Algorithmic Biases and Machine Learning. Can Artificial Intelligence Exacerbate Health Inequalities?

INTRODUCTION
Popular culture is replete with examples of technologies capable of providing automated medical assistance. One of the best-known is Baymax, a character from Disney’s Big Hero 6 (Gavero & Guerrero, 2017). This robot is designed to identify the need for medical attention—for example, by recognizing that its user is in pain—and to offer a diagnosis and prescribe treatment accordingly (similar to a family doctor, but available 24/7, with unlimited access to medical knowledge and all existing diagnostic methods). At the time of its release in October 2014, devices of this kind seemed distant and even unattainable for many people. However, in recent years, with the arrival of technologies like ChatGPT and other systems based on generative artificial intelligence (AI), what once seemed like science fiction is beginning to emerge as a real possibility, even in the short or medium term.

While there are high expectations for the potential of these tools in the health sector—for example, to achieve faster and more accurate diagnoses, expand access to services, or accelerate research and innovation—it is also necessary to recognize that they carry significant risks. These risks not only include potential direct harm to people’s health (for example, through inappropriate recommendations), but can also have negative consequences for society, such as widening inequalities in access to and quality of care or even reproducing and deepening existing gaps between different groups (WHO, 2021b).

Ilustration: Monserrat García Silva

In this article, we aim to raise this debate, highlighting both the opportunities AI offers in healthcare and the risks it entails, with particular attention to its impact on equity. To make the reading more engaging, we have organized the text in a question-and-answer format. We will address the following questions: How does AI work in simple terms? What are its main applications in the healthcare field? What problems can its indiscriminate use generate, and what ethical dilemmas does it raise, particularly those related to a potential increase in inequality? And finally, what recommendations should be considered so that users can reap its benefits with the least possible risk of negative consequences?

HOW DOES AI WORK?
The World Health Organization (WHO) defines AI as “the ability of computer programs to perform tasks that are normally associated with intelligent beings” (WHO, 2021a). To fulfill this function, AI relies on algorithms, a series of procedures or instructions that allow it to solve a problem without having to invent a solution each time it arises (Abiteboul & Dowek, 2020). An algorithm can be compared to a recipe: it receives input information (the ingredients for making bread), processes it following a set of precise rules (the recipe itself), and produces an output result (the finished bread) (Abiteboul & Dowek, 2020).

To generate these algorithms, AI feeds on vast amounts of data that allow it to identify patterns and apply them to tasks of interest to humans (WHO, 2021a). Large language models (LLMs) like ChatGPT are one example, requiring enormous volumes of information for their training. Thanks to this process, they are extremely adept at recognizing language patterns and completing sentences. If you type “no por mucho madrugar” (meaning “it doesn’t matter how early you get up”), their internal algorithms clearly respond with “amanece más temprano” (meaning “it doesn’t dawn earlier”) and add an explanation of the meaning of this saying in Mexican popular culture.

WHAT ARE ITS MAIN APPLICATIONS IN THE HEALTH FIELD?
Just as AI can be trained and used to predict something as simple as the words that complete a proverb, it can also be applied—and in fact has been for many years—to solve specific medical problems. There are numerous examples of its use in this field dating back to the second half of the 20^th century (Kaul, Enslin & Gross, 2020). Since the 1990s, there have been articles on the use of neural networks—computational models inspired by the workings of the human brain that learn to recognize patterns from data—to help diagnose acute myocardial infarction (Baxt, 1991). Another example is expert systems—computer programs designed to mimic the reasoning of a human specialist using predefined rules—with the aim of producing diagnoses similar to those made by an internist (Miller, Pople & Myers, 1982).

AI applications in medicine are numerous and continue to expand (WHO, 2021a). There are different ways to classify them, and one of the most widely accepted is by their end use: there are direct applications in healthcare, such as supporting radiological diagnoses (Rajpurkar et al., 2018); in research, to advance precision medicine and make drug development more efficient (Kant, Deepika & Roy, 2025); and in the organization of healthcare systems, by identifying those who require the most urgent care (Porto, 2024). These are not the only uses, and it is important to note that they are not limited to healthcare professionals, as there are increasingly more tools aimed directly at patients and the public (Lee, Goldberg & Kohane, 2023). Therefore, their presence is becoming more common every day and is expected to become even more so in the near future.

Ilustration: Monserrat García Silva

TWO MAIN RISKS THAT AI CAN POSE IN THE HEALTHCARE SECTOR: THE DIGITAL DIVIDE AND BIASES OR FORMS OF ALGORITHMIC DISCRIMINATION

WHAT PROBLEMS CAN ITS INDISCRIMINATE USE CAUSE?
Like any technological innovation, its benefits can be accompanied by significant risks, especially if adopted without understanding its fundamental limitations (WHO, 2021a). At the individual level, there is a risk of overestimating AI’s capabilities in tasks such as medical diagnoses or recommendations, which could delay professional care or lead to inappropriate indications. There are also concerns about the ongoing collection of personal data by many technologies and the uncertainty surrounding long-term data security. In addition to these issues, focusing on equity, we want to highlight two main risks that AI can pose in the healthcare sector: the digital divide and biases or forms of algorithmic discrimination.

The so-called digital divide refers to the differences in access to and use of information technologies between groups, whether between countries or between sectors of the same population (WHO, 2021a). In practice, this means that while some people can benefit from advanced AI systems, others—due to a lack of infrastructure, training, qualified personnel, or other structural deficiencies—are limited or even excluded, which can exacerbate existing inequalities in the health sector. One example of this problem is ageism and its relationship with AI. The WHO defines ageism as stereotypes, prejudice, and discrimination against people based on their age. In fact, the WHO recently published a specific report on the topic (Ageism in Artificial Intelligence for Health, WHO, 2022) (Figure 1), in which it calls for ensuring that the design and application of these technologies do not reinforce the digital exclusion of older people.

The second problem relates to the presence of biases in algorithms and their capacity to generate, perpetuate, and even exacerbate discrimination (WHO, 2021a). These biases can arise from the data used to train AI. If the data comes primarily from countries or sectors with greater economic resources, the algorithms will mainly reflect the needs of those populations, which can lead to less accurate or even erroneous results for certain groups, such as women or ethnic minorities, and worsen pre-existing inequalities in access to healthcare.

A frequently cited example, though outside the field of AI, is found in nephrology (Eneanya, Yang & Reese, 2019). When a physician wishes to assess kidney function, they typically do so through calculations based on laboratory parameters and patient characteristics such as age and sex. However, these formulas previously included an adjustment for “Black race” that relied on biased data, leading to a systematic overestimation of kidney function in these populations, which delayed diagnoses and even access to transplants. This led to the abandonment of this variable in current formulas, but the case clearly demonstrates how the use of biased information can exacerbate inequity rather than correct it.

WHO report on ageism in artificial intelligence applied to health

WHAT RECOMMENDATIONS ALLOW US TO TAKE ADVANTAGE OF THE BENEFITS OF AI WITH THE LEAST POSSIBLE RISK?
The WHO recently proposed six principles to guide the use of AI in health:

1. Protect people’s autonomy and their right to decide on their care (it is important that patients always have the option to accept or reject recommendations generated by an AI system).
2. Promote human well-being, safety, and public interest (for example, ensuring that algorithms are validated before clinical use to avoid harm).
3. Ensure transparency and intelligibility of the systems (explain in simple terms how a diagnosis or prognosis was reached.)
4. Foster responsibility and accountability (clearly define who is responsible if an automated recommendation causes a medical error).
5. Ensure inclusiveness and equity in its development and application (incorporate data from women, older people, and minorities to avoid bias in the results).
6. Promote responsive and sustainable technologies that can be monitored, updated and adapted over time, also considering their environmental impact.

These guidelines offer us a practical framework whose objective is not to hinder innovation or limit the use of these technologies, but to serve as a guide to take advantage of the opportunities that AI offers in the field of health in a safe, equitable and sustainable way.

IF THE DATA COMES PRIMARILY FROM COUNTRIES OR SECTORS WITH GREATER ECONOMIC RESOURCES, THE ALGORITHMS WILL MAINLY REFLECT THE NEEDS OF THOSE POPULATIONS

CONCLUSIONS
AI is one of the most promising innovations in healthcare today. Its applications, which span the entire care process, have the potential to transform medicine as we know it. However, the risks associated with its use should not be underestimated, particularly the digital divide and algorithmic biases. In this regard, the WHO proposes a series of principles to guide the responsible use of these technologies. These principles are not intended to delay their implementation, but rather to serve as a safeguard that allows us to maximize their benefits and minimize their risks. 

---