Behind the Keyboard: What AI Tools Could Not Tell Us
Authors: Karina Faria de Souza1,2, Ana Paula Almeida Brito1,2, Karina Sichieri1,2
1. Hospital Universitário - USP
2. JBI Brazilian Centre for Evidence-Based Health Care
This year, when we came across the theme for World EBHC Day, we did something that felt fitting: we asked AI itself the very question posed to us as a theme. We put the same prompt to four different AI tools – What is the role of artificial intelligence in evidence-based healthcare? – and the answers were startlingly similar. All these AI tools replied that they could assume the role of sourcing and presenting evidence-based content from around the world to guide us in delivering healthcare aligned with the latest knowledge and developments. Which, of course, raised the most important question of all: who is behind the keyboard, inputting healthcare-related questions to AI tools, and do they know how to critically evaluate what they are reading?
We do not ask that question from a distance. We are nurses in the Neonatal Unit at the University Hospital of the University of São Paulo – and we are also, we must confess, Artificial Intelligence enthusiasts. We have used AI for statistical analysis and to locate academic references, organise content, and brainstorm work schedules. Its applicability is undeniable. But so are its risks. And it was inside a real implementation project – one built on evidence, collaboration, and a great deal of human judgement – that we found our most honest answer to that question.
The hospital discharge of a preterm newborn is one of the most emotionally charged moments in neonatal care. For families, it brings hope – and fear. For us, it used to bring frustration. Despite our commitment to quality care, our discharge process was fragmented, documentation was scattered, and instructions were delivered to exhausted, overwhelmed parents in the middle of a busy ward. We were offering information when what families needed was clarity and confidence.
That tension pushed us to act. Using the JBI framework for evidence implementation, we built a multidisciplinary team and set out to redesign discharge planning for preterm newborns. We conducted a baseline audit, identified our gaps, reviewed the best available evidence, and developed a Discharge Planning Guideline for Preterm Newborns. We also created an illustrated booklet in plain language for families, centralised documentation across the team, and used an educational video and WhatsApp to disseminate the content of the Discharge Planning Guideline – aiming to build a new culture of shared responsibility.
But one challenge remained: how do you ensure a technical guideline genuinely reaches the professionals responsible for its implementation? Busy nurses juggling multiple demands; physicians on a night shift; physical therapists, speech-language therapists, nutritionists, and psychologists in multiple wards – a multidisciplinary team that has seen many guidelines come and go?
That’s where artificial intelligence entered our story.
As highlighted by Trinkley and colleagues (2024), leveraging AI offers significant opportunities to advance implementation science. So, we decided to use an artificial intelligence tool to produce an awareness-raising video to support the implementation of our Discharge Planning Guideline, directed at the healthcare professionals involved in neonatal discharge. The purpose was to engage our team and communicate the recommendations in a way that would capture attention, spark reflection, and motivate change.
And here we had to make a very deliberate choice about what AI would – and would not – do for us. The content was entirely ours. The script, clinical recommendations, language choices, communication strategy – all of that came from the project team, grounded in the evidence we had synthesised and the barriers we had identified through our audit. AI did not generate the evidence. It did not assess our local context. It did not understand the nuance of our team dynamics or the specific resistance patterns we had observed. Those were human judgements, and they had to stay that way.
What AI did was bring that content to life. The visual layout, the avatar, the voice narration, and the background music were all produced with AI tools – quickly, affordably, and with a quality that genuinely surprised us. The result was an engaging audiovisual resource that professionals actually wanted to watch. We then disseminated it through social media, extending its reach beyond our unit. It was a perfect combination.
But let’s be honest about the risks. We understand why many professionals and researchers almost demonise artificial intelligence. Complex databases are accessed day and night to answer questions that are not always asked by people who know how to interpret the answers. The risk of distortion is real: a poorly framed question, an analysis that overlooks non-obvious biases, or a confident-sounding response that is simply wrong – what we call hallucination. These are not hypothetical concerns. They align with the ethical and governance issues raised by the World Health Organization regarding AI, and we take those concerns very seriously.
In our project, our main safeguard was structural: we never asked AI to generate or interpret clinical content. AI entered the process only after the evidence had been appraised, the recommendations had been defined, and the script had been written and reviewed by the team. AI was working downstream of our critical thinking – not upstream of it. That boundary was not accidental. It was a conscious decision rooted in exactly the kind of critical appraisal that evidence-based healthcare demands of us every day.
This is what we believe: AI is a powerful supporting tool, but the protagonist is always the person behind the keyboard – the one who knows what to ask, how to evaluate the response, and how to use that output as an ally to human judgement, observation, and creativity. Many people want to use AI as a shortcut, becoming hostages to its algorithms. In science and in clinical practice, that is not the intended purpose of AI.
When we asked those four AI tools about the role of artificial intelligence in evidence-based healthcare, the answers were startlingly similar. But our answer to that question looks different – because it was built in a neonatal unit in São Paulo, as a team, one careful and human decision at a time.
References
Trinkley, K. E., An, R., Maw, A. M., Glasgow, R. E., & Brownson, R. C. (2024). Leveraging artificial intelligence to advance implementation science: Potential opportunities and cautions. Implementation Science, 19(1), 317. https://doi.org/10.1186/s13012-024-01346-y
World Health Organization. (2021, June 28). Ethics and governance of artificial intelligence for health. WHO. https://www.who.int/publications/i/item/9789240029200
To link to this article - DOI: https://doi.org/10.70253/JXIL3410
Links to additional resources
The video mentioned above is available at https://canva.link/2gixnwozge5kh0k.
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 it imply endorsement by the aforementioned parties.
