AI Biases in Healthcare: Exploring With Real-Life Examples
Published: 14 Jun 2025
AI is now helping doctors to make faster and smarter decisions. It reads X-rays, checks patient records and even suggests treatments. Sounds amazing, right?
But there is a problem, AI can make mistakes, and sometimes those mistakes are unfair.
In one real case, an AI tool used in U.S. hospitals gave less care to Black patients than white patients even when both were equally sick. Why? Because the AI learned from old data that contains more detail of white patients as compared to black patients. That’s called AI bias in healthcare and it’s a growing concern.
This article will show you how and why AI makes unfair choices. We will also discuss real-life examples that reveal the truth behind the screen.
What Is AI Bias in Healthcare?
AI is being used more and more in hospitals. It helps doctors find patterns, make diagnoses and customize personalized treatments. But while AI can do amazing things, it’s not always fair. Sometimes, it makes decisions that favor one group over another and that can be dangerous.
This unfair behavior is called AI bias in healthcare. It means the AI tool doesn’t treat every patient equally. Even if it’s not done on purpose, the results can hurt real people.
Here’s how AI bias can show up in healthcare:
- Unequal care
Some patients may get less attention or fewer tests just because of their race, gender or age. - Missed diagnoses
AI might fail to spot problems in certain groups if it wasn’t trained on enough diverse data. - Wrong treatment plans
The system may suggest treatments that work better for one group but not others. - Inaccurate risk scores
AI tools may wrongly predict who is more at risk of getting sick or needing care. - Fewer healthcare resources
Some patients may be left out when AI decides who gets extra help or follow-up care.
What Causes AI Bias in Healthcare?
AI systems in healthcare don’t create bias on their own. They reflect the information and assumptions built into them by data, developers and decisions.
The root cause is Imbalanced data and poor design choices.
When AI tools are built using data that lacks diversity, they often ignore or misjudge people from underrepresented groups. These tools might seem “smart,” but they carry the same errors that already exist in our healthcare systems.
Below are five common causes of AI bias in healthcare that experts and hospitals must take seriously:
1. Historical Bias in Healthcare Data
Most healthcare AI tools are trained on historical data from electronic health records, billing systems or insurance claims. But if these records reflect a past filled with discrimination or unequal access to care, the AI will learn those same patterns.
If minority patients historically received fewer scans or tests, AI might learn to recommend less care for similar patients in the future.
2. Lack of Representation in Training Data
AI models need large datasets to learn. But often, these datasets overrepresent certain populations like white, middle-aged males and underrepresent others, such as women, the elderly or ethnic minorities.
Without equal representation, the AI struggles to make accurate decisions for everyone.
A dermatology AI tool trained mostly on images of light skin tones performed poorly when diagnosing skin conditions on darker skin tones.
3. Poor Validation Across Diverse Populations
Before using an AI tool, it must be tested or validated on a wide range of patient types. But in many cases, developers skip or limit this step. As a result, the tool may work well in a lab setting but fail when used in diverse, real-world clinical environments.
An AI model might predict heart disease accurately in men but miss warning signs in women if it was not tested thoroughly on both.
4. Biased Design Assumptions
Some tools are built with flawed logic. For example, using healthcare spending as a measure of health needs sounds practical but spending is influenced by access, insurance and income not just sickness.
In a case study, A widely used risk-scoring algorithm in the U.S. assigned lower risk scores to Black patients compared to white patients with the same health conditions. Why? Because it assumed patients who cost more were sicker. In reality, Black patients were receiving less care not because they were healthier but because of systemic inequities.
5. Unintentional Bias from Developers and Institutions
Even well-meaning data scientists and healthcare leaders can overlook how bias sneaks into algorithms. Without a strong focus on fairness and ethics, these blind spots lead to flawed tools that affect real lives.
AI reflects the choices of those who build it. If fairness is not a goal from the start, bias becomes a built-in feature not a bug.
Real-Life Case Studies of AI Bias in Healthcare
AI tools in healthcare promise better decisions, faster results and more personalized care. But when built on biased data or flawed assumptions, they can cause serious harm. Below are five real-world case studies that reveal the dangers of AI bias, especially when lives are on the line.
1. Racial Bias in Hospital Risk Predictions
An AI tool used across U.S. hospitals was built to identify patients who needed extra medical care for chronic conditions like diabetes or heart disease. The idea was simple: flag high-risk patients and offer them more support.
But the problem was that it gave lower risk scores to Black patients, even when their health conditions were just as serious as white patients.
Why it happened:
The algorithm used past healthcare spending as a way to guess how sick someone was. The logic was that sicker people spend more on care. But here’s the hidden bias Black patients have historically received less care, so their spending data was lower. The AI saw this and mistakenly assumed they were healthier.
What was the impact?
Because of this flaw, the system underestimated the needs of Black patients. In fact, one study found that only 20% of the Black patients who should have received extra help were flagged by the tool. In contrast, 46.5% of white patients were identified for special programs.
This means thousands of people didn’t get the care they urgently needed just because the data was biased.
Source & Findings:
- Reported in Wired discussing a 2019 Science study from Obermeyer et al.
- The algorithm was in use for tens of millions of patients
2. AI Dermatology Tools Struggle on Dark Skin
AI tools designed to detect skin conditions like melanoma or eczema are being used more and more by dermatologists and telehealth apps. These tools analyze photos of skin and compare them to thousands of images in their database.
But most of those images are of light-skinned people.
Why it happened:
Many datasets used to train AI systems in dermatology don’t include people with darker skin. Some datasets had fewer than 5% images of non-white patients.
What was the impact?
When these tools are used on patients with darker skin, the AI often fails to recognize key signs of disease. One study in Nature Medicine found that performance dropped by nearly 40% for dark skin tones.
That’s dangerous. It means life-threatening diseases like melanoma could go unnoticed in people of color just because the AI wasn’t taught to see them.
Source & Findings:
3. Kidney Test Algorithm Delays Transplants for Black Patients
The eGFR test is used to check how well a patient’s kidneys are working. It plays a key role in deciding when someone needs to see a specialist or join a transplant list.
But for decades, the formula used a “race correction” that adjusted the result upward for Black patients. It made them appear to have better kidney function than they actually did.
Why it happened:
The original formula was based on an old assumption that Black people have higher muscle mass. But this generalization wasn’t always true.and it caused harm.
What was the impact?
Because of this race adjustment, many Black patients waited longer for referrals and transplants. In some cases, patients were told they were ineligible for a kidney transplant when they actually were.
After experts removed race from the formula, one hospital (Penn Medicine) found that over 14,000 Black patients were affected. Many were reclassified as needing care sooner.
Source & Findings:
NJEM/New England Journal research & JAMA report
4. VBAC AI Tool Lowers Options for Women of Color
Hospitals used an AI-powered calculator to estimate the chances of a successful Vaginal Birth After Cesarean (VBAC). It asked for details like age, weight and number of past births but it also included race and ethnicity.
Why it happened:
The tool was built using past hospital data, which already had racial disparities in outcomes. The calculator predicted lower chances of success for Black and Hispanic women even when their medical histories were similar to white women.
What was the impact?
Doctors were less likely to recommend VBAC to women of color. They pushed for cesarean deliveries instead. Many women didn’t even know the tool was used to influence their birth plan.
After public pressure, several hospitals dropped the race factor from the tool.
Source & Findings:
PubMed study on VBAC calculators
5. Pulse Oximeters Give Inaccurate Readings on Darker Skin
Pulse oximeters are devices used in nearly every hospital and home to check oxygen levels. They were widely used during COVID-19 to detect low oxygen.
But research found that these devices are less accurate for people with dark skin. The light sensors can’t always read correctly through higher melanin levels.
Why it happened:
The devices were tested and calibrated mainly on people with lighter skin tones. This lack of diversity in testing led to biased results.
What was the impact?
During the pandemic, some Black patients were sent home even though their oxygen levels were dangerously low. The oximeter showed false-normal results. This led to delayed care, longer hospital stays and even deaths.
A 2020 study in the New England Journal of Medicine found that Black patients had three times the risk of hidden hypoxia.
Source & Findings:
- Wikipedia & multiple studies
- FDA issued a warning in response
What Are the Risks of AI Bias in Healthcare?
AI is helping doctors make faster and better decisions. But when AI tools carry hidden biases, they can cause serious harm. Let’s look at some of the major risks:
Patients May Not Get the Care They Need
Biased AI tools may ignore or overlook certain patients. For example, if a Black patient has the same health issue as a white patient but the AI scores them as lower risk, they might not get treatment in time. This can lead to worsening health or even life-threatening delays.
As we saw earlier, an algorithm wrongly told hospitals that many Black patients didn’t need extra care, when they actually need.
Some People May Get Wrong or Delayed Diagnosis
When AI misreads skin conditions, breathing problems or other symptoms especially on dark skin tones, it may give the wrong answer. This can lead to wrong treatment or no treatment at all.
Pulse oximeters gave normal oxygen readings to patients who were actually in danger. These false results meant people didn’t get urgent care.
Trust in Healthcare AI May Go Down
When patients see AI making unfair or unsafe choices, they may stop trusting it. Doctors might also feel unsure about using AI tools. This creates a barrier between tech and care and slows down helpful progress.
Trust is the heart of healthcare. If people stop trusting AI, it becomes harder to bring new tools into hospitals and clinics.
It May Widen Health Gaps
Bias in AI can make health differences worse between groups. People of color, women, older adults and people with disabilities may already face barriers to good care. AI bias adds to that problem by giving unequal answers.
Instead of making healthcare more equal, biased AI may do the opposite, making the gap bigger.
How Can We Fix AI Bias in Healthcare?
AI bias is not just a tech problem, it’s a people problem. To build fair and safe healthcare tools, we must work together. Here’s how we can reduce AI bias and make care better for everyone.
Use Better, More Diverse Data
Many AI tools learn from old hospital records. But if those records mostly come from white, male or wealthy patients, the AI will miss others.
What to do:
Add more patients from different races, ages, genders and backgrounds. This helps AI tools understand everyone, not just a few.
Example:
Adding skin images from people of color helped improve how AI spots skin diseases in darker tones.
Test AI on All Groups
An AI tool might work well in one group but fail badly in another. That’s not fair and it’s not safe.
What to do:
Before using AI in hospitals, test it on all kinds of people, young and old, men and women, people of every race and background.
Example:
During COVID-19, pulse oximeters worked poorly on darker skin tones. Testing early on different skin colors could have prevented that.
Involve Doctors, Patients and Tech Experts Together
Doctors know patients. Patients know their needs. Tech experts know AI. When they all work together, they can spot problems early.
What to do:
Include real voices in every step of AI development from design to testing to real use.
Why it matters:
Many tools that failed because no one asked the right people the right questions.
Make AI Rules and Checks
We can’t fix bias if there are no rules. Right now, many AI tools get used without strong checks.
What to do:
Governments, hospitals and tech companies must create clear rules and safety checks. These can guide what’s allowed, what’s not and how to measure fairness.
Good sign:
Groups like the FDA and WHO are already working on AI ethics in healthcare. That’s a strong step forward.
Conclusion
AI can do amazing things in healthcare. It can spot diseases faster, help doctors make better choices and even save lives. But when it’s built on unfair data or limited thinking, it can also make mistakes especially for people from different races, ages or backgrounds.
We’ve seen real examples where biased AI tools gave the wrong answers, delayed care or widened the gap in health services. That’s why it’s so important to fix these issues now, not later.
Let’s build AI that works for everyone. Because healthcare should treat all people equally and AI should too.
Frequently Asked Questions about AI Bias in Healthcare
Here is the list of FAQs:
Most patients don’t know when AI is being used because doctors often don’t mention it. You can ask your doctor directly if any AI tools are helping with your diagnosis or treatment plan. Some hospitals are starting to inform patients but it’s not required everywhere yet.
In most cases, you can ask your doctor not to use AI tools for your care. However, some AI systems are so integrated into hospital operations that avoiding them completely might be difficult. Your best option is to discuss your concerns with your healthcare provider and ask about alternatives.
Currently, there are very few specific laws protecting patients from AI bias in healthcare. Most regulations focus on general medical device safety rather than fairness across different groups. However, organizations like the FDA are working on new guidelines for AI tools in medicine.
Fixing AI bias can take months or even years depending on the complexity of the system. The process involves collecting new data, retraining the AI and testing it thoroughly before it can be used again. Some quick fixes like removing race factors from calculators, can happen faster.
Many doctors don’t realize the AI tools they use might be biased because the bias is not obvious during normal use. Medical schools are just starting to teach about AI bias, so many practicing doctors haven’t learned about these issues yet. This is why ongoing education and awareness are so important.
Yes, children can be especially vulnerable to AI bias because many AI systems are trained primarily on adult data. Pediatric conditions may be missed or misdiagnosed if the AI hasn’t learned enough about how diseases appear in children. Kids from minority backgrounds face double risk from both age and racial bias.
AI bias can be more problematic in hospitals that serve diverse populations but use AI tools trained on less diverse data. Safety-net hospitals and community clinics which often treat more minority patients, may see bigger impacts from biased AI systems. However, the bias exists in the AI tool itself, not necessarily the hospital.
Creating fair AI systems requires significant investment in diverse data collection, extended testing and ongoing monitoring. While exact costs vary, companies may need to spend 20-50% more on development to ensure fairness. However, the cost of biased AI through lawsuits, harm to patients and lost trust can be much higher.
Currently, no AI system is completely free from bias but some are better than others. The key is continuous testing and improvement rather than claiming perfection. Patients should look for healthcare providers who are transparent about their AI use and actively work to address bias issues.
Start by discussing your concerns with your doctor and asking for a second opinion if needed. Document your experience and consider reporting it to your hospital’s patient advocacy department. You can also file complaints with medical boards or contact patient rights organizations for guidance on next steps.
