The practical decision is whether to adopt the intervention, adapt it locally or generate new evidence. An LLM can interpret the research and produce an initial classification, while a decision tree applies transparent rules to the ratings. However, researchers and local experts must examine the reasoning, challenge unsupported assumptions and decide whether the recommendation makes contextual sense. Their role is not ceremonial approval; it gives the analysis meaning. 

When high recall becomes “include almost everything” 

I saw this need while evaluating AI-assisted screening in EPPI-Reviewer for a systematic review on misinformation. The AI assessed titles and abstracts against predefined criteria and recommended inclusion or exclusion. It worked quickly, but systematically over-included studies mentioning misinformation, communication or digital media even when they did not answer the review question. 

Technically, the model favoured recall over precision. Protecting recall is sensible because missing an eligible study can bias a review. However, a system that includes almost everything transfers the workload back to humans and can stretch the review beyond its protocol. We must measure false positives and false negatives, inspect disagreements and compare AI decisions with a human reference standard. 

The reviewers did not simply correct the AI when it was wrong. We asked it to explain every recommendation using the eligibility criteria, so we could see why it included or excluded a study. Those explanations showed us where the prompt was unclear or the model misread the population, phenomenon, study design or outcome. We revised the prompt and tested it again. This aligns with the Responsible use of AI in evidence SynthEsis, or RAISE, guidance, which emphasises task-specific evaluation, contextual validation, transparent limitations and researcher accountability (Thomas et al., 2025a, 2025b, 2025c). It turns “human-in-the-loop” from a reassuring slogan into a testable methodological process. 

Tools strengthen validation. The Data Extraction and Evaluation Toolkit, or DEET, compares AI-extracted data with a human reference standard, including at individual-field level (DESTINY Evidence, 2026). Although DEET addresses data extraction rather than title-and-abstract screening, its logic applies across evidence synthesis: define the task, establish a trusted reference, quantify performance, examine error patterns and improve the workflow. The question is not merely whether AI can complete a task, but whether we have enough evidence to trust it for that use. 

Human-in-the-loop, or human at the edge? 

“Human-in-the-loop” has become a comforting phrase in responsible AI discussions. Yet placing someone at the end of an automated workflow and asking them to click “approve” does not make it people-centred. If they cannot see how an output was generated, investigate errors or reject it, they are not meaningfully in the loop. They are standing at its edge. 

People-centred AI involves people throughout its lifecycle. Patients and communities should help define worthwhile problems. Health workers should identify operational realities that technical systems overlook. Researchers should design validations, interpret performance metrics and monitor accuracy across populations, languages and document types. Policymakers must question data provenance, bias and accountability. Young people should participate not only as users, but as designers, researchers and decision-makers. 

For Africa, this participation is particularly important. Our continent should not become merely a testing ground or final market for health technologies designed elsewhere. African evidence leaders must help determine what AI systems learn, which outcomes they optimise, whose knowledge they recognise and how they are governed. Otherwise, we risk automating the same inequalities that evidence-based healthcare is meant to address. 

Better partnership, not less AI 

These experiences have not made me less excited about AI. They have made me more precise about what I expect from it. I want AI that helps evidence teams manage growing research volumes without lowering methodological standards; tools that make evidence more accessible to policymakers, practitioners and communities; and systems that express uncertainty instead of hiding it behind polished language. Most importantly, I want AI that strengthens human judgement instead of quietly replacing it. 

Achieving this requires more than adding an ethics statement after a system has already been built. It requires diverse participation from problem definition to deployment, transparent documentation of models and prompts, evaluation against human standards, and continuous monitoring after implementation. Local evidence must be treated as essential rather than supplementary, and affected communities must have genuine influence over decisions. 

The future of evidence-based healthcare should not be humans versus machines. It should be machines that expand what people can do, people who challenge what machines think they know, and communities with the power to decide what progress looks like. AI may be able to read all the evidence in the world, but before we allow it to shape healthcare, we must make sure it has learned to read the room. 

The Author

Kinlabel Tetamiyaka Okwen, IT Manager eBase and Team lead at iCODE Abakwa, Ghana.
Kinlabel has a passion for bridging the gap between technology and African innovation and she is dedicated to tackling the challenges across the country. Through her work at iCODE Abakwa, a team of African coders breaking down barriers with innovative tech solutions, she has contributed to designing tools that adapt global technologies to solve local problems in meaningful ways.

References 

DESTINY Evidence. (2026). Data Extraction and Evaluation Toolkit (DEET). https://destiny-evidence.org/data-extraction-evaluation-toolkit/ 

Thomas, J., Flemyng, E., Noel-Storr, A., et al. (2025a). Responsible use of AI in evidence synthesis (RAISE) 1: Recommendations for practice. Open Science Framework. https://doi.org/10.17605/OSF.IO/FWAUD 

Thomas, J., Flemyng, E., Noel-Storr, A., et al. (2025b). Responsible use of AI in evidence synthesis (RAISE) 2: Building and evaluating AI evidence synthesis tools. Open Science Framework. https://doi.org/10.17605/OSF.IO/FWAUD 

Thomas, J., Flemyng, E., Noel-Storr, A., et al. (2025c). Responsible use of AI in evidence synthesis (RAISE) 3: Selecting and using AI evidence synthesis tools. Open Science Framework. https://doi.org/10.17605/OSF.IO/FWAUD