# Gemini ## Introduction and Setting the Stage The provided material begins with a standard legal disclaimer from the University of Melbourne. This notice serves to inform the audience that the content they are about to experience is protected under Australian intellectual property law, specifically Section 113P of the Copyright Act 1968. This section of the act pertains to educational use, allowing institutions to make copyrighted material available for educational purposes under specific conditions. The disclaimer also directs individuals to the university's copyright website for further clarification, ensuring legal compliance and awareness of intellectual property rights. Following this formal notice, a brief, informal conversation takes place among attendees before the lecture officially begins. This dialogue provides context, indicating that the event is a live, in-person lecture. The speakers discuss a previous talk from the same institute, clarifying that it was a different event and not part of the "annual lecture series," suggesting a recurring set of talks. One person mentions having a two-hour lecture on their phone that they plan to listen to in segments, highlighting the modern, on-demand nature of content consumption. This casual exchange sets a relaxed, academic atmosphere before the formal proceedings commence. The official start of the lecture is marked by a host (Speaker 3), who formally introduces the guest lecturer, Doctor Brian Chapman. The host notes that Dr. Chapman gave a similar talk the previous year and was invited back, indicating the positive reception of his prior presentation. The introduction establishes Dr. Chapman's credentials and the scope of his expertise. He is presented as a colleague and friend who works in the field of **medical informatics**, a discipline that involves the application of information science, computer science, and data analysis to healthcare. The host further specifies that Dr. Chapman has "philosophical interests and ethical interests," which primes the audience to expect a talk that goes beyond purely technical details to explore the deeper implications of technology in medicine. ## The Speaker's Framework: Ethos, Pathos, and Logos Dr. Brian Chapman begins his lecture by engaging the audience directly, a common rhetorical strategy to build rapport. He first asks for a show of hands from those with expertise in healthcare, followed by a question about who has been a patient. This establishes a shared foundation: while few may be healthcare professionals, everyone in the room has experience with the healthcare system as a recipient of care. This makes the topic universally relevant. He then clarifies his own position, stating, "I am not an ethicist." This is a strategic move to manage expectations and frame his role as that of a practitioner and informed observer rather than a formal philosopher. He shares his personal ethical guideline, derived from a dialogue of Plato where Socrates asserts, "It is better to suffer an injustice than to do an injustice." By referencing this classical philosophical tenet, he signals a deep-seated concern with moral action, even while disclaiming formal expertise in ethics. This principle, which prioritizes moral integrity over personal gain or avoidance of harm, suggests a value system focused on the rightness of actions themselves, a concept often associated with deontological or virtue ethics. His stated goal is to present the context of healthcare and specific examples of Artificial Intelligence (AI), inviting the audience to engage in a dialogue about the ethical implications. Dr. Chapman explicitly structures his presentation using the classical rhetorical framework of Aristotle: **Ethos, Pathos, and Logos**. This framework is a powerful method for constructing a persuasive argument. * **Ethos** refers to the speaker's credibility and character. Dr. Chapman explains that he must first establish why the audience should listen to him. * **Pathos** refers to the emotional appeal of the argument. He aims to get the audience "emotionally involved" and to recognize the importance of the topic. * **Logos** refers to the logical and rational content of the argument. This will form the core of his presentation, where he delivers the "logic of it." ### Establishing Ethos and Pathos: A Dual Foundation of Expertise To build his **ethos**, Dr. Chapman presents two distinct but complementary sets of credentials. The first is his professional and academic background. He holds a PhD in medical informatics, which he defines in shorthand as "the application of artificial intelligence to health care." He has over 30 years of professional experience in this domain, specifically in the sub-fields of **medical imaging** (interpreting visual data like X-rays and CT scans) and **natural language processing** (enabling computers to understand and process human language, such as from doctors' notes). The second, and arguably more powerful, foundation for his credibility is what he calls "lived experience." He shares a deeply personal story: he was diagnosed with cancer at the age of seven. This has resulted in over 50 years of continuous engagement with the healthcare system, dealing with both acute (short-term, severe) and chronic (long-term) conditions. This personal history provides him with an intimate, firsthand understanding of the patient's perspective. Furthermore, having lived in Australia for the past five years after spending most of his life in the United States, he has a comparative perspective on two different major healthcare systems. He concludes this section by stating that these experiences have profoundly "shaped what I have wanted to do as a professional," linking his personal journey directly to his professional motivations and lending a powerful sense of **pathos** to his introduction. He also mentions a poignant comment from his PhD advisor, who noted that Chapman's "mistake" was wanting "to know what it all meant and not just solve my technical problem." This anecdote reinforces his identity as a deep thinker, interested in the philosophical and humanistic dimensions of his technical work. ## The Context of Healthcare: A System Ripe for Disruption Having established his credibility and the importance of the topic, Dr. Chapman moves to the **logos** of his argument, beginning with a detailed description of the context of modern healthcare. He outlines several fundamental characteristics of the system that create both the need and the opportunity for AI intervention. ### 1. Healthcare is Extremely Expensive The first critical point is the immense cost of healthcare. Dr. Chapman presents a graph comparing healthcare expenditure as a percentage of Gross Domestic Product (GDP) — the total value of all goods and services produced by a country — against military spending for various developed nations. The United States serves as a stark example, with its healthcare costs approaching a staggering 20% of its entire GDP. This expenditure is five to six times larger than its military budget, which is the largest in the world and includes a massive nuclear arsenal. This comparison dramatically illustrates the scale of healthcare spending. While other countries like Australia spend less, the graph shows that for all developed nations, healthcare represents a massive and growing financial burden. This economic reality creates a powerful incentive to explore technologies like AI that might help control or reduce these costs. ### 2. Healthcare is Dangerously Risky The second point is that, despite its cost, healthcare is a surprisingly dangerous activity. Dr. Chapman displays a graph from the Harvard School of Public Health that categorizes activities by their level of risk. On one end are "ultra-safe" activities like flying on commercial airlines. In the middle are "reasonably safe" activities like driving a car. Shockingly, healthcare is placed in the "dangerous" category, alongside high-risk recreational activities like mountain climbing and bungee jumping. This means that whenever a person engages with the healthcare system, there is a substantial, non-trivial risk of something going wrong, sometimes with fatal consequences. The status quo is therefore a system that is both incredibly expensive and inherently dangerous, a combination that urgently calls for improvement. ### 3. Medical Data Comes at a Cost and a Risk The data used for medical diagnosis is not benignly collected; its acquisition involves cost, discomfort, and risk. Dr. Chapman uses the example of a Computed Tomography (CT) scan, a powerful imaging technique that provides detailed cross-sectional views of the body. He explains that CT scans use **ionizing radiation**, a form of energy strong enough to knock electrons out of atoms, which can directly damage a person's DNA. This DNA damage creates a risk of developing cancer later in life. He shows a graph estimating the number of cancers caused by these routine diagnostic scans, indicating that the very act of gathering information to diagnose a disease can itself be a source of future disease. This paradox highlights the need to maximize the value derived from any data that is collected at such a risk to the patient. ### 4. Healthcare is Not Uniform Globally The delivery and nature of healthcare vary dramatically around the world. Dr. Chapman illustrates this with several graphs. * **Physician Density:** A graph shows the number of physicians per 1,000 people. Countries like Australia and many in Europe have around four physicians per 1,000 people. In stark contrast, many African nations like Rwanda and Mozambique have a number so low it is "functionally zero." This means that the healthcare systems in these regions cannot be centered around physicians; they must rely on other clinicians, such as nurses, who have less extensive training. Any AI tool designed for these environments must account for this reality. * **Technology Availability:** Another graph shows the number of high-tech medical devices, like CT scanners or mammography machines, per million inhabitants. Japan, for instance, is highly technology-oriented, while other developed countries like Sweden and Australia have comparatively fewer machines per capita. The implication is that an AI algorithm designed to work with data from an advanced technology like Magnetic Resonance Imaging (MRI) might be highly effective in an urban center like Melbourne but would have no value in a place like Mozambique where such machines are unavailable. The data that can be generated is dependent on the available technology. * **Consultation Time:** A particularly striking graph shows the average time a patient spends with a General Practitioner (GP). In Sweden, this is about 23 minutes. At the other end of the spectrum, in countries like Pakistan, Bangladesh, and China, the average consultation is a mere 2.5 minutes. This tenfold difference in time has profound consequences. A doctor in India, as Dr. Chapman learned from speaking with alumni, simply does not have the time for the detailed documentation or use of complex software tools that are standard in Australia. This means an AI tool must be designed to fit within the workflow constraints of its intended user, which differ vastly from one country to another. ### 5. Healthcare is Incredibly Complex Dr. Chapman highlights the procedural complexity of healthcare with a single, powerful example: ordering a medication in a hospital. Studies have shown that there are approximately 50 distinct steps involving different people and systems between the moment a doctor orders an antibiotic and the moment a nurse administers it to the patient. This long and convoluted chain of events creates 50 potential points of failure where an error can occur. This inherent complexity suggests a role for AI in managing, tracking, and verifying these processes to improve safety and efficiency. ### 6. Medical Knowledge is "Impoverished" A crucial philosophical point is that medical knowledge is fundamentally different from knowledge in fields like physics. Dr. Chapman, an electrical engineer by training, notes that medicine is built almost entirely on **statistical associations**, not deep **mechanistic explanations**. He introduces the term "impoverished knowledge," coined by philosopher of medicine Miriam Solomon. The primary method for generating medical knowledge is the **Randomized Controlled Trial (RCT)**. In an RCT, a population is split into two statistically matched groups; one receives a new treatment (Treatment A), and the other receives a placebo or standard treatment (Treatment B). By comparing the outcomes, researchers can determine *if* the treatment is effective. However, as Dr. Chapman explains using a quote from Solomon, RCTs are not designed to discover *how* or *why* an intervention works. They establish a correlation (e.g., taking aspirin is correlated with a lower rate of heart attacks) but often do not reveal the underlying biological mechanism. This leaves medicine operating in what philosopher Nicholas Rescher called a **"statistical fog,"** a state of inherent uncertainty where we know *that* things work for a certain percentage of people, but not precisely why, or how to make them work better. ### 7. The Human Brain is Overwhelmed by Data Compounding all these issues is the exponential growth in the amount of data relevant to any single medical decision. Dr. Chapman presents a graph showing that in the 1990s, a typical decision involved around five facts. This number is significant because it aligns with the approximate capacity of human **working memory**—the "mental workspace" we use for immediate tasks. However, by the 2020s, the number of relevant facts per decision has grown into the hundreds. This volume of information biologically exceeds the capacity of the human brain to process effectively. This cognitive overload creates a clear and compelling need for tools that can help manage and synthesize this vast amount of data. ## The Human Brain and the "Extended Mind" Having established the limitations of the healthcare system and the human mind's ability to cope with its complexity, Dr. Chapman introduces a philosophical concept to frame the role of AI: the **extended mind**. He cites the work of philosopher and cognitive scientist Andy Clark, who provocatively claims the human brain is "bad at logic and good at frisbee." Clark's point is that our biological brains are not optimized for complex, sequential, logical reasoning but are brilliant at pattern recognition, perception, and controlling physical actions. This profile may be excellent for a surgeon performing a delicate operation but is ill-suited for the kind of logical derivation required for complex diagnosis in the face of hundreds of data points. So, what makes human intelligence special? According to Clark, it is our unique ability to "enter into deep and complex relationships with non-biological constructs, props, and aids." We are masters at creating tools that augment our cognitive abilities. Dr. Chapman calls these **"information processing mergers."** Simple examples include using paper and pencil to perform long division, a task most people cannot do in their heads, or using books and writing to create a vast cultural memory that far exceeds what any individual could store in their brain. From this perspective, AI is not about creating a standalone artificial person. Instead, it is the next logical step in this long history of tool-making. AI systems are advanced information processing mergers designed to extend our capacity for reasoning, logic, and memory. This idea aligns with a vision of AI articulated as early as 1960 by J.C.R. Licklider, who hoped for a future where "human brains and computing machines will be coupled together very tightly," creating a partnership that could "think as no human brain has ever thought." This vision is not about replacing humans but about creating a human-machine symbiosis that transcends the capabilities of either alone. ## A Brief History of AI in Medicine Dr. Chapman provides a concise history of AI in medicine, outlining three major phases or "flavors" of the technology. 1. **Probabilistic AI (1960s, with a resurgence in the 1990s):** The earliest attempts, such as a 1961 paper on diagnosing congenital heart disease, were based on **Bayes' Theorem**. This mathematical formula provides a principled way to update the probability of a hypothesis being true as new evidence becomes available. In theory, it is the optimal way to reason under uncertainty. However, these early systems struggled because it was incredibly difficult to acquire the accurate probability estimates needed to make them work, especially in the pre-digital era. 2. **Logic-Based Expert Systems (1970s-1980s):** Due to the challenges of probabilistic methods, the field shifted to expert systems. The core idea was **knowledge engineering**: researchers would interview a human expert (e.g., an oncologist) and meticulously codify their decision-making process into a set of "if-then" rules. A famous example is **MYCIN**, a system developed at Stanford to recommend appropriate antibiotics. These systems aimed to directly replicate the reasoning of a human expert. 3. **Machine Learning (2000s-Present):** This is the dominant modern paradigm. Instead of being explicitly programmed with rules, machine learning models learn relationships directly from vast amounts of data. They are trained to find implicit or explicit patterns that link data inputs (like the pixels of a medical image) to classification outputs (like "cancer" or "no cancer"). ## Case Studies: AI in Action and its Ethical Dilemmas The core of the lecture consists of a series of real-world examples and thought experiments that illustrate the practical applications and ethical complexities of AI in healthcare. ### Case Study 1: The Overly-Dutiful Pharmacy Robot In 2010, a San Francisco hospital installed a $7 million pharmacy robot designed to automate the process of dispensing medication, with the stated goal of "eliminating the potential for human error." The robot would accurately pull pills, package them, and send them to the correct patient floor. The system's flaw was exposed in the case of a 16-year-old patient, Pablo Garcia. A doctor, due to a user interface error, accidentally prescribed a dose of an antibiotic that was 39 times the correct amount. The robot, programmed for precision, dutifully and perfectly counted out 38.5 pills, packaged them, and sent them for administration. A human pharmacist or technician, upon receiving such an unusual order, would likely have recognized it as an anomaly and questioned it. Their common sense would have served as an informal but critical safety check. The robot, lacking this common sense, executed its instructions perfectly but mindlessly. This case demonstrates a critical danger of automation: by removing humans from the loop, we can also remove their invaluable capacity for common-sense reasoning and anomaly detection. ### Case Study 2: The Manipulated Medical Image Cybersecurity researchers demonstrated a chilling vulnerability in the digital healthcare pipeline. Using deep neural networks, they were able to take a CT scan of a patient's lungs and either digitally *add* fake cancerous nodules or *remove* real ones. The manipulated images were visually indistinguishable from the originals to expert radiologists. In a controlled experiment, they intercepted images from a scanner, altered them, and sent them to radiologists for interpretation, who were completely fooled. The ethical implication is profound. As healthcare becomes increasingly data-driven, the integrity and **provenance** (the trusted origin and history) of that data become paramount. This example reveals a new vector for harm, whether malicious (e.g., insurance fraud, terrorism) or accidental. We must have absolute certainty that the data our AI systems are analyzing is an authentic representation of the patient. ### Case Study 3: The Imperfect Mole Classifier Dr. Chapman shares a personal story of a recent skin check. A dermatologist identified a suspicious mole. They used an AI tool—a **convolutional neural network (CNN)** trained on thousands of images—to classify it. The AI, which outperforms human dermatologists in classifying static photos, confidently declared the mole "benign." However, Dr. Chapman and his doctor decided to remove it anyway. The reason lies in the information the AI *did not have*. Dermatologists use a rule known as ABCDE to assess moles: Asymmetry, Border, Color, Diameter, and **Evolution** (change over time). The AI was trained only on the static ABCD features of a single image. Dr. Chapman, however, had photos from 15 years prior which showed that the mole was new. For a person his age, a new mole is highly suspicious. Furthermore, he had a history of previous melanomas that also appeared normal by ABCD criteria but were new. The AI had no concept of "E for Evolution" or the patient's broader medical history. The mole was indeed found to be a melanoma. This case powerfully illustrates that AI models are fundamentally limited by the data they are trained on. They operate within a narrow, pre-defined context and can be blind to crucial information that falls outside that context. This highlights the indispensable role of the human clinician, who can integrate disparate sources of information—the patient's history, longitudinal data, and clinical intuition—to make a final judgment. ### Case Study 4: AI as a Second Reader in Mammography Screening mammograms for breast cancer is a difficult task, and a significant percentage of cancers are missed by a single radiologist. A common practice in well-resourced countries is **double reading**, where two radiologists independently interpret each scan to improve accuracy. This, however, doubles the cost. An emerging application of AI is to serve as the second reader. A human radiologist and an AI tool interpret the scan, and if they disagree, the case is flagged for further review. Sweden has already adopted this as a national policy. This raises an ethical question: could we take it a step further? Could we use the AI to triage the scans, automatically clearing those it deems "normal" with high confidence so that a human never has to look at them? This would save significant resources, but it introduces the risk of an AI error leading to a missed cancer. This forces a difficult trade-off between efficiency, cost, and the acceptable level of risk. ### Case Study 5: The Problem of Data Portability Dr. Chapman discusses two historical examples that reveal a fundamental problem with data-driven AI: **databases don't travel.** * In the 1970s, a researcher named F. T. de Dombal developed a highly accurate Bayesian AI tool in Leeds, England, for diagnosing acute abdominal pain. When the tool was implemented in a hospital in Copenhagen, Denmark, its accuracy plummeted. The reason was not that Danish and English physiology are different, but that the *practice of medicine* was different. The questions doctors asked, the tests they ordered, and the patient populations they saw were different, meaning the statistical patterns in the data were different. * A more recent example involves a sepsis detector developed by a major software vendor in the United States. The deep learning model worked remarkably well on US patient data. When ported to Australia, using the exact same software, it performed poorly. Again, the reason was differences in clinical practice. A key lab test used for early sepsis detection in the US is ordered later and less frequently in Australia. The AI had learned to rely on this early data point, which was simply absent for most Australian patients. These examples lead to the conclusion that clinical data is not an objective reflection of reality; it is **"formed, not found."** It is the product of specific human decisions, cultural norms, and institutional processes. This means that an AI model trained in one location may not be generalizable to another, a phenomenon captured by the title of one paper: "The Ghost in the Machine Has an American Accent," highlighting the dominance of US-centric data in AI development. ### Case Study 6: The AI Scribe The final example is the emerging technology of AI scribes. These tools use ambient listening to record a doctor-patient conversation and then use a Large Language Model (LLM) to automatically generate a clinical summary note. This could be a game-changer in time-pressured environments like India or Bangladesh. Dr. Chapman shares an experience where he and a colleague simulated a consultation using such a scribe. The resulting note was about 80% accurate, which was impressive. However, the 20% it got wrong included major errors, such as misstating his cancer history. While the technology is not yet perfect, studies suggest that even if it doesn't save time (because the clinician must carefully review and correct the note), it can significantly reduce **cognitive load**. By automating the tedious task of documentation, it frees up mental energy and may help reduce clinician burnout. ## Conclusion: Automating Thought to Free the Mind Dr. Chapman concludes with a provocative quote from philosopher Alfred North Whitehead: "Civilization advances by extending the number of important operations we can perform without thinking about them." Whitehead argues against the truism that we should always think about what we are doing. Instead, he suggests that we should automate routine operations to conserve our finite cognitive resources—our "fresh horses"—for the "decisive moments." This provides a powerful philosophical framework for the role of AI in healthcare. The goal is not to create a perfect, all-knowing machine to replace flawed humans. Rather, the goal is to build tools that can handle the routine, data-intensive, and logical tasks that our brains are not well-suited for. This frees up the human clinician to focus on what they do best: integrating complex context, exercising common sense, communicating with empathy, and making the final, nuanced judgments that require wisdom. The ultimate vision is a symbiotic partnership, coupling imperfect AI tools with imperfect human decision-makers to achieve a system of care that is, hopefully, slightly less imperfect. During the subsequent Q&A, the concept of **abduction** is introduced. Described as a third mode of inference alongside deduction (logical proof) and induction (learning from patterns), abduction is "inference to the best explanation." It is the creative, hypothesis-generating process of seeing a set of facts (e.g., symptoms) and reasoning backward to the most likely cause. This form of common-sense reasoning is central to medical diagnosis and represents a significant gap in the capabilities of current AI systems, further underscoring the need for human-machine collaboration.