Quantum machine learning (QML) has emerged as a defining area of intersection between quantum computing and artificial intelligence, yet significant confusion persists regarding its core definition. While the term gained prominence following the 2013 establishment of the Quantum Artificial Intelligence Lab by Google and NASA, its usage has become ambiguous, often encompassing everything from accelerated classical algorithms to hardware experiments that lack fundamental quantum properties. The central question remains: what specifically makes a machine learning model quantum? The answer lies not in speed, neural network architecture, or vague claims of quantum advantage. Fundamentally, QML is defined by the way information is represented, transformed, and measured. It requires the use of quantum mechanics rules rather than classical computation across three distinct dimensions. First, data representation differs from classical models. Instead of bits or floating-point numbers, QML utilizes quantum states, described as complex vectors or density matrices. These states exist in superposition, encoding information in complex-valued amplitudes. However, loading classical data into these states can be resource-intensive, and retrieving information is limited by the laws of quantum physics. Second, the models themselves operate through quantum evolution. Classical machine learning applies mathematical functions to data, whereas QML models apply unitary transformations, typically structured as parameterized quantum circuits. These circuits function by evolving the system's state over time, guided by a Hamiltonian. This creates a hypothesis space structurally different from classical counterparts, offering unique capabilities even if the training loop appears similar. Third, the measurement process is integral to learning in QML. Unlike classical systems where reading an output does not affect the model, quantum measurement is probabilistic and destructive. Outputs are determined by repeated executions, known as shots, to estimate results statistically. Consequently, the gradients used to update model parameters are derived from sampling noise rather than exact calculation, embedding uncertainty directly into the learning process. Many current applications labeled as QML do not meet these criteria. Approaches that merely run classical algorithms on quantum hardware or use quantum-inspired physics without true quantum substrates often fail to leverage the unique properties of quantum mechanics. A useful test is to determine if the quantum component can be replaced by a classical one without altering the model's mathematical structure. If so, the approach is not fundamentally quantum. Currently, the field is constrained by noisy, small-scale hardware. Most research is exploratory, focusing on learning theory, model classification, and understanding the impact of noise. Despite the lack of immediate speedups over classical methods, QML remains vital for rethinking the foundations of learning in a quantum context. It expands the definition of what learning can mean, preparing the scientific community for future fault-tolerant hardware. As quantum computers evolve, clarifying the boundaries between hype and reality will be essential for translating theoretical promise into practical technological advancement.