The trust deficit we cannot automate
Author: Zoe Jordan1
1. JBI, School of Public Health, Adelaide University, Adelaide, SA, Australia
I have been contemplating the theme of this year’s World Evidence-Based Healthcare Day, Evidence and AI: People at the Centre, for some weeks now. As artificial intelligence (AI) becomes increasingly embedded in evidence synthesis, there is much discussion and debate about accuracy, hallucinated citations, or the risk of a black box producing conclusions no one can interrogate or verify. Those concerns are valid, of course, and our sector is right to take them seriously. But at times I wonder if we are solving the wrong problem, or at least not the one that will determine whether our work actually changes anything.
At face value, it appears as though AI is doing exactly what we hoped: the volume of systematic reviews available to inform decisions is growing quickly. In fact, there is now an abundance of evidence syntheses available to decision makers; perhaps even (dare I say it) an overabundance? However, that is not where the trust deficit lives; that lives downstream, in the systems where decisions are actually made. We can produce more evidence, engage more closely with the practitioners and policymakers who need it, and still watch that evidence sit unused, because the systems around decision-making have not evolved to receive it.
More reviews is not the same as more uptake
It is important to acknowledge that real progress has been made on both the supply and the demand sides. We are increasingly better than we used to be at working with interest holders to shape synthesis questions around what they actually need to decide, rather than what researchers find interesting. That has been a particular focus of the Evidence Synthesis Infrastructure Collaborative (ESIC), a bold effort to mobilise the global synthesis community to work more effectively together across the supply and demand sides to deliver trustworthy evidence at scale and pace. Equally, the increased collaboration among global networks to produce the Responsible use of AI in evidence SynthEsis (RAISE) guidance, is a notable step toward ensuring that AI-assisted synthesis is transparent and ethical.
But asking the right question and answering it quickly, while important, does not guarantee the answer reaches the point of decision, still less that it shapes that decision-making. Perhaps this is truer for clinical practice than it is for policy. The infrastructure that would let a clinician routinely locate, appraise, and apply a relevant review at the moment a decision is being made [largely] does not exist. Reviews continue to accumulate in repositories built for researchers, not in the real-time workflows of the people making health system or practice decisions. The output has scaled, but the absorption capacity around it has not. Furthermore, that gap is not evenly distributed. It is often widest in under-resourced health systems, where the infrastructure to receive and apply evidence was already thinnest before AI accelerated the supply of syntheses.
Why this isn’t a tech problem, but a human one
It is tempting to frame the risks of AI in evidence synthesis primarily as a governance question: addressing whether the methods are transparent, whether human oversight is adequate, whether disclosure is sufficient. Questions on those topics matter, and initiatives like ESIC and RAISE are addressing them directly. But governance of production does not solve a bottleneck in use. Even a perfectly governed, transparent, rapidly produced review changes nothing if the system it is meant to inform has no mechanism, no individual, positioned and supported to use it.
That last part matters more than the infrastructure metaphor usually allows. AI does not create trust. A systematic review, however rigorous, does not create trust. People create trust, and people make decisions. A clinician acts on a recommendation because a person [or an organisation] they trust has interpreted it, contextualised it, and stood behind it, not because an algorithm assembled it quickly. Any implementation infrastructure we build (registries, decision support tools, embedded evidence roles, institutional routines) is unlikely to succeed if it is designed around the evidence rather than around the people who have to act on it. The interface between synthesis and decision is not, at its core, a technical interface. It is a human one; it’s a relational one; and it needs to be built, staffed and supported accordingly.
What this means for those of us leading these organisations
For those of us running global collaborative evidence networks, this should reshape where we put our effort. Demand-side engagement, working with interest holders to define the right synthesis questions, is necessary but no longer sufficient on its own. We need to turn as much attention to the systems on the receiving end as we have turned to the systems producing evidence. That means partnering with health systems and policy bodies not just to ask better questions, but to build the mechanisms that let answers land somewhere real.
It also means being honest that AI's contribution, however significant, addresses a part of the evidence ecosystem that was already improving. The part that has not kept pace is not a technical problem we can solve with a model. It is an institutional one, and it will require the same kind of sustained, cross-sector, collaborative effort that has been invested in synthesis, applied to implementation infrastructure instead.
The task ahead
Our sector has spent several years getting better at asking the right questions at the right time of the right interest holders to produce better evidence. These are genuine achievements to be celebrated. The next task is less glamorous and perhaps harder to solve; that is, building the systems, inside the places where decisions actually happen, that can receive what we are now capable of producing. That is the work in front of us.
References
Evidence Synthesis Infrastructure Collaborative. (2025). ESIC. https://www.evidencesic.org
Flemyng, E., Noel-Storr, A., Macura, B., Gartlehner, G., Thomas, J., Meerpohl, J. J., Jordan, Z., Minx, J., Eisele-Metzger, A., Hamel, C., Jemioło, P., Porritt, K., & Grainger, M. (2025). Position statement on artificial intelligence (AI) use in evidence synthesis across Cochrane, the Campbell Collaboration, JBI, and the Collaboration for Environmental Evidence 2025. JBI Evidence Synthesis, 23(11), 2162–2166. https://doi.org/10.11124/JBIES-25-00480
Flodgren, G., O'Brien, M. A., Parmelli, E., & Grimshaw, J. M. (2019). Local opinion leaders: Effects on professional practice and healthcare outcomes. Cochrane Database of Systematic Reviews, 6, CD000125. https://doi.org/10.1002/14651858.CD000125.pub5
Grimshaw, J. M., Eccles, M. P., Lavis, J. N., Hill, S. J., & Squires, J. E. (2012). Knowledge translation of research findings. Implementation Science, 7, 50. https://doi.org/10.1186/1748-5908-7-50
Hoffmann, F., Allers, K., Rombey, T., Helbach, J., Hoffmann, A., Mathes, T., & Pieper, D. (2021). Nearly 80 systematic reviews were published each day: Observational study on trends in epidemiology and reporting over the years 2000-2019. Journal of Clinical Epidemiology, 138, 1–11. https://doi.org/10.1016/j.jclinepi.2021.05.022
Oliver, K., Innvar, S., Lorenc, T., Woodman, J., & Thomas, J. (2014). A systematic review of barriers to and facilitators of the use of evidence by policymakers. BMC Health Services Research,14, 2. https://doi.org/10.1186/1472-6963-14-2
To link to this article - DOI: https://doi.org/10.70253/LFEX9561
Conflict of interest
Zoe is a member of the World EBHC Day Steering Committee.
Disclaimer
The views expressed in this World EBHC Day Blog, as well as any errors or omissions, are the sole responsibility of the author and do not represent the views of the World EBHC Day Steering Committee, Official Partners, or Sponsors; nor does its publication imply endorsement by the aforementioned parties.
