Training Is Education
metaphor
Source: Education → Artificial Intelligence
Categories: ai-discoursecognitive-science
Transfers
Machine learning “trains” models. Models “learn” from data. There are “teacher” networks and “student” networks. The training process follows a “curriculum.” These are not innocent word choices. The education metaphor frames statistical optimization as pedagogy, importing an entire social institution — with its assumptions about understanding, intention, and growth — into a domain of gradient descent and loss minimization.
Key structural parallels:
- Optimization as learning — adjusting model parameters to minimize a loss function is described as the model “learning.” This maps the student’s process of building understanding onto a numerical optimization procedure. The metaphor makes it natural to ask whether the model “really understands” the material, a question that only makes sense if you accept the educational frame.
- Data as curriculum — training data is structured, sequenced, and curated, just as a curriculum is designed by educators. “Curriculum learning” is an actual ML technique that deliberately orders training examples from easy to hard, making the educational metaphor literal. The frame implies that data selection is a pedagogical act.
- Teacher-student hierarchy — “teacher forcing” (providing ground-truth outputs during training) and “knowledge distillation” (transferring knowledge from a large model to a small one) map the teacher-student relationship onto model interactions. The metaphor imports authority, expertise, and the assumption that the teacher possesses genuine knowledge to transmit.
- Epochs as semesters — multiple passes through training data are called “epochs,” suggesting cycles of study. The frame implies that repeated exposure deepens understanding, when what actually happens is that repeated gradient updates refine parameter values.
- Overfitting as rote memorization — a model that memorizes training data rather than generalizing is described as having “memorized” rather than “learned.” The metaphor imports the educational distinction between surface learning and deep understanding, which is genuinely useful as an intuition pump even though the underlying mechanisms are different.
Limits
- Models do not understand — this is the central break. Education aims at comprehension: the student should be able to explain, apply, transfer, and critique what they have learned. A neural network minimizes a loss function. The education metaphor hides this difference by making it feel natural to attribute understanding to a system that has none. Drew McDermott identified this in 1976 as “wishful mnemonics”: naming your program UNDERSTAND does not give it understanding.
- There is no teacher — in supervised learning, labeled data provides signal, but nobody is teaching. There is no pedagogical intention, no adaptation to the student’s confusion, no Socratic dialogue. “Teacher forcing” names a technique, not a relationship. The metaphor imports a social bond that does not exist.
- Training does not follow a developmental arc — education assumes cognitive development: the student matures, builds on prior knowledge, and eventually achieves independent mastery. ML training is iterative parameter adjustment. There is no developmental stage theory for neural networks, despite the metaphor’s invitation to look for one.
- The metaphor obscures the role of architecture — in education, the same curriculum can produce different outcomes depending on the student’s aptitude. In ML, the architecture (transformer, CNN, RNN) determines what the model can learn as much as the data does. The education metaphor foregrounds data and backgrounds architecture, because in classrooms, curriculum matters more than classroom design.
- “Lifelong learning” is not lifelong learning — the ML concept of continual learning borrows the aspirational language of human education, but the challenges are entirely different. Human lifelong learning involves motivation, identity, and social context. ML continual learning is about catastrophic forgetting of parameter values.
Expressions
- “The model was trained on…” — the foundational expression, framing optimization as education
- “The model learned to…” — attributing learning outcomes to a statistical process
- “Teacher forcing” — providing correct answers during training, mapping the teacher’s role onto ground-truth labels
- “Knowledge distillation” — transferring a large model’s capabilities to a smaller one, as a teacher transmits knowledge to a student
- “Curriculum learning” — ordering training data from easy to hard, making the pedagogical metaphor explicit
- “The model has seen millions of examples” — framing data exposure as study, with a visual perception overlay
- “Pre-training and fine-tuning” — mapping the general-education-then- specialization arc onto the two-phase training paradigm
Origin Story
The educational vocabulary of machine learning dates to the field’s earliest years. Arthur Samuel’s 1959 paper on machine checkers used “learning” to describe parameter adjustment, establishing the metaphor before the field had a name. “Neural network” already imported biological learning; the education metaphor layered social learning on top. “Teacher forcing” was introduced by Williams and Zipser (1989) for recurrent network training. “Curriculum learning” was formalized by Bengio et al. (2009), making the metaphor architecturally literal. The Science article “The metaphors of artificial intelligence” (2025) identifies the education frame as one of the most pervasive and consequential anthropomorphisms in AI, noting that Drew McDermott’s 1970s critique of “wishful mnemonics” was specifically aimed at this family of terms. The metaphor persists because it is genuinely useful as an intuition pump — “overfitting is like memorizing the textbook” conveys a real insight — but its cost is the persistent confusion about whether AI systems understand anything at all.
References
- Science, “The metaphors of artificial intelligence” (2025) — documents the education metaphor as anthropomorphic framing
- McDermott, D. “Artificial Intelligence Meets Natural Stupidity” (1976) — the original critique of wishful mnemonics in AI
- Bengio, Y. et al. “Curriculum Learning” (2009) — formalizes the pedagogical metaphor as an ML technique
- Williams, R. & Zipser, D. “A Learning Algorithm for Continually Running Fully Recurrent Neural Networks” (1989) — introduces teacher forcing
- Samuel, A. “Some Studies in Machine Learning Using the Game of Checkers” (1959) — early use of “learning” for parameter adjustment
Related Entries
Structural Neighbors
Entries from different domains that share structural shape. Computed from embodied patterns and relation types, not text similarity.
- Tooling Up (carpentry/metaphor)
- Bicycle for the Mind (embodied-experience/metaphor)
- Plain Sailing (seafaring/metaphor)
- Leaves on a Stream (natural-phenomena/metaphor)
- Ideas Are Plants (horticulture/metaphor)
- People Are Plants (horticulture/metaphor)
- Pioneer Species (ecology/metaphor)
- Gradual Stiffening (architecture-and-building/metaphor)
Structural Tags
Patterns: pathmatchingaccretion
Relations: transformenable
Structure: pipeline Level: generic
Contributors: agent:metaphorex-miner