The doctor the machine drew first
Author: Bianca Pilla1
1. JBI, School of Public Health, Adelaide University, Adelaide, SA, Australia
Earlier this year, I asked a generative AI tool to help me design an infographic about the gap between how research evidence is written and communicated and what patients actually need from that evidence. My first prompt was brief. I described the topic but said nothing about diversity, accuracy, or who should appear in the image. The result below is what the tool produced when left to its defaults. The expert it drew was a white man, blond, bespectacled, confident. His patient was an elderly Western woman in pearls, visibly overwhelmed. Between them sat a wall of jargon and a scatter of puzzle pieces carrying statistical terms, several of them misspelled and/or erroneous. I had asked for an illustration of the pitfalls of communicating science to patients, and the tool had inadvertently (and ironically) reproduced many of them in this image.
Figure 1: AI-generated infographic about communicating evidence to patients. The image contained accuracy and inclusion errors.
I keep returning to this image because it captures, more economically than I could, the question at the heart of this year’s World EBHC Day theme, Evidence and AI: People at the Centre. As AI becomes part of how we generate, translate, and communicate evidence, who does it represent and who does it leave out?
I come to this question from two directions. For more than 20 years, I have lived with seven autoimmune and chronic conditions, including endometriosis and ulcerative colitis – two decades of consultations, scans, and medical information that often seemed written for someone other than me. For more than a decade, I also worked alongside rural communities in the Global South, where the question of whose knowledge is recognised and whose is overlooked was ever-present. These experiences are part of why I came to research inclusive science communication, and why I write here as one of the lived-experience researchers on a recent narrative review of AI and inclusive science communication in evidence-based healthcare.
The question behind the methods section
Patients, in my experience, rarely ask about the methodological rigour of studies. They ask what the evidence means for them: whether a treatment is likely to help, what it might cost them, and whether it was ever intended for someone in their circumstances. Inclusive science communication takes this seriously. It reframes exclusion not as a failure of individual comprehension but as something produced structurally through the language we use, the knowledge we treat as legitimate, and the assumptions we build into how evidence is shared. On this account, communication is not a neutral channel through which evidence passes. It is part of how evidence comes to mean something, and for whom.
What the image revealed
This is why the infographic gave me pause. Its first version did not merely look cluttered; it encoded a particular set of defaults about who counts as an expert and who is cast as the confused recipient of expertise. The science was compromised too; terms such as ‘risk ratio’ and ‘subgroup analysis’ appeared garbled, and a non-significant p-value was presented as though it settled the question at hand. Accuracy and inclusion had failed together, in the same image, which is itself instructive.
I produced a second, better version through deliberate prompting. It is tempting to describe that revision as a matter of making the people more diverse, but that would misrepresent what was improved and would understate the problem. The substantive gains were threefold: the statistical and methodological terms were corrected, so the science was sound; the clutter was removed, so the central message was legible and accessible; and the barrier between clinician and patient was reframed from a heap of jargon into what it more accurately represents – differences in language, priorities, and context.
Figure 2: AI-generated infographic produced after deliberate prompting to address accuracy and inclusive representation issues.
Yet fixing what was broken is not the same as fixing what was missing. The second infographic corrected the science and cleared the clutter, but it did so on my terms: I decided what counted as accurate, what counted as legible, and what counted as a fairer depiction of the clinician–patient divide. None of that touches the harder question of who gets to make those calls in the first place, or whether a more considered output changes anything about the hierarchy the first version made so visible. That distinction – between a better image and a genuinely inclusive process – is easy to miss.
Representation is not inclusion
In our recent narrative literature review, my colleagues and I describe a pattern we term proxy participation, or the appearance of inclusion. That is, when a more diverse image, a personalised tone, or a translated phrase stands in for the redistribution of authority that inclusion actually requires, while leaving the underlying hierarchy of whose knowledge counts intact. A more representative cast arranged in front of the same exclusionary assumptions does not make communication inclusive. It makes it appear inclusive, which is not the same thing, and can be harder to detect precisely because it reassures.
The burden of noticing
I want to be careful not to claim too much for my own intervention. I noticed the defaults in that first image because of my vantage point of years as a patient, time spent with communities that dominant systems tend to overlook, and science communication training that taught me to look for these patterns. The AI tool, by contrast, defaulted to the white, male expert first, and it took sustained and deliberate effort on my part to redirect the tool’s approach. Even then, my corrections operated downstream of design decisions I could neither see nor change – the training data, the optimisation choices, and the architecture that made that doctor the most probable output in the first place.
This points to a lesser-known form of inequity than the one we most often discuss. We tend to worry, reasonably, about unequal access to AI. The difficulty here is different, as the work of noticing AI’s exclusions tends to fall to the people it is most likely to exclude, and not everyone is positioned or resourced to undertake this work. For those of us communicating evidence, the role is shifting from author to interrogator, as we are increasingly answerable not only for what we produce but also for understanding and disclosing the limits of systems built by others. For instance, De Vito et al. (2024) found that an AI chatbot answered HIV prevention questions with high technical accuracy while using outdated terminology and erasing transgender and non-binary people. A clean accuracy check, in other words, is not the same as inclusive communication.
Putting people at the centre
None of this is an argument against using these tools. The potential is real: AI can translate evidence across languages, adapt it to different needs, and reach people the health system has too often reached last. The argument is rather that the potential is conditional. It depends on how these tools are designed, governed, and reviewed.
In practical terms, and consistent with the guidance our review works towards, this means reviewing AI-assisted outputs for framing and representation and not only for factual accuracy; validating anything ‘personalised’ with the people it claims to describe, rather than trusting the model’s approximation of them; disclosing when AI has been used and what is beyond it capability; and, wherever possible, developing communication with the communities whose health is at stake, rather than for them. These are not new principles. They are the long-standing commitments of inclusive science communication, applied to a setting that makes them easier to neglect.
Where to now?
Keeping people at the centre can mean consulting patients once a design is finished, or it can mean treating patients and communities as partners in shaping what is built. The difference is not merely rhetorical; it determines whether AI becomes a route to more inclusive communication or a faster way to reproduce the defaults we already know how to recognise. Our review sets out the foundations for integrated, validated guidance to help science communication practitioners make that choice deliberately, and the next stage of this work is now underway to build a consensus on such guidance with practitioners, community representatives, and AI ethics scholars.
The machine drew a particular doctor first. The question we need to ask ourselves, and the one I think this year’s theme asks of all of us, is not who it drew, but who it did not draw – and what we intend to do about that.
References
Canfield, K., & Menezes, S. (2020). The state of inclusive science communication: A landscape study. Metcalf Institute, University of Rhode Island.
De Vito, A., Colpani, A., Moi, G., Babudieri, S., Calcagno, A., Calvino, V., Ceccarelli, M., Colpani, G., d'Ettorre, G., Di Biagio, A., Farinella, M., Falaguasta, M., Focà, E., Giupponi, G., Habed, A. J., Isenia, W. J., Lo Caputo, S., Marchetti, G., Modesti, L., Mussini, C., … Madeddu, G. (2024). Assessing ChatGPT’s potential in HIV prevention communication: A comprehensive evaluation of accuracy, completeness, and inclusivity. AIDS and Behaviour, 28(8), 2746–2754.
Finlay, S. M., Raman, S., Rasekoala, E., Mignan, V., Dawson, E., Neeley, L., & Orthia, L. A. (2021). From the margins to the mainstream: Deconstructing science communication as a white, Western paradigm. Journal of Science Communication, 20(1), C02.
Hansen, A. (2016). The changing uses of accuracy in science communication. Public Understanding of Science, 25(7), 760–774.
Hendriks, F., Barel-Ben David, Y., Banse, L., Fick, J., Greussing, E., Klein-Avraham, I., Rakedzon, T., Taddicken, M., & Baram-Tsabari, A. (2025). Generative AI in science communication: Fostering scientists’ good working habits for ethical and effective use. Science Communication.
Luna, D. S., Broer, I., Bilandzić, H., Taddicken, M., Schuller, B. W., & Bürger, M. (2025). Quality in science communication with communicative artificial intelligence: A principle-based framework. Public Understanding of Science, 34(8).
Pilla, B., Mohan, A., Elanga Elanga, A., Dean, L., & McCulloch, H. (2026). Power, voice, and accountability: How artificial intelligence reshapes inclusive science communication in evidence-based healthcare – A narrative literature review. Frontiers in Communication, 11, 1879185.
To link to this article - DOI: https://doi.org/10.70253/YUEV7721
Conflict of interest
Bianca is the Chair 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.
