Philosophical Foundations of Intelligence

Introduction

"Can machines think?" is the most fundamental philosophical question in artificial intelligence. This article explores the Turing test and its critiques, the Chinese Room argument, the symbol grounding problem, the hard problem of consciousness, and the philosophical distinction between strong and weak AI.

Related content: Embodied Cognition Theory

1. The Turing Test and Its Critiques

1.1 Turing's Question (1950)

In "Computing Machinery and Intelligence," Turing proposed: rather than asking the vague question "Can machines think?", we should substitute an operationalizable test -- the imitation game.

If a human judge cannot reliably distinguish the machine from a human in text-based conversation, the machine has demonstrated intelligence.

1.2 Turing's Responses to Objections

Turing anticipated and responded to numerous objections:

Objection	Content	Turing's Response
Theological	Thinking is a function of the soul	God could endow machines with souls
Head-in-the-sand	The consequences are too frightening	An emotional reaction, not a rational argument
Mathematical	Goedel's incompleteness limits machines	Humans also have limitations
Consciousness	Machines lack conscious experience	Solipsism applies equally to humans
Creativity	Machines cannot truly innovate	Creativity may just be complex computation

1.3 Criticisms of the Turing Test

Ned Block's China Brain argument: a system of a billion people, each simulating one neuron, could pass the Turing test but clearly has no understanding
ELIZA effect: simple pattern matching can deceive some people, suggesting the test standard is too low
Tests only language: ignores vision, motor skills, and other dimensions of intelligence
Cultural bias: unfair to test subjects from non-English cultures

2. The Chinese Room

2.1 Searle's Argument (1980)

Premise 1: Programs are purely syntactic (manipulate symbols)
Premise 2: Human minds possess semantics (understand meaning)
Premise 3: Syntax is insufficient to produce semantics
Conclusion: Programs cannot produce genuine understanding

Key distinction:

Strong AI: an appropriately programmed computer truly possesses cognitive states (understanding, beliefs, consciousness)
Weak AI: computers merely simulate intelligence without truly understanding

2.2 Major Rebuttals

Rebuttal	Argument	Searle's Response
Systems reply	The room as a whole understands Chinese	Let the person memorize all the rules; still no understanding
Robot reply	Adding perception and motor abilities would give understanding	Perceptual input is still just more symbol manipulation
Brain simulator reply	A brain-simulating program should have understanding	Simulating digestion cannot digest food
Other minds reply	We cannot confirm that other people understand either	But other people have similar causal architectures

2.3 Modern Significance

Large language models like ChatGPT have reignited the Chinese Room debate:

Do LLMs "understand" language?
Can statistical pattern matching produce semantic understanding?
If an LLM behaves as though it understands, is there a difference from true understanding?

3. The Symbol Grounding Problem

3.1 Harnad's Challenge (1990)

How can symbols in a symbol system acquire meaning? Symbols cannot be defined solely through other symbols (circular definition); they must be grounded in perceptual experience.

Traditional AI:
  "cat" → defined as "small furry four-legged animal"
  "animal" → defined as "living organism"
  ... (infinite loop of symbolic definitions)

Grounded AI:
  "cat" → visual patterns + tactile experiences + sound memories + ...
  Symbols are bound to perceptual experiences

3.2 Connection to Embodied Cognition

Embodied cognition holds that:

Intelligence is not just computation in the brain, but also depends on the body and environment
Conceptual understanding is grounded in sensorimotor experience
Abstract thought is built on embodied metaphors (e.g., "warm person," "heavy problem")

This suggests that true intelligence may require:

A physical body interacting with the environment
Multimodal perception (vision, touch, proprioception)
Active exploration and manipulation of the world

4. The Problem of Consciousness

4.1 The Hard Problem

Chalmers (1995) distinguished between the "easy problems" and the "hard problem" of consciousness:

Easy problems (explaining functions, solvable in principle):

How does attention focus?
What is the difference between wakefulness and sleep?
How is information integrated?

Hard problem (explaining subjective experience, extremely difficult):

Why do physical processes give rise to subjective experience (qualia)? Why does "seeing red" have a particular feeling?

4.2 Implications for AI

Position	View	Stance on AI Consciousness
Functionalism	Consciousness = functional organization	In principle, AI can be conscious
Biological naturalism	Consciousness requires specific biological substrate	AI cannot be conscious
Panpsychism	Consciousness is a fundamental property of matter	AI may have some form of consciousness
Integrated Information Theory (IIT)	Consciousness = integrated information \(\Phi\)	Depends on system architecture

5. Strong AI / Weak AI / AGI

5.1 Conceptual Distinctions

Concept	Definition	Current Status
Weak AI	A tool that simulates certain intelligent behaviors	Widely achieved
Strong AI	A system that truly possesses understanding and consciousness	Philosophically debated
AGI	Artificial General Intelligence, capable of any intellectual task	Not yet achieved
ASI	Artificial Superintelligence, exceeding humans in all intellectual dimensions	Theoretical discussion

5.2 The AGI Debate

Optimists:

Computationalism: intelligence is fundamentally computation; silicon and carbon are just substrate differences
Scaling hypothesis: sufficiently large models + sufficient data → AGI
Evolutionary analogy: natural selection produced general intelligence, and so can artificial methods

Pessimists/Cautious:

The consciousness problem is unsolved; computation may not be enough
Current AI lacks causal reasoning and common sense
Emergent capabilities may be illusory (is there true understanding?)
Safety and control concerns

6. How Philosophical Positions Influence AI Research

Philosophical Position	Implication for AI
Rationalism	Knowledge can be encoded a priori → Expert systems, logic programming
Empiricism	Knowledge comes from experience → Machine learning, data-driven approaches
Pragmatism	Focus on results rather than essence → Whether it "understands" doesn't matter, as long as it solves problems
Embodiment	Intelligence requires a body → Embodied AI, robotics
Phenomenology	Emphasizes first-person experience → Consciousness research

References

"Computing Machinery and Intelligence" - Alan Turing (1950)
"Minds, Brains, and Programs" - John Searle (1980)
"The Symbol Grounding Problem" - Stevan Harnad (1990)
"Facing Up to the Problem of Consciousness" - David Chalmers (1995)
"Philosophy of Artificial Intelligence" - Stanford Encyclopedia of Philosophy