We have seen how LLMs could help generate IFUs, package leaflets, explore health literacy in lesser-known languages and support cross-cultural communication with a nuance and speed that would never have been possible just a few years ago. The next part of the story takes us to a more sensitive frontier: regulated content, proprietary data and reputational risk.

COA Content and Concept Work

Coming out of our global health communication work, I began thinking about COA development, particularly concept elaboration and translatability assessments. These are the backbone of patient-centric measurement tools used in clinical trials.

There is a catch here: COA instruments are owned. Its content is licensed, copyrighted and protected. You cannot train an AI on it unless you own it or have explicit permission. So, we developed a ‘bank of best practice’ where we proactively compiled concept elaborations of individual vocabulary items frequently used in COAs. I have nearly 20 years’ worth of non-proprietary data to work with. This enabled us to experiment with our own small, controlled and very useful datasets:

·       We fed in individual terms and phrases.

·       We asked the model to explore nuance.

·       We developed an AI-assisted glossary of key COA concept definitions (super useful!)

·       We were to run ‘agnostic’ translatability assessments, without referencing protected content and then dig deeper: what translation difficulties could arise with ‘X’ term in Congolese French or Bahasa or Sesotho, etc.?

The results were stunning, insightful and more precise than – seriously – any human input we have received, particularly for translatability assessments.

Could this go further? Could we one day use LLMs to pre-identify problematic phrases that cause cultural or linguistic friction points and only focus on those in the testing phase? Maybe, but we are not there yet, and it would require a lot of regulatory buy-in. But the potential benefits of AI in the COA-sphere are very clear.

AI and Regulatory Affairs: Proceed with Caution

One area where the boundaries must be very clear is regulatory intelligence, especially in the labelling and packaging regulations of any kind of product, whether it is a device, a drug or a chocolate bar. At GRC, we do a great deal of labelling work, particularly IMP (Investigational Medicine Product) labelling. We do not – and will not – use public LLMs to support regulatory labelling.

However, there is definitely a valid use case for internal, AI-assisted tools, similar to the ‘bank of best practice’ I described earlier. You could build a searchable AI-assisted database of previously approved country labelling requirements. But anything new – any jurisdiction that you have not worked in – must be researched manually by qualified humans with domain expertise.

Why? Because LLMs hallucinate!

I once asked a model to provide updated IMP labelling guidance for several Caribbean countries. It cited a law in one of those countries that did not exist. I knew this because I knew better – I had already led on the manual labelling reviews for these countries. The law it cited was a complete fabrication, an invention, a fiction. Imagine if I had taken that at face value and submitted it to a client?

Also, LLMs cannot keep on top of in-country regulatory updates.

On another occasion, I tested LLMs on regulatory intelligence for Armenia. It cited an article I had seen before. I wrote it in 2015 and nowhere in that piece was Armenia mentioned. The model just seemed to reach for anything that mentioned ‘IMP’. There is a twist as well: we took the article down a few years ago. It must have been cached and somehow the LLM dredged it up as a “source”. This is nonsense.

In addition, how would AI assistance for labelling check for regulatory updates? This is a very important issue to consider.

In this case, it might look sophisticated as a marketing tool to say that you use AI for labelling reviews. However, I really believe that this activity is fraught with potential pitfalls  and ones that have very serious repercussions if the AI is wrong and the human quality checks are deficient. It is a Swiss Cheese potential for systemic failure just waiting to happen.

LLMs are not a source of truth. Do not use it to look anything up. They are only language prediction machines. For regulatory work, we do not want a predictor. We need a guarantor. An experienced human fills that need.

Branding: Where LLMs Can Help

One aspect of regulatory and commercial work where AI has exciting potential is brand name checking. There is an entire industry built around testing new product names, for medical devices, medicines, food, drinks, cosmetics – any product. These tests would look for linguistic, cultural or regulatory conflicts in international markets.

We have all heard and chuckled over naming fails. These are words that sound like slurs or taboo references in another language. It doesn’t even have to be vulgar. I’ve seen a perfectly upbeat UK-based company name that means ‘grief’ or ‘bereavement’ in Spanish. I am reliably told that their Google Analytics regularly shows people from Spanish-speaking countries (one third of the planet, mind you) mistakenly thinking their website is a grief counselling service. Another example is an acronym for a company that turned out to be the same as a notorious secret police unit in another country – one they were trying to sell into.

This is where LLMs are eminently useful. You can spin a name through dozens of languages and dialects to flag problems with candidate brand names early. Is it a substitute for human review? No, but it is an early filter, one that works fast and at scale.

We are already exploring this space because it is not too far removed from the experimenting we have already done around health communication and translatability work. It is just another layer of cultural and linguistic vetting and AI performs very well.

Looking Ahead at the Patient Voice

One final frontier that I want to touch on: Patient Voice research, particularly AI-assisted thematic analysis. This is a topic I will explore in another article entirely, but it is worth mentioning here.

We have barely scratched the surface of how AI can help with qualitative data analysis, patient sentiment mapping and real-time insight generation from interviews, transcripts, or open-ended survey data. There is huge scope here and it demands its own dedicated space.

But as with everything else I have shared in these three articles, one principle remains: AI can only succeed professionally when paired with human transparency, steer and intention.

So much of the conversation about AI still circles around either blind enthusiasm or total suspicion. I hope this series of articles on AI has offered something more balanced. AI is not coming to get us. It is just changing what is possible.

When it is used carefully and collaboratively, it can help us do better work, with deeper focus and impact.

Thank you for reading,

Mark Gibson

Leeds, United Kingdom, Easter 2025