I was asked recently to contribute to a conference presentation about AI and health communication in global health for APHA. That’s what today’s episode is about. We’ll consider some of what’s beneath the surface of AI, in a way that may touch on problems you’re dealing with in AI and healthcare or public health.
Hi, everybody. This is 10 Minutes to Better Patient Communication from Health Communication Partners. I’m Dr. Anne Marie Liebel. Since 2017, we’ve been digging into communication problems and issues in healthcare and public health, and increasingly across sectors because communication touches everything. We’re here to learn and get inspired, but mostly make the difference we got into our jobs to make. Did you know we’ve got a newsletter? It’s good. You should get it. Visit healthcommunicationpartners.com.
Yes, thank you to my co-presenters. They certainly did the heavy lifting. And they gave me a whole lot to think about when it comes to AI and health communication. You know by now that AI relies on existing data sets. We’ve all heard about AI’s biases, its really inflated tone, its overconfidence, the predictable sentence structure, the wooden word choice, blah, blah, blah. This presentation pushed further, into AI and power dynamics and global health.
So I’m going to talk specifically about how AI wields power through prose. Prose, like texts, words, phrases, sentences, bigger chunks of written language. AI was trained on a lot of prose. And now when we interact with AI, of course, we’re reading a lot of its prose outputs.
But it’s more than just words we’re taking in. There is a whole host of other assumptions and judgments embedded in those words or hidden behind them. For instance, like the line of logic behind the sentences, behind AI output. It asks us implicitly to accept:
- its logic;
- the priorities behind its phrases and sentences;
- its considerations;
- its judgments about what’s important or unimportant.
It also asks us to accept that. The texts and the data sets that AI was trained on reflect the worldview of the authors. So those worldviews were transmitted along with the vocabulary and the grammar and the sentence structure, because they’re encoded in it.
This isn’t unique to AI. All language use works this way. You, me, every day, the language we use, the language that gets used on us, reflects our values, assumptions, rationales, priorities, worldviews. And we rely on these, and the way that language works, almost automatically.
For example, I’ve talked before about how it feels if you watch or listen to a news source that’s different from your usual. Whatever your favorite news source, when you tune into a different one, it can sometimes put your teeth on edge, because of the way issues are handled. It just seems like ridiculous or wrong to you. Or in an extreme case, like people that you consider heroic or demonized, problems that you see as systemic in nature get treated as purely individual issues, so on.
There are subtle everyday ways that our language use does this. Linguists have shown that our language reflects what we think is important, normal, or taken for granted. Our everyday language use encourages certain ways of thinking, and discourages others. It privileges some concepts over others.
And it’s true for the way we use words. It’s also true for the way that organizations use words.
For example, at your workplace. Think of how “employee engagement,” for example, gets talked about or written about at work. I don’t know if any words or phrases jump to mind. But some of them would indicate that your workplace sees employee engagement as an individual issue. So it’s a motivation issue or an attitude problem in particular, specific people. But other words or phrases would indicate that they see employee engagement as a systems issue. Maybe it’s a reflection of management, or workload, or employee voice.
So when AI models are trained on texts, it’s not just the words it learns. It’s the worldview behind the words that it’s also taking on. And then spitting back to us!
What’s interesting is that this can seem normal, or natural, or neutral. The fact that it’s coming from a machine, or a faceless program, can amp up the illusion that these words are objective. They’re value-free, sanitized even. But in reality, they reflect the very real worldview, assumptions, values of the people who wrote the majority of the texts the model was trained on.
Sure, AI is sycophantic. As much as the models are programmed to please us, keep us interacting with it, it inescapably still ingests the words and the worldview of the dominant group, those people responsible for the majority of the text that the models are trained on.
And we’re swimming in it, also ingesting it, often going along with it. If you happen to be from the same social group, you might not notice anything odd because it’s basically your worldview more or less reflected back to you.
We’re all going to need to become critical readers and critical consumers of AI
But if you’re from a different social group, AI outputs can sometimes feel off. Maybe you’ve spotted biases. I know I have! Sometimes it’s more subtle and it just kind of slips by. But when it’s our health at issue, the stakes can be high and it’s worth taking careful notice.
Sometimes we use AI to sort things through for us. We might let it make decisions for us. Both of these tasks rely on AI’s reasoning, which again can seem neutral because it’s a piece of software. But there’s implications for people and groups who are different from the dominant group. This is what my co-presenters were mainly concerned with.
Though many people are very excited about the potential AI has to help close care gaps around the world, there’s a real fear of conflict with local health frameworks, exclusion of people, exclusion of knowledge sources, and digital colonization. AI can be yet another force for marginalizing local health and language practices.
I’m very grateful to the presenters for what I learned from their research, and their contributions left me thinking some thoughts.
- What are we asking of our organizations in terms of norms and governance when we use AI agents?
- What do we need to be aware of, if AI is in the picture when we interact with people from non-dominant groups around issues of their health?
Because it seems inescapable: we’re all going to need to become critical readers and critical consumers of AI. Constantly questioning it, interrogating it, not letting it just slide by.
But this is going to be hard, because most of what we do with language escapes our conscious awareness. And that includes reading, and writing to, and interacting with AI agents.
Plus, AI sounds so convincing. That eerily creepy persuasiveness, the overconfident walls of text. And we’re all so busy, and AI helps us get through things we need to get off our plates.
There are frameworks and recommendations coming from many corners. We can begin by helping each other keep an eye on AI.
If you’re wondering where to start as an individual, I’ll humbly suggest you trust your judgment. When that spidey sense starts telling you something is off, believe it. Maybe notice what phrase or sentence got your attention. Try keeping track just for a week. You’re looking for patterns, so that you can begin to interrogate: what’s going on under the surface here?
This has been 10 Minutes to Better Patient Communication from Health Communication Partners. Audio engineering and music from Joe Liebel. Additional music from Alexis Rounds.
Thanks for listening to 10 Minutes to Better Patient Communication from Health Communication Partners, LLC. Find us at healthcommunicationpartners.com.
