First things first, I won’t write here about the existential threat of Artificial Intelligence (a common topic these days), as I don’t want to risk a Dunning-Kruger effect. You might want to listen to the related experts on that issue. Here, instead, I will explore opportunities for the intersection between a complex tool, Generative AI, and a similarly complex process and dynamic, Knowledge translation.
As an AI ‘first-timer’ (and like so many other enthusiasts with a rather basic appreciation of new technology), I’m enthralled with the progress Artificial intelligence has made over the last 60+ years. I watch with fascination as experts and commentators refer to how close AI has gotten to human-like characteristics in the last ten years – or not. Some of the clues can be found in AI terminologies, such as Human General Intelligence (HGI) or Artificial General Intelligence (AGI), which make you wonder about the respective realms of computers and humans. (ps: AGI is “a machine that can perform any intellectual task that a human can”. We aren’t there yet, but some claim that’s a matter of some years only)
OpenAI, whose product ChatGPT – an example of Generative AI (Gen AI) – has generated a lot of interest in Artificial Intelligence in the last three years, defines its mission as “to ensure that AI systems are generally smarter than humans”, while making sure at the same time that they “benefit” humanity. ChatGPT, basically a chat box, is quite intuitive. It not only makes it easy for one to have a wealth of “information at one’s fingertips” but it also attempts to provide you with an interactive guide (/companion) through different thought processes, with the help of scenarios (real ones or otherwise). Policymakers have probably already made use of this tool by now. Nevertheless, making highly impactful decisions requires more than the intuitions ChatGPT provides – particularly when the contexts are different from those for which the data used was collected. And of course, you also have to keep in mind the obvious AI biases and errors already flagged in various research and media articles in recent months.
Over to Knowledge translation (KT), then, defined here as “the iterative and dynamic identification, synthesis, dissemination, exchange, and application of evidence to accelerate the benefits of global and local innovation in strengthening health systems and improving people’s health.” I work in this field since a number of years. As the Covid pandemic showed, when facing an abundance of rapid evidence syntheses and under huge pressure to take decisions in conditions of uncertainty, policymakers often ended up ignoring the evidence presented to them, or just completely misunderstood the information. Lessons from the Covid-19 response, as debated in the UK parliament supplement already “known” information: that just because the evidence synthesis is available and communicated, it won’t necessarily be used (properly).
Lessons from KT related to informing complex and obtuse policy processes with the best available evidence, could provide a “blue-print” for working at the intersection of AI and KT, as they typically assess the utility of scientific evidence and data. It is critical to ensure that the KT process embeds interactions and relationships between and within “active” networks of stakeholders, ànd is sustained . The question then, is how can AI be optimally leveraged for this (KT) endeavour?
AI tools have been described to elicit a near “radioactive” feeling – e.g., “Working with these tools feels like being bitten by a radioactive spider and gaining a suite of superpowers.” . Superlative descriptions, perhaps, but I guess the point is that understanding and leveraging their full potential can (and will?) make one feel invincible- if there exists such a thing, obviously. Interaction with AI aims to improve the intricate and intimate understanding of problems, processes, and dynamics. However, as mentioned earlier, the complexity of KT does not reside in “one” problem only. It is layers of problem(s), networks, processes, and dynamic(s).
Additionally, the rapid developments in AI put focus on the importance of interdisciplinary and transdisciplinary approaches when dealing with complex problems. Similarly, knowledge translation also requires an interdisciplinary, transdisciplinary approach between stakeholders, from research academia, civil society organizations, funders, communities, to policymakers. KT is more about awareness of the importance of context, culture, knowledge, attitudes, leadership, values, ethics, and economics and how these can affect the application of evidence into policy processes. Similarly, although data scientists have perhaps focused on the latest computation prowess, there is also an increased appreciation of the relevance of contextual awareness in the development of AI tools – which could hopefully improve their adoption and utility. In the optimistic scenario, the two processes could feed each other and tailor support to institutions and individuals, possibly improving the application of evidence in policy processes.
As an example, imagine civil servant X working at the Ministry of Health in a country in Africa, who needs to decide whether to use community health workers in contact tracing during an outbreak of disease Q in community W, living in the mountains. Should the Minister present to Parliament the option to recruit more community health workers external to the community, or use the few already present there, but hand them more responsibilities during this period? The technocrat is expected to respond to the Minister of Health within seven days.
KT is about improving the technocrat’s engagement with scientific evidence. In simple terms, it is about the decision-maker’s ability to understand the evidence, where to find it, synthesize and translate it, and apply to his/her context (eg. keeping in mind gender considerations, cultural values, norms and attitudes towards the disease and external people). To do this, the technocrat also needs to have the relevant knowledge and skills, and be aware of these values and norms, in order to be able to engage with each of the different steps within a KT process. KT hubs, for example those conducting rapid response services for policymaking, have improved their capacity to engage policymakers within short time periods with the evidence, through iterative processes of clarifying policy issues, co-creating the policy question, and translating it.
In the coming years, it seems likely that AI tools like ChatGPT (or its successors) will improve the productivity and efficiency within evidence-informed policymaking. However, AI enthusiasts will have to work with KT practitioners and other disciplines to ensure it is relevant and useful. At least, if done well and with the caveats and constraints mentioned above.
Well, what does all this have to do with the reinstatement of the Open AI chief from a few weeks ago, you might ask? Well, perhaps just that the chaos and mystery of what really happened during that weird saga, feels more than a bit like the messy and obtuse policy process a KT expert could relate with.
Disclaimer: I am obviously excited to be part of a consortium funded by the Bill and Melinda Gates Foundation, made up of experts in knowledge translation in Africa, including the Pan African Collective Evidence in South Africa, Center for Rapid Evidence Synthesis in Uganda, and Ethiopian Public Health Institute in Ethiopia. The consortium will explore the opportunities (caution!) of digital solutions, including Gen AI in knowledge translation. As this journey starts this year, the development of new and exciting tools provides a worthwhile opportunity to explore another dimension of AI for the future.