Imagine asking an AI a simple question.
“Name the five Black films every kid should watch.”
Now imagine it answers:
- Driving Miss Daisy
- The Help
- Green Book
- Crash
- The Blind Side
Technically, none of those answers are wrong. They’re all films connected to Black experiences. But if you grew up in Black communities, something immediately feels…off.
The problem isn’t that AI failed to recognize Black films. The problem is that it couldn’t distinguish between movies about Black people and movies that shaped Black culture. It knows the titles. It doesn’t know what those stories meant around our dinner tables, in our classrooms, at family reunions, or in conversations that stretched late into the night.
That distinction turns out to matter far beyond movies.
RLiS Quick Take
What is a Large Language Model (LLM)? An LLM is a type of artificial intelligence trained on enormous collections of text so it can recognize language patterns and generate human-like responses. Tools such as ChatGPT, Claude, Gemini, and Microsoft Copilot are all examples of LLMs. Rather than “thinking” like a person, an LLM predicts language based on patterns it learned from the information it was trained on (Bommasani et al., 2021).
That means an LLM doesn’t know something because it experienced it.
It knows something because it has seen enough examples to recognize a pattern.
AI Learns From the History We Preserve
As AI becomes more integrated into healthcare, those patterns matter more than ever. Researchers are using AI to summarize scientific literature, clinicians are beginning to rely on AI-assisted documentation, and health systems are exploring how these tools can support everything from clinical decision-making to patient communication. Every one of those systems depends on language. And language always carries history.
One of the biggest lessons emerging from the Rethinking Language in Science Literature Review is that today’s scientific language wasn’t created in a vacuum (Dopson & Gray, 2026). Many of the terms, definitions, and communication practices we now treat as “standard” were developed within institutions that often centered professional expertise over community experience. Disability justice, racial justice, and rare disease communities have spent decades challenging those assumptions—not because words alone change the world, but because words shape how the world understands people.
RLiS Quick Take
What is training data? Training data is the collection of books, research papers, websites, medical records, images, and other information used to teach an AI system how to recognize patterns. An AI model can only learn from the information it is given, which means the strengths—and the gaps—in that data shape the answers it produces (National Institute of Standards and Technology [NIST], 2024). In healthcare, that data may include:
- Medical journals
- Clinical notes
- Electronic health records
- Practice guidelines
- Health websites
- Government reports
- Published research
AI doesn’t decide what’s included. It learns from whatever people choose to write, publish, and preserve. That’s why we often say: AI doesn’t invent bias. It inherits it.
But inheritance isn’t destiny. Communities can reshape what future AI systems learn by changing what gets documented, published, and preserved.
RLiS Quick Take
What is a pattern? In AI, a pattern is a relationship a model learns after seeing many examples. The model does not necessarily understand why the pattern exists; it learns that certain words, ideas, or outcomes commonly appear together. When historical examples reflect bias, exclusion, or unequal investment, AI may reproduce those patterns as though they are neutral or complete (Bommasani et al., 2021).
Communities often recognize harmful patterns long before institutions name them. One goal of the RLiS Framework is to help AI learn from community knowledge, not only from the institutional record.
If medical literature repeatedly describes disability as a burden, AI learns that pattern. If generations of research underrepresent Black patients or overlook certain rare diseases, AI doesn’t recognize those absences as problems; it simply assumes that’s what the world looks like.
This is one of the most important shifts in how we should think about AI. The question isn’t whether the technology is biased. The better question is: What history has it inherited?
AI Doesn’t Have Cultural Memory
Communities carry memories that never make it into a dataset.
Black communities remember Tuskegee, not simply as a historical event, but as part of a larger story about trust, research, and who gets protected when science goes wrong. Disability communities remember why Rosa’s Law mattered and why debates around identity-first versus person-first language continue today. Rare disease advocates remember years of fighting to have conditions recognized, funded, and taken seriously.
Those memories shape how communities interpret language.
AI doesn’t carry those memories.
It carries whatever parts of those memories people choose to write down.
RLiS Quick Take
What is cultural memory? Within RLiS, we use cultural memory to describe the shared history, lived experience, and collective wisdom communities pass from one generation to the next. Cultural memory shapes how people understand language, trust institutions, interpret events, and assign meaning to words.
People inherit cultural memory. AI does not.
The difference becomes especially visible when we look at healthcare.
Take sickle cell disease and cystic fibrosis. Both are serious genetic conditions. Both deserve excellent research, equitable investment, and compassionate care. Yet advocates have long pointed to major differences in public awareness, funding, and research attention. Those historical patterns influence what gets studied, what gets published, and ultimately what becomes part of the evidence AI systems learn from. The disparity isn’t simply about funding. It’s also about which stories become part of scientific knowledge and which remain primarily within community memory.
AI doesn’t choose which stories become visible.
It reflects the stories our institutions have already chosen to tell.
That realization changed how we think about RLiS.
At first glance, Rethinking Language in Science appears to be a project about words. In reality, it’s a project about context. Our literature review found important language frameworks already exist, but few bring together disability justice, racial justice, rare disease advocacy, and the emerging realities of healthcare AI into one community-informed approach (Dopson & Gray, 2026).
That’s the opportunity we see.
Rather than asking AI to simply use different words, we believe we need to teach it why those words matter in the first place.
RLiS Quick Take
What is the RLiS Framework? The RLiS Framework is a community-informed resource being developed to help researchers, clinicians, policymakers, advocates, and AI developers make more thoughtful language decisions. Rather than prescribing one “correct” way to communicate, the framework is designed to provide:
- Community principles and values
- Language and pattern libraries
- Decision-making guidance
- Implementation strategies for healthcare and AI
- Practical resources that organizations can integrate into their own work
Healthcare AI will continue to evolve. That’s no longer the question. The question is whether the communities most affected by these technologies will help shape the knowledge they learn from.
The RLiS Framework exists because AI shouldn’t have to guess who we are. It should learn from us.
The Conversation Continues
If AI is learning from our history, what parts of our history are still missing from the data?
References
Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., … Liang, P. (2021). On the opportunities and risks of foundation models. Stanford Center for Research on Foundation Models. https://arxiv.org/abs/2108.07258
Dopson, R., & Gray, K. (2026). Rethinking Language in Science Literature Review. The NAMED Advocates.
National Institute of Standards and Technology. (2024). Artificial Intelligence Risk Management Framework (AI RMF 1.0). U.S. Department of Commerce. https://www.nist.gov/itl/ai-risk-management-framework
