Why Generative AI Falls Short: Embracing Causal Reasoning in Medicine

A few years ago I began experiencing unexplained symptoms — persistent fatigue, heart palpitations, and resistance to weight loss — despite my efforts. I consulted multiple physicians, underwent extensive testing, and each time received the reassurance that my laboratory results were normal. The implicit message was that the issue lay within me.

Through independent research and a willingness to dig deeper, I learned that my thyroid’s impaired conversion of hormones stemmed from a genetic variant and transient stress, conditions not captured by routine TSH testing. Compounding iron and vitamin D deficiencies and other nutrient shortfalls, the full picture emerged only when I examined the interactions rather than isolated markers. I was confronting insulin resistance linked to my PCOS genetics, yet the situation was more complex. Each abnormality alone appeared trivial, but together they revealed a distinct narrative.

The clinicians I encountered were not negligent; they operated within a system ill‑equipped to address the depth of information patients now present. Individuals arrive for primary‑care appointments bearing data from wearables, comprehensive lab reports, genetic panels, and microbiome analyses — information that often exceeds the capacity of doctors to interpret without additional resources.

The sequencing revolution has produced the most detailed map of human biology to date. Yet we continue to struggle to integrate that complexity into meaningful clinical reasoning at the point of care.

Standard laboratory reference ranges wield considerable influence, yet they should not be accepted uncritically. These ranges were derived from populations that often failed to reflect diverse life stages, genetic backgrounds, and hormonal profiles. Consequently, a normal TSH reading can mask symptomatology; research indicates that many women with normal TSH still exhibit low T3, reflecting inadequate conversion of T4. This is not an anomaly but a systemic limitation of the tools we rely on, affecting men as well.

The more precise inquiry — central to personalized medicine — asks not whether a value falls within a reference interval, but whether it represents an optimal state for that individual, considering genetics, history, and system interactions. Here, AI can play a transformative role: rather than measuring biomarkers against a generic population average, it can integrate them with a patient’s genetic context to define what truly optimal looks like. Biomarkers considered in isolation cannot achieve this, nor were reference ranges designed to. Moreover, those ranges mirror the health of the populations from which they were derived; using them as preventive targets tacitly accepts a floor that has already proven too low given the rise in chronic disease over the past two decades.

The consequence is what clinicians increasingly term “faithful hallucinations” — outputs that appear confident and clinically plausible yet are factually inaccurate. A recent study found that 91.8 % of clinicians using AI tools encountered such hallucinations, and 84.7 % believed they could jeopardize patient safety. For instance, an AI might claim that DHEA serves as a precursor to both cortisol and sex hormones; while this sounds authoritative to a non‑specialist, the underlying pathway does not support the statement. Detecting such errors can erode trust, whereas a failure to detect them may lead to harmful actions. The alternative lies in AI that operates with transparent, causal reasoning, grounding each recommendation in validated biological mechanisms, illustrating the logical pathway, and providing citations.

Clinicians are aware that a primary‑care physician typically has only 15 minutes per patient and is trained to address population‑level risk rather than to pursue extensive panel testing for patterns that standard protocols overlook. The existing system was not engineered for the cross‑system reasoning that personalized care demands, nor does it afford patients visibility into the extensive behind‑the‑scenes analysis that occurs beyond face‑to‑face encounters.

Consequently, the depth of care a patient receives often hinges on geography and financial resources. A functional‑medicine specialist who can simultaneously analyze endocrinology, immunology, and genomics remains out of reach for many. Currently, such comprehensive care is largely a concierge service. A primary‑care physician in rural Arizona possesses the same potential as a specialist in Manhattan, but lacks the supporting infrastructure — a disparity that is bridgeable.

Transparent, causal AI — systems that link biomarkers to genetic context, disclose their reasoning, and cite sources — does not supplant clinical judgment; it enhances it. By providing primary‑care providers with cross‑system pattern recognition that would otherwise require years of subspecialty training, such AI makes advanced diagnostic capability accessible at the point of care, irrespective of geographic location.

Trust in clinical AI will not arise from superior design or smoother interfaces alone; it will be earned by providing clinicians with transparent tools that cite sources and demonstrate biological reasoning, allowing physicians to interrogate the logic before acting and to identify errors. Such infrastructure expands the realm of what is possible at the point of care. It does not replace clinical judgment but equips clinicians with resources that match the complexity of the data patients now present. The traditional framework that treats the body as a collection of isolated components, benchmarked against a non‑universal population average, is not the ceiling of medical possibility — it is merely the point from which we have evolved. The gap between the current state and what can be achieved is tangible, measurable, and warrants decisive action.

Photo: peterhowell, Getty Images

Elena Ikonomovska, PhD, serves as Co‑Founder and Chief Executive Officer of Diadia Health, a platform that employs causal reasoning to assist clinicians in managing complex chronic diseases. With nearly two decades of experience, she holds a doctorate in machine learning. Her professional trajectory includes positions at Google, where she helped shape technology that later formed the basis of BigQuery, and at Reddit, where she functioned as the company’s inaugural data scientist, developing content recommendation systems for identifying harmful behavior.

Among her other endeavors, she co‑founded Nuntio Labs, an AI‑driven writing platform that attracted clients including Wells Fargo and Square, served as Head of AI at Change.org, establishing and directing the machine‑learning and AI department, and co‑founded Mnemonic, Inc., assuming the roles of CEO and Chief AI Officer to create AI‑enhanced blockchain analytics. Following a prolonged personal health struggle that resulted in repeated dismissal of her symptoms, she launched Diadia Health to address the gaps she experienced within conventional medical paradigms.

Source link