Cognition encompasses processes involved in acquiring, processing, storing, and applying information. Human cognition is a useful reference for interpreting complex behavior, but it does not define every possible form of intelligence or how it must be implemented. section 5.1.1
Thought extends beyond language. Spatial reasoning, visual imagination, motor planning, and emotional inference do not necessarily depend on words. Children can reason before acquiring speech, and experts may make accurate judgments they cannot easily articulate. Conversely, fluent speech does not by itself establish understanding. section 2.1.3
Artificial intelligence offers ways to investigate cognition as well as perform tasks. Symbolic reasoning treats thought as rules; probabilistic models explore reasoning under uncertainty; learning-based systems investigate what can be acquired from experience. These approaches reflect different hypotheses about thinking rather than a settled account of it. section 1.6.1
Observable behavior and underlying mechanism must be distinguished. Humans and artificial neural networks can produce similar outcomes through different processes, so coherent responses or successful task performance alone do not settle whether a system understands. At the same time, opaque internal processes make that judgment difficult in practice. section 5.4.6 section 1.5.5
Table 5.1 Key structural differences between human cognition and artificial neural networks. Similar behaviors can emerge in both systems, but they arise from fundamentally different mechanisms governing goals, learning, memory, time, and representation.
|
Dimension
|
Human cognition
|
Artificial neural networks
|
|
Goal formation
|
Arises from internal processes shaped
by needs, intentions, and external context
|
Externally specified as optimization
targets during training
|
|
Learning signal
|
Consequences of actions and
interactions with the environment
|
Numerical loss minimized during
training
|
|
Memory
|
Continuously reshaped through use and
recall
|
Parameters remain fixed between
training runs
|
|
Temporal organization
|
Learning and action unfold together
over time
|
Training and inference generally occur
in separate phases
|
|
Representation
|
Concepts refer to objects and events
in the world
|
Representations capture statistical
structure in data
|
|
Adaptation
|
Development accumulates along an
irreversible history
|
Models can be retrained, reset, or
copied
|
Compare human cognition and artificial neural networks across goals, learning, memory, time, and representation. section 5.4.6
Brain models serve different purposes. A model designed to explain biological behavior is not necessarily the best engineering system, while a model built for efficiency may omit much of the brain’s complexity. Biological fidelity, explanatory understanding, and task capability should not be treated as interchangeable achievements. section 5.3.1
Current AI systems show reasoning, adaptation, and problem solving, but their abilities remain uneven and depend heavily on training data, prompts, tools, and surrounding systems. They lack features associated with human cognition, including embodied experience and a continuous lived history. Their performance is therefore neither proof of human-like thought nor easily dismissed as empty imitation. section 9.5.6 section 2.1.3



