System 0 and Rehabilitation Medicine: When Artificial Intelligence Becomes an Extension of Clinical Reasoning

The interview that follows stems from a dialogue between Dr. Franco Molteni, Scientific Director of the Villa Beretta Rehabilitation Research Innovation Institute (VBRRII), and Adjunct Professor Massimo Chiriatti (Università Cattolica del Sacro Cuore, LUISS), Chief Technical & Innovation Officer at Lenovo Italia. The discussion is drawn from a webinar in the Abductive Reasoning Lectures series held at the end of October 2025.

The conversation is inspired by the paper “System 0: Transforming Artificial Intelligence into a Cognitive Extension”, in which Professor Chiriatti introduces System 0, a model that describes artificial intelligence as a pre-cognitive extension capable of analyzing data, identifying relationships, and suggesting interpretations that expand human decision-making capabilities.

In his dialogue with Dr. Molteni, these themes are applied to clinical and rehabilitation practice: from data management and privacy to the role of professionals in interpreting results, and the training and organizational implications for healthcare institutions.

Franco Molteni – Has artificial intelligence changed significantly compared to a year ago? If so, how? And do you have any advice for those who are beginning to approach this field?

Massimo Chiriatti – Yes, there have been major changes, as we can read in virtually every journal and on every website. The first point is that we may be attributing excessive capabilities to this tool. The second is that we need to establish a shared definition of artificial intelligence; otherwise, everyone forms their own personal interpretation.

We can define AI as a discipline that uses very powerful computers to analyze data that has been previously collected. Without data, artificial intelligence does not exist. The task we assign to machines is to read data and, based on it, make predictions.

In recent years, these machines have also begun generating content—texts, images, audio, and video — often of very high quality. But one thing must be made clear: machines do not know what they are doing. We are the ones who initiate a dialogue with them—often a smooth, articulate dialogue—and they respond, but without any real understanding of the words. Recently, we have seen the quality of responses improve, becoming somewhat more precise and better aligned with our requests.

Why do machines respond better and better? Because they have learned the relationships between the words we have embedded in code. However, much remains to be discussed regarding semantics and meaning: some believe machines understand nothing about semantics, while others think a small degree of understanding may be emerging.

What is certain is that nearly one billion people now interact with these machines, asking them extremely complex questions or requesting tasks that we ourselves are unable to perform. This is because the data we now have access to is far beyond human scale. Thanks to these machines, we can do things that would otherwise be impossible.

Franco Molteni – There is a difference between connecting to widely available AI systems that anyone can subscribe to, and saying: “I am in my department, I have highly complex data, and I want to analyze it using artificial intelligence”. To be concrete: if I am managing a patient who brings with them social and economic information, assessment scales, blood tests, imaging exams, and even a self-description, to whom should I give all this data? To a machine within my department or to one located outside it?

Massimo Chiriatti – This is an issue that is receiving increasing attention. If we are dealing with public data, it makes sense for a company website to be hosted on a cloud platform, allowing easy and secure access. But when it comes to sensitive data—patient data, or innovative elements related to marketing, communication, or core business processes—my recommendation is to deploy systems within the organizational perimeter.

In this way, the question, or prompt, remains inside the system, which performs the analysis using its own hardware and software—often available as open source or open access. The response then returns directly to the user, without a single bit of processing leaving the organization. This approach addresses both privacy concerns and cost predictability.

With such a system, you do not pay per token—the basic unit of text typically used as currency for these services—and you can conduct all the experimentation you need. Today, hybrid models are also available, combining on-premises systems with cloud-based technologies, some purchased, and others leased. This hybrid approach prevents external access to data or the duplication of business models, while keeping system and application expertise—an important competitive factor—within the organization.

Franco Molteni – It is certainly easier to think about external subscription-based systems, but here we are talking about professional, organizational uses of artificial intelligence, which are very different from individual use. This hybrid model can therefore become a concrete clinical reality, ensuring privacy. It also raises a key question: with these systems, we are doing things that would otherwise be impossible. Beyond analyzing vast amounts of data, can we also generate something genuinely new? And how do we manage this extension of the mind without losing control?

Massimo Chiriatti – The foundation lies in monitoring patient progress. Consider a patient whose clinical history spans ten years. No human being can analyze in detail everything that has happened over that time. Historically, humans have always externalized certain functions: wheels to enhance mobility, tools to shape materials, and more recently, cognitive extensions. Even a notebook is an extension of memory; a calculator extends our mathematical capacity.

Together with an international team, we developed the theory of System 0. From an externalization perspective, we need a machine that has read the entire past and now monitors the patient, suggesting possible behaviors. For example, if a patient performed a specific action every day at 1:20 pm in the past, the machine could suggest an action that falls within the probability distribution of what that person might repeat in the future.

This is an algorithmic suggestion, not a verdict. Precisely because the machine processes enormous volumes of data much faster than we can, its output requires even stronger, broader, and interdisciplinary human judgment. That is why we need technicians, philosophers, economists, and of course healthcare professionals to interpret results that are far more than simple calculations.

Franco Molteni – So we are dealing with suggestions, not certainties. Medicine traditionally operates under uncertainty: the more complex the problem, the greater the uncertainty. Can System 0 help us reduce uncertainty, or does it risk creating a false sense of confidence? And how decisive is the human component in interpreting these hybrid systems?

Massimo Chiriatti – The paradigm differs from classical computers, which operate deductively by applying explicit rules. Here, models are trained inductively on historical data and produce probabilistic inferences. The result is not a necessary truth, but a statistical estimate. This can expand knowledge, but it always requires critical interpretation, as correlations may be spurious and lack causal value. The machine lacks context.

We call it System 0 because, as Daniel Kahneman observed, humans rely on intuitive or reflective systems for decision-making. System 0 acts as an additional “suggestor”, not an oracle, providing statistical outputs.

Moreover, machines lack a body in the biological and experiential sense. They do not feel pain, fatigue, or vulnerability, nor do they experience time as lived duration. They can optimize means, but not understand ends. Responsibility therefore remains entirely human: we must decide how and when to use these suggestions.

Franco Molteni – It may be true that machines do not have a body, yet results often emerge that make us think: “Would I have done better than this?” In many cases the machine works well, but I am convinced that it is crucial to establish at which point in our reasoning process we introduce the algorithmic suggestion.

For example: do I use the AI’s suggestion to prepare for a discussion, or do I consider it after having already reasoned through a complex problem? In clinical practice, one option could be to introduce it after reviewing the patient’s data, when it would be useful to have something predictive to support decision-making.

However, this also means that we ourselves must be able to identify meaningful relationships. So, my question is: what kind of training is needed to recognize significant relationships in one’s own field? And can artificial intelligence also help us in this regard?

Massimo Chiriatti – I am not a physician, so I will try to answer with a hypothesis that should be validated. It is useful to distinguish between efficiency and effectiveness.

If I have to examine one thousand radiographic images in a single day, after a certain number I will inevitably become tired, distracted, perhaps even disengaged. If my goal is efficiency, I will tell the machine: “You perform a first analysis of the thousand images, then I will review the results and interpret them”.

If, instead, my goal is effectiveness, I will first analyze fifteen or twenty images myself and then submit them to the machine to obtain its assessment.

Training, investment, and work on multidisciplinary skills are the three factors that help us understand this cognitive extension, which is not a mind in itself. We are extending our capabilities in a pre-cognitive way, through the machine’s observation of our past data.

One warning is essential: without data, nothing works. The system must operate on your data, integrate into your clinical and organizational processes, and be “trained” alongside your daily activity. Specific skills are required to understand algorithmic outputs and integrate them into clinical practice; otherwise, after the proof-of-concept phase, the solution will fail.

Franco Molteni – With this line of reasoning, we move beyond the logic of “saving time.” We are approaching artificial intelligence with a model similar to the one applied when machines replaced humans in repetitive tasks, freeing people to do something else.

Now, however, we are entering a phase of human evolution through hybrid reasoning systems. This evolution is influenced partly by objectives and partly by continuous interaction with artificial intelligence—especially now that AI is beginning to generate outputs autonomously. How can we keep this trajectory under control? How can we understand how artificial systems themselves are evolving?

Massimo Chiriatti – It is true that we initially used machines for automation and timesaving. But using artificial intelligence solely to save time is not ideal. The best approach is to use it to do something new, something we could not do before.

AI began in the consumer domain: people started generating texts and images. Then it moved into enterprises. Today, machines can significantly reduce time in individual processes, but I return to the initial point: we must be extremely careful about the data we input. If a collaborator enters data without awareness—poor-quality or insufficient data—the outcome will be compromised. And data must always be interpreted.

You have raised a crucial issue regarding responsibility. We need ethical and standards committees, because results depend on the values with which we use the machine. You and I can use AI for positive and beneficial purposes—to care for people and increase accessibility—but it can also be used for harmful goals: attacking individuals, creating monopolies, discriminating against those without access to technology.

One of today’s major problems is precisely the formation of monopolies, due to the concentration of data and computational power. For this reason, your idea of hybrid systems seems the most appropriate approach for optimizing treatments and healthcare resource management.

Franco Molteni – Among the many ideas you have outlined, you highlighted one key element: data quality. This is a topic we are also reflecting on internally.

Today, we already have systems capable of reading the texts we write, summarizing them, and extracting information—not only in terms of content, but even in terms of writing quality or frequency of certain terms. If we bring this into clinical practice, what we currently experience as bureaucratic work—writing in electronic clinical records—can become a valuable source of information for the system, helping it understand how a patient is described by different professionals, today versus in the past.

Thus, what each individual team member writes into the electronic medical record becomes part of the dataset on which predictions are built. This forces us to rethink how we write in clinical records and what we really want to convey. Otherwise, this large volume of documentation risks distorting the system’s reasoning. Once again, the hybrid system adds responsibility to us—it does not take it away. Am I wrong?

Massimo Chiriatti – You are right, and this also allows me to respond more clearly to your earlier question: should we accept results as they are?

We suffer from a cognitive bias—a systematic preference that leads us to think: “If the machine said it, then it must be correct.” This happens because we are used to a deductive model, like a calculator that produces a certain result. But this machine does not work deductively: its output is probabilistic. We therefore need to change our mindset and interpret the result.

One of the most effective approaches I am seeing today is to let the machine challenge us. For example: a patient may have a specific disease, I have formed an initial hypothesis, and I ask the AI: “Can you provide all alternative hypotheses that would falsify my idea?”

In doing so, the machine may generate combinations that go beyond those typically considered by a human operator, exploring the space of statistically compatible possibilities. This “novelty” is not intentional creativity, but probabilistic combinatorial exploration, which can nonetheless be cognitively stimulating for those interpreting the results.

This is helpful only if the professional understands the domain. If the machine diagnoses a disease and I am completely ignorant of medicine, blindly accepting its output would be disastrous. We must always remember that the machine does not know what truth is—and, in fact, neither do we in an absolute sense. This is a philosophical issue that has been debated for millennia, and truth must never be delegated to a machine.

Franco Molteni – These aspects must also be incorporated into training. We should ask ourselves how this will influence continuing education and initial training. We often say that, in theory, every five years we need to “regraduate” because knowledge evolves so rapidly. How will AI affect continuing medical education and initial professional training? Should these systems be systematically integrated into both?

Massimo Chiriatti – From my perspective, absolutely yes. However, there is a difficulty: the time required for training often exceeds the pace at which solutions become obsolete.

For example, we might spend years improving an AI-based blood analysis system through feedback loops that reduce errors, only for a new system with a different approach to emerge, requiring entirely new competencies.
I am convinced this is an opportunity. Every technology is a blessing for humanity, insofar as we choose how to use it. It opens unprecedented possibilities, but no technology is neutral: it embodies worldviews, organizational models, economic interests, and values. It can therefore become either a tool for emancipation or for inequality.
The outcome is not written into the technology itself, but into the political, ethical, and organizational choices that guide its adoption. That is why, even though I am a technologist, when I engage with companies or public administrations, I never start from technology.
I always start from the human being: from motivation to learn, to study, to organize collective work. Without that, no technology can truly help us.

Franco Molteni – I would like to conclude by looking toward the future. Are the artificial intelligences we currently have on the verge of a radical transformation? And finally, when will quantum computers come into play?

Massimo Chiriatti – Experts in quantum computing tell me that we will likely talk about it again in a few years. What we will see much sooner are major developments in humanoid robotics, which are also highly relevant for care and therapy.

I speak with people around the world who are developing joints, software, and hardware to assist humans. We need extensions that support care and assistance. I am optimistic: even in this case, machines will not be able to personalize their behavior the way a human can.

We no longer give machines only spatial commands using XYZ parameters. We ask them to observe reality and understand how to move in environments populated by human beings—doctors, nurses, patients. We are approaching what is known as Physical Artificial Intelligence.

An integration is underway between large language models, which can now communicate with us, and computer vision. We have already improved verbal interaction; through technologies such as automotive radar systems, spatial perception is also advancing. What remains is the physical dimension. We expect these tools to become more widespread and affordable, and I look forward to it—we truly need them.

Franco Molteni – Humanoid robots represent one of the frontiers of Rehabilitation Medicine, alongside wearable robots that already interact with patients. Given the current trajectory of healthcare systems, do you believe that organisations that fail to integrate artificial intelligence into healthcare management within the next three years risk becoming “second tier” compared to those that do?

Massimo Chiriatti – Increased productivity can lead to better resource allocation, shorter waiting times, and more personalized care, but this outcome is not automatic. It depends on how artificial intelligence is integrated into clinical processes, how decision flows are governed, and how well the relational dimension of care is preserved. Without conscious governance, efficiency gains may even produce outcomes opposite to those intended.

Franco Molteni – This conversation should prompt serious reflection: technological evolution is not a luxury. It risks becoming a dividing factor between those who have access to technology and those who do not. It is something vital, because the individual’s decision—to use or not use artificial intelligence—will be decisive.

We probably need to build an internal ethical structure that allows us to externalize important functions consciously, aware that they will have significant interactions with ourselves and with others. This is not a matter of efficiency, but of evolution.