By early 2023, after seeing how AI had become a quiet force in our back-office operations, I began exploring its role in something far more complex: health communication, specifically how we communicate across languages, cultures, and health literacy levels.

It started with curiosity. I wanted to understand what an LLM could do if tasked with generating QRD-compliant patient information in UK English. So, I invented fictional angina medications and asked a public LLM to write a package leaflet.

The results were surprisingly solid. They were on average 80% accurate (in terms of QRD compliance and leaflet content): not perfect, but the foundations were there. I corrected some elements, rewrote others. Given that the content was entirely fictional and there was no SPC behind it, I was not expecting pharmacological accuracy. What I was looking for was structure, clarity and tone. I saw that generating first drafts of leaflets were feasible through LLMs.

This made me want to delve deeper.

Pushing Boundaries with Language

I decided to explore how LLMs handled generating QRD-compliant templates in lesser-known languages, using ChatGPT 3.5. I chose four: Galician, Sorani Kurdish, Frisian and Papiamentu. I had native speakers review the outputs:

·       Galician: around 85% accurate

·       Sorani Kurdish: 34% accurate

·       Frisian and Papiamentu: over 90% accurate.

Then, some months later, using GPT 4 (Pro version), I ran the same task again. The jump in quality was extraordinary:

·       Galician: 98%

·       Sorani Kurdish: 90%

·       Frisian and Papiamentu: 98 and 99%, respectively.

These did not require entire rewrites anymore. They just required careful editing.

This was just an experiment, an anecdote. It only had one reviewer per language and no further validation process. But it was interesting, nonetheless.

LLMs can generate first drafts of patient-facing information that rival – if not exceed – those written by many professionals.

This is not a throw-away comment. I have spent years reading the landscape and testing a lot of health information that is not great. While some health communication is decent, there is a lot that could be better.

Slot Machine Messaging in Health Communication

As the experiments grew more elaborate, a new model began forming in my mind, what I now call Slot Machine Messaging.

The idea for this is simple: it is customisation on demand, with dimensions tailored to public health communication. Here is what the LLM tools I used could respond to:

·       Selecting language variant: Not just ‘Spanish’ or ‘French’, but ‘Spanish of Equatorial Guinea’ or ‘French of Gabon’, etc.

·       Selecting audience: adolescent girls, older caregivers, children, peri-urban community health workers, etc.

·       Selecting topic: HIV prevention, stroke symptoms, breastfeeding, family planning, etc.

·       Cultural considerations locale-specific, context-sensitive: gender roles, religious taboos, oral literacy traditions, etc.

Click spin. Build the message.

These were not translated messages but entirely generated, localised health communications. They were directed by our insight, tuned into context and yielded a very good first draft. The process allows the designer of the information to anticipate cultural sensitivities from the outset. We can know before engaging a community that a certain metaphor might offend, or if a topic is taboo. Forewarned is forearmed. We can use a tool like this to start a targeted health communication campaign from a position of being prepared, rather than being reactive.

Of course, it would never replace human input, but kickstarts it.

Global Health Glossaries

This led us to build out multilingual glossaries on specific health and social topics in lesser-known languages. We began generating and refining vocabulary sets in dozens of languages like Uyghur, Sango, Kirundi, Tamasheq, Guaraní, Greenlandic, Nauruan and lots more. We covered topics such as heart health, malaria prevention, COVID-19 (because it had such a rich corpus), HIV awareness, stroke detection and response, sexual and reproductive health and so on.

Here, LLMs offered tremendous scope. The capacity to generate base material for localisation was very powerful.

But one thing must be clear: native speaker input is non-negotiable. The risk of hallucinations is high. Misleading terms and nonsense phrasing occur all the time. But with newer LLMs, we noticed that accuracy of generated items is improving. And the ability to customise across hundreds and hundreds of languages and sub-varieties in seconds is something that no traditional workflow can match. Again, it could be useful to get resources in a given language started quickly, making it ready for native speaker input. This could be useful in a Crisis and Emergency Risk Communication scenario: natural disasters, disease outbreaks, extreme weather events.

What’s The Point of all This?

None of what I described here is hypothetical. They are not toy model or thought experiments. We have done them and others can emulate this to help reshape our approaches to health education and public-facing medical communication across multi-country, multilingual communities.

AI gives us tools that can free us up to do real work: connecting, testing, refining, in the spirit of co-creation. AI, in the instances I described in this article, is not writing for us, but with us, across the globe, across contexts and should always be with human steer. AI can run most of the race, even over 90% of it, but we must be the ones to take the baton across the line.

In the next article, I explore the frontier of regulation, proprietary data and brand risk. These are high-risk areas where AI can clearly help, but where boundaries must be crystal clear.

 Thank you for reading,

Mark Gibson

Leeds, United Kingdom, Easter 2025