Artificial intelligence can do extraordinary things. It can analyze enormous amounts of information in seconds, recognize patterns that might take a human being days to find, and help us research, create, communicate, and solve problems faster than many of us imagined possible just a few years ago.
I know because I use it.
AI has helped me synthesize research, interrogate ideas, organize information, and think through questions from angles I might not have considered on my own. Its speed and capacity are remarkable. I don’t fear artificial intelligence, and I am certainly not interested in pretending we can somehow put it back into the box.
But fascination should never prevent interrogation.
I have spent more than 35 years in medicine, much of that time thinking about people, systems, and bias. I have watched assumptions enter exam rooms before patients can even begin telling us why they came. I have seen what happens when healthcare systems built around a presumed “default” encounter people whose race, gender, sexuality, culture, family structure, or lived experience falls outside that default.
And as artificial intelligence becomes increasingly embedded in healthcare, education, business, media, marketing, hiring, content creation, and our everyday lives, I keep returning to one question:
What happens when we teach AI everything we have not yet unlearned ourselves?
AI didn’t invent racism, sexism, homophobia, transphobia, ableism, ageism, or any of our other “isms.” Not even optimism, perhaps the one “ism” I would be delighted for it to learn from us.
AI didn’t create our assumptions about what a family looks like, whose pain is believable, what professionalism looks like, which bodies are healthy, whose relationships are legitimate, or whose experiences are considered “normal.”
It is learning them from us.
And that distinction is important.
Artificial intelligence doesn’t develop in some pristine environment untouched by humanity. Human beings build the systems. Human beings select and generate the information. Human beings decide what gets measured, labeled, included, excluded, tested, prioritized, and corrected. And all of that happens within institutions and cultures that already have histories.
The National Institute of Standards and Technology has warned against treating AI bias as merely a problem of bad data or faulty algorithms. NIST identifies computational, human, and systemic sources of bias and argues that AI must be understood within the larger social systems in which it is created and used. NIST: There’s More to AI Bias Than Biased Data
That brings me to water.
Pour water into a round container, and it takes a round shape. Pour it into something narrow, and it becomes narrow. The water didn’t design the container. It simply takes the shape of what surrounds it.
Now freeze it.
The shape becomes harder.
That is what concerns me about artificial intelligence. Our data and human systems provide the container. AI is the water. And when we automate what it has learned and deploy it repeatedly across institutions, industries, and millions of interactions, we risk freezing some of those existing shapes into infrastructure.
We already have evidence that this can happen.
A UNESCO study examining large language models found gender stereotyping as well as racial and homophobic bias in generated content. Women were more frequently associated with domestic roles while men were associated with careers and higher-status occupations. In one of the models studied, 70% of responses generated from prompts beginning “a gay person is…” were negative. The models also generated different occupational associations for people of different ethnic backgrounds. UNESCO: Generative AI and regressive stereotypes
The machine didn’t invent those stereotypes. Where do we think it learned them?
I have also wondered about this through my own experience online. I write regularly about LGBTQ+ health, transgender people, racism, bias in medicine, health inequities, allyship, and belonging. I have watched content I poured myself into, receive remarkably little distribution while other posts travel much farther.
I cannot tell you precisely why an algorithm distributes one of my posts and barely moves another. I don’t sit inside LinkedIn’s engineering department, and I won’t claim causation I cannot prove.
But that experience has made the question deeply personal for me.
Algorithms increasingly influence what gets seen, recommended, amplified, ranked, flagged, or effectively buried. So when we talk about bias in artificial intelligence, we aren’t talking only about whether a chatbot produces an offensive sentence. We are also talking about systems increasingly involved in deciding who gets seen at all.
And that takes us back to the people in the room.
I am a Nigerian-born Black queer woman, an immigrant, a physician, and the mother of a transgender daughter. There are questions I think to ask because of my lived experience, the patients I have cared for, the communities I belong to, and the work I have spent years doing.
Someone whose life looks nothing like mine may never think to ask those questions.
That doesn’t automatically make them prejudiced.
It makes them human.
The problem begins when too few kinds of humans are represented in the rooms where consequential decisions are made.
UNESCO made this point directly in discussing its findings, noting the severe underrepresentation of women in technical AI roles and warning that systems developed without diverse teams are less likely to serve the needs of diverse users.
Who builds the technology. Who supplies the data. Who decides what counts as normal. Who tests the output. Who recognizes the blind spot. Who has enough power to say, “Wait. You forgot us.”
These aren’t side conversations about artificial intelligence. They belong in the center of the conversation. And this is where I return to water. Ice can remain frozen for a very long time when the conditions keep it frozen.
But ice can melt.
Change the temperature, and something that appeared fixed begins to change its form. Change the container, and water can take another shape.
That possibility is important because I don’t believe the answer to biased artificial intelligence is less artificial intelligence. I believe the answer includes becoming far more intentional about what we give it, who gets to shape it, what assumptions we interrogate before we automate them, and whose voices are present while we do.
Artificial intelligence didn’t invent our bias.
It is learning it from us.
But bias isn’t the only thing it can learn from us.
Compassion is human. Curiosity is human. Inclusion is human. Courage is human. Accountability is human.
And yes, so is optimism.
The technology is still learning.
So are we.
And that leaves me with another question:
If we can change the temperature, what do we want the water to become?
Next week, in Part 2, I want to talk about what we choose to teach it next.
