When I say … artificial intelligence

Artificial intelligence (AI) looms large in popular imagination, from Shelley's Frankenstein to Kubrick's HAL9000. But AI also has been a significant topic of research for computer scientists, health informaticians, educational technologists and social scientists over many decades. This broad engagement means that it is often unclear what the term AI refers to; definitions vary markedly across fields of endeavour.1 While generative AI and other large language models (LLMs) have grabbed the headlines, there is a lack of clarity about what AI is and what it might do, both broadly in society and specifically within professional educational settings. In order to clarify different modes of thinking about AI, we categorise different conceptualisations and associated definitions. These are not exclusive nor is one approach is better or more correct. Rather, they serve different purposes. We start with technical definitions (what AI is) and move to capability definitions (what AI does), before exploring relational definitions (how AI works within a social system). We end by introducing the 'AI interaction', a relational conceptualisation that may be particularly valuable for health professional practice and education. Technical definitions (describing what an AI is) provide the most straightforward approach. For example, there are two main algorithmic labels for AI: pre-set approaches ('expert systems', which rely on known rules) and pattern recognition approaches ('machine learning' trained on existing datasets). The latter are in use in LLMs like ChatGPT; they employ machine learning, where statistical weighting allows software to predict users' desired patterns of text, image or audio. For those who feel that these technologies are too mysterious (and indeed they often feel magical), you may wish to think of them as highly sophisticated predictive text generators, like the ones completing your sentences on your smartphone. Technical definitions allow people to understand how AI technologies work, but they are limited in defining what an AI contributes to a particular task or situation. From its early inceptions, AI has been defined by its capabilities.2 In 1980, Searle3 classically separated out 'weak' AI, which is where the AI acts as a technological tool under the control of humans, from 'strong' AI which 'can literally be said to understand and have other cognitive states'. Strong AIs—or conscious machines—remain the stuff of science fiction. Current AIs are generally designed as technological tools and therefore are often defined by what they are capable of doing (rather than by the underlying algorithms). Medical education scholars tend towards capability definitions. For example, Tolsgaard et al4 cite the Oxford dictionary, describing the capabilities of AI as '… to perform tasks normally requiring human intelligence, such as visual perception, speech recognition, and decision-making …'. Indeed, many capability definitions focus on decision-making. A classic undergraduate computer science textbook defines AI as a technology that seeks to identify the 'best possible action in a situation'.5 Moreover, particular AIs are also defined by their specific capabilities, such as image classifiers in pathology or driverless cars. Technical and capability definitions are useful because they clearly delimit AI's scope, but in a decontextualised manner. Relational definitions address how AI and people work together within practice situations. These tend to draw from the theoretical foundations of science and technology studies (STS), which generally position all technologies as being social actors. From this perspective, technologies can be understood as relational, in the sense that their meaning and function is found in the ways that they are put to use.6 Along this line, Johnson and Verdicchio1 propose an 'AI system' as '… sociotechnical ensembles which are combinations of artefacts, human behaviour, social arrangements and meaning'. This type of relational definition makes most sense when thinking about pedagogy in medical education. Our students must learn to engage with a complex messy world of care—full of ambiguity, flesh-and-blood experiences and rife with emotions—and simultaneously work with and around AI.6 A relational framing conceptualises AI as situational and dynamic rather than a fixed artefact. Importantly, it foregrounds the issue of ambiguity. And AI use is surprisingly ambiguous. For example, a 2019 ethnography exploring radiologists' responses to the introduction of AI to perform bone age assessments, suggests that AI increased the doctors' sense of uncertainty amidst the burden of deep care for their patients. One participant said: 'Sometimes it (the AI) would give me bone ages that would make me re-think what I said and go, "Ok, maybe". And I would adjust closer to it (the AI assessment). But sometimes, I think, "This is way off". So I don't know. I just don't know …'.7 This is not isolated: a 2024 ethnography describing similar challenges of AI in pathology underlines the need to prepare our graduates for the reality of AI in patient care rather than the promise.8 The 'AI interaction'9 is a concept that focuses on the dynamic and indeterminate relationship between people and AI. This relational definition is concerned with what happens within a particular moment of use between a human and an AI and therefore can help with managing the realities of AI in practice. To give a specific example, a calculator is not generally considered an AI, but a 4-year-old must trust a calculator's outputs without any way of knowing whether it is right or wrong. We suggest, therefore, that when such a child uses the calculator, this is an AI interaction. But when an adult uses the calculator, it is not. Thus, 'AI' is not dependent on the technological specifications or even what it might do, but on the relationship between the human and the technology. The formal definition of an AI interaction is when: 'in the context of a particular interaction, a computational artefact provides a judgement to inform an optimal course of action and that this judgement cannot be traced'.9 This is not a matter of a person knowing whether a technology is producing a right answer or not. This is about knowing how to trace whether the technology is right or wrong. In this way, an AI is as much dependent on who is using it, as what the technology is or what its capabilities are. A further example, if a doctor asks an LLM about a medical treatment in which they are the world expert, it will not be an AI interaction, because the doctor is already familiar with the sources that the LLM is drawing from. But for a layperson, they must take the answer on trust; at that particular moment, there is no way of knowing, no manual at hand, no means of looking inside the 'black box'. In other words, any AI interaction by definition must involve a leap of faith. This is a definition that may feel uncomfortable. How can any technology contribute to an AI interaction? And what if the human thinks they can trace a judgement but cannot or vice versa? We argue these are immaterial: the concept of an AI interaction centres on the leap of faith in a moment in time. It does not matter what the technology is or is not—or what it can or cannot do. It foregrounds how, with AI, people need to take something on trust. Our ethnographic examples describe how radiologists and pathologists alike have found their practices were fundamentally altered in unexpected ways by AI and that much of this was a matter of trust (or distrust). Thus, thinking about an AI interaction may be just as helpful for understanding AI in practice, as learning to use a particular AI software, or understanding the reliability of their algorithmic underpinnings. The 'AI interaction' introduces at a definitional level, the role of trust and doubt,9 the contextualised nature of practice and the need to grapple with the ethical implications of working with a technology where one cannot know. It eschews often unanswerable questions such as 'is this AI accurate?' to ask ourselves and our students to critically consider whether a particular AI interaction is meaningful, useful or even harmful. It attunes us to the need for students to develop distinctly human capabilities as part of these interactions. Thinking about AI interactions highlight how our curricula may need to emphasise learning to discriminate quality9 or underline compassionate care in an AI-mediated world.6 We need to direct our students to these types of interactions and hence to the fundamental problems of AI in practice. Such things can be done through simulations or case studies that emphasise complexity, ambiguity and clinical responsibility. We can counter the characterisation of the technological as necessarily superior or rational but rather enable a critical foregrounding of how work is co-produced by humans and machines in all the situated messiness of health professional practice and education. Margaret Beaman led writing the primary draft; Rola Ajjawi contributed to writing the primary draft; both critically reviewed and edited the final version. Open access publishing facilitated by Deakin University, as part of the Wiley - Deakin University agreement via the Council of Australian University Librarians. Data sharing is not applicable to this article as no new data were created or analyzed in this study.