Artificial intelligence (AI) is rapidly touching nearly every aspect of our modern life, from simple predictive models for text completion when composing messages and emails to highly sophisticated content creation, such as prompt-based realistic video generation. Yet despite AI’s widespread prevalence, there remains unawareness or misunderstanding about what AI truly is and how it operates. AI is often perceived as a “black box,” its workings and decision-making processes being not fully explainable. However, AI still represents an advanced form of mathematical modeling at its core. The AI model has been made powerful in recent years due to the vast computational power and large volumes of data. The black box perception of AI largely stems from the limitation in analyzing models of large size, trained with a large volume of data drawn from real-world (vs synthetic data with known generative processes that are easier to model and thus study).
A Common Misconception About AI
A common misconception about AI is the belief that it deeply understands the data it processes. However, whether the AI models truly understand information (ie, they have a world model in the data space where they operate) is debatable. The debate exists in part because defining what understanding means is unclear and is a deep philosophical question itself. However, what can be unequivocally said is that AI models have become superbly adept, even at a superhuman intelligence level, at identifying patterns within the data to produce outputs/decisions accordingly. The increasing capability of the AI models with larger model size and data-driven training has come at the cost of transparency and explainability. Early models of AI were symbolic AI, relying explicitly on human experts encoding knowledge in rule-based forms. Such modeling could align AI decision-making with human understanding.1 Modern data-driven modeling approaches lack such explicit transparency and make it challenging to intuitively grasp how decisions emerge from an AI model.1 This lack of understanding, while the AI applications grow exponentially in availability and ease-of-use to make people over-rely on AI systems, potentially makes one bypass their own better judgment; this could increase confusion and sow mistrust.
Before the AI-Based Modeling Prevalences: The Traditional Hypothesis-Driven Approach
In traditional mathematical and statistical practices, the standard model development approach typically begins with a well-defined hypothesis (a clear modeling relation) tested through experimentation. Researchers design experiments to evaluate the validity of their hypothesis, meticulously controlling what variables are measured and how confounders are controlled. The experiment outcomes confirm or disprove the hypothesis. For example, a researcher might hypothesize that increased daily exercise correlates with improved cardiovascular health. They would then systematically collect and analyze data statistically to test that precise hypothesis. This classical approach is inherently structured and constrained by design. Hypotheses explicitly direct the investigative process, including the data and confounders collected, and outcomes confirm or reject delineated ideas (
Machine Learning: An Exploratory Paradigm
Machine learning, the core technology behind recent advances in AI, departs from the traditional hypothesis-driven paradigm. Rather than imposing a rigid assumption or hypothesis, machine learning-based inferences/discoveries are mostly data-driven. The machine learning algorithms are usually provided with large datasets to find patterns relating inputs to desired outputs, but without explicit instructions or priors on what patterns are to be expected (the hypothesis). The algorithms “roam freely,” learning patterns from the data, independently discovering relationships and correlations that humans might never have anticipated or specifically sought out.1 For instance, a machine learning-based model analyzing consumer purchasing habits does not begin with a specific hypothesis such as, “Customers of a particular age group who buy diapers also buy beer.” Instead, given sales data, the model might uncover relationships that might not have been trivial for human analysts to consider.
Furthermore, machine learning modeling that powers modern AI, particularly deep learning with artificial neural networks (ANNs), is characterized by algorithms that can improve performance through continuous training on large datasets.1,2 Unlike symbolic AI, deep learning architecture becomes increasingly proficient through iterative processes as more data becomes available, dynamically adjusting its internal parameters without direct human intervention.2 In all forms of machine learning-based modeling, the model’s true capability is usually assessed with model validation in an unseen dataset. As increasingly large machine learning models are considered, it is easier to find several spurious patterns in data-driven search. However, only those patterns that are truly predictive in an unseen dataset (the model has never seen this dataset to find patterns being tested) are likely to provide value.
AI as Advanced Statistical Analysis
Despite the methodological differences between machine learning-based modeling powering modern AI and conventional approaches such as symbolic AI or hypothesis-driven investigations, it is crucial to recognize that machine learning, and therefore today’s AI, is fundamentally statistical in nature. Like traditional statistics, AI models operate by identifying patterns, making predictions, and quantifying uncertainty. Techniques such as affine transformations, perceptrons, decision trees, regression, prediction, and clustering have existed for decades within statistical practice. Modern AI primarily expands upon these foundational methods by leveraging advanced computational techniques, larger model size, larger datasets, and enhanced computing power.
Thus, rather than viewing AI as a mysterious “black box,” it is more accurate to consider it an advanced statistical toolbox. The novelty is not in the basic concept of pattern identification, but rather in the scale, speed, complexity, and freedom with which these patterns can now be uncovered (
Bridging Conventional Statistics and Machine Learning
While traditional statistical modeling and machine learning share a foundation in probability theory and data analysis, their philosophical orientation toward knowledge generation differs. Conventional statistics is primarily hypothesis-driven. It begins with a hypothesis and seeks to test relationships between variables through probabilistic modeling and estimation. In contrast, machine learning is prediction-driven, emphasizing model performance and pattern recognition over formal hypothesis testing. Yet, these approaches are not opposites but points along a continuum. Statistical inference provides interpretability and theoretical grounding, while machine learning offers enhanced capability, scalability, flexibility, and adaptability to complex, high-dimensional data. Conceptually, machine learning can be understood as computational statistics, an evolution of classical methods that leverages computational power to model patterns beyond human-specified constraints.
The Continuum From Supervised to Unsupervised Learning
This continuum between hypothesis-driven and exploratory approaches is clearly illustrated in the spectrum of learning paradigms used in today’s machine learning-based AI modeling, from supervised to unsupervised learning.3 In supervised learning, algorithms are trained on labeled data, mirroring the traditional scientific process in which hypotheses (labels) guide model construction and validation. Semi-supervised and reinforcement learning occupy a middle ground, combining elements of external guidance (active/passive labels) with adaptive, data-driven learning. At the far end, fully unsupervised learning operates with minimal or no predefined structure/guiding signal, identifying latent relationships and emergent groupings within data, effectively allowing the data to “speak for itself.” This spectrum underscores that modern AI does not replace statistical reasoning but extends it, enabling the discovery of complex, nonlinear, and multidimensional associations that traditional methods might not have the power to model.
Predictive and Generative AI: Examples of Statistical Pattern Recognition
The machine learning-based modern AI today generally can be categorized into two broad categories: predictive and generative. Both use statistical patterns, but they do so in distinct ways.
Predictive AI functions power many tools we commonly encounter, such as autocomplete on a smartphone keyboard or predictive search suggestions from Google. These AI systems analyze historical data to estimate or “predict” future outcomes. For example, a text prediction model learns common linguistic patterns from millions of users to anticipate the next word a user might type.
Generative AI takes statistical pattern recognition a step further. Instead of models that can merely predict future data points based on past patterns, generative AI models are capable of modeling the underlying pattern that generates data to infinitely create new content from the modeled pattern. This demonstrates a more sophisticated application of statistical pattern recognition. The capabilities of generative AI differ substantially from traditional AI in their potential to produce new outputs for value creation.3 For example, generative AI can create realistic images, text, music, or code. Platforms like ChatGPT, which generate coherent and contextually relevant text in a conversational setting, exemplify this approach. These systems statistically analyze vast quantities of text, learning the structure, context, and language style to model the underlying language patterns. The model can then synthesize new sentences or paragraphs based on the learned patterns (
AI in Health Care
As the AI modeling progression, the integration of AI into health care represents an exemplary convergence of human and artificial intelligence, enabling unprecedented precision and efficiency in medical decision-making and patient care. Machine learning-based AI applications are revolutionizing clinical workflows, from radiology and pathology to mental health, driven by data-rich environments and advanced computational capabilities.4,5
AI applications have particular promises in health care due to the complexity of diagnosis, the variability of symptom presentations, and the challenges inherent in traditional diagnostic methods. Machine learning algorithms can analyze extensive clinical datasets, including electronic health records and patient-reported outcomes, to identify subtle patterns and predictors of illnesses/recovery, thus enhancing early detection, precision in diagnosis, and individualized treatment planning.4
However, the integration of AI in health care raises critical considerations also. Ethical issues surrounding patient privacy, data security, algorithmic transparency, and interpretability must be rigorously addressed.5 The future of AI in health care will depend significantly on collaborative efforts between clinicians, technologists, ethicists, and policymakers to ensure that AI applications not only enhance clinical capabilities but also uphold ethical standards and equitable access to care (
Real-World Implications of AI Adoption
AI’s integration into our daily lives and society at large significantly impacts productivity and labor management. AI can substantially enhance workplace productivity by automating routine tasks, allowing human workers to focus on more strategic or creative responsibilities.6 This productivity shift underscores the practical significance of AI in optimizing work processes, while also raising concerns about employment displacement, data privacy, and environmental impacts.7,8
It is essential to critically examine these transformations, such as AI’s anticipated impacts on employment, social and political discourse, the nature of teaching and learning, etc, in the broader socio-political context that shapes AI deployment.9 These socio-economic considerations highlight that AI technologies do not exist in a vacuum but reflect and reinforce existing societal structures and tensions.
Concluding Thoughts
Machine learning-based AI modeling is increasingly becoming a central methodology across diverse fields, reflecting not just technological innovation brought by the approach but a broader transformation in scientific inquiry itself.1,10 The effective integration of resulting AI applications into society requires interdisciplinary collaboration, as well as careful consideration of their ethical implications, transparency, and interpretability.10 Identifying these dimensions will enable us to leverage AI’s capabilities fully, ensuring technological advancement and alignment with societal values.
Understanding AI as a fundamentally statistical model helps demystify its current perception as a fully black-box offering. AI models, though increasingly complex, where model interpretation and scrutiny techniques are catching up with the modeling innovations, are still an evolution of established statistical methods based on advances in computing power, increased data availability, and algorithmic innovation. Recognizing modern AI as an advanced statistical model helps clarify AI’s strengths, limitations, and potential uses, allowing us to apply these powerful tools thoughtfully, ethically, and effectively.
Ms Hakam is a medical student at Texas A&M’s EnMed program, where she conducts research in nanomedicine drug delivery systems and postpartum mental health, and is actively engaged in mental health advocacy.
Dr Lamichhane is a research scientist at Houston Methodist Academic Institute (HMAI) in the Department of Psychiatry and Behavioral Health, working to advance research in mental health. Before moving to HMAI, he was an assistant research professor at Rice University, working within the Digital Health Initiative
Dr Sabharwal is the Ernest D. Butcher Professor at Rice University’s Electrical and Computer Engineering department. His research interests are at the intersection of machine learning, behavioral sciences, and medicine.
Dr Moukaddam is a professor of psychiatry in the Department of Psychiatry at Baylor College of Medicine and the director of outpatient psychiatry at Harris Health System. She also serves on the Psychiatric Times Editorial Board.
References
1. How artificial intelligence works. European Parliament. 2019. Accessed January 7, 2026.
2. LeCun Y, Bengio Y, Hinton G.
3. Jindal JA, Lungren MP, Shah NH.
4. Topol EJ.
5. Rajpurkar P, Chen E, Banerjee O,Topol EJ.
6. Al Naqbi H, Bahroun Z, Ahmed V.
7. Maphosa V.
8. Resnick M. Generative AI and creative learning: concerns, opportunities, and choices. An MIT Exploration of Generative AI. March 2024. Accessed January 7, 2026.
9. Deranty JP, Corbin T.
10. Jordan MI, Mitchell TM.
