Philosophical and Theoretical Approaches to AI: Four Classical Views, Major Schools and a Worked Example

An “approach to AI” may mean a criterion for intelligence, an engineering school, or a philosophical claim about machine understanding. The four classical approaches differ by whether they target reasoning or action and human likeness or rationality, as shown by a route-choice example.

Philosophical and theoretical approaches to AI: three different questions

A success criterion asks what counts as intelligent behaviour. An engineering school asks how a system produces it. A philosophical position asks what it implies about mind and understanding.

The Approaches to AI: Four Approaches and a Worked Example gives the full four-quadrant and Turing Test treatment. That quadrant answers only what counts as intelligence; engineering schools explain mechanisms, while philosophical positions examine machine understanding. A neural network can face an acting-humanly test or serve a rational agent, and a symbolic program does not imply strong AI.

Four classical approaches to AI: human thought, human action, rational thought and rational action

Russell and Norvig arrange four approaches as a 2 × 2: internal reasoning versus observable action, and human likeness versus rationality.

• Thinking humanly, or cognitive modelling: model how people think, using response times, errors or cognitive experiments. Psychological accuracy need not be optimal.

• Acting humanly: produce behaviour a human judge cannot reliably distinguish from human behaviour. Alan Turing’s Turing Test is the classic reference, but performance does not prove consciousness.

• Thinking rationally, or laws of thought: derive valid conclusions from explicit premises. Formal validity is transparent, but uncertainty, incomplete knowledge and computational cost trouble pure deduction.

• Acting rationally, or the rational-agent approach: choose the action expected to best achieve the objective with available information and constraints. Rational does not mean omniscient or always successful.

Theoretical schools of AI: symbolic, connectionist, probabilistic, evolutionary and embodied

These schools describe mechanisms rather than targets.

For a delivery robot, a symbolic controller applies if obstacle then detour; a connectionist model recognises camera data; a probabilistic planner uses P(corridor open)=0.70; and an embodied controller reacts to a sensed trolley. A hybrid agent connects them.

Symbolic AI uses formal representations and inference. Propositional and Predicate Logic: Truth Tables to Proofs teaches that foundation, but propositional logic is not all of symbolic AI.

Four AI approaches worked example: choosing a delivery route

A campus delivery robot must choose before entering either corridor. Route A takes 4 minutes when open with probability 0.70, and 10 minutes when blocked with probability 0.30. Route B always takes 6 minutes. In 10 human demonstrations, 3 people chose A and 7 chose B.

The novice model says, “Choose the guaranteed route when blockage probability is at least 0.25.” The rational agent minimises expected delivery time.

1. Acting humanly: imitate the majority. Since 7 of 10 people chose B, predict B.

2. Thinking humanly: apply the novice model’s threshold rule. Since 0.30 ≥ 0.25, predict B. This tests fidelity to the stipulated model, not a rule followed by every person.

3. Thinking rationally: use Open(A) → time(A)=4 < 6 → choose A and Blocked(A) → time(A)=10 > 6 → choose B. Neither condition is known categorically before entry, so deduction alone selects no route without an uncertainty rule.

4. Acting rationally: compute expected time:

• E[time(A)] = 0.70 × 4 + 0.30 × 10 = 2.8 + 3.0 = 5.8 minutes

• E[time(B)] = 1.00 × 6 = 6.0 minutes

• Since 5.8 < 6.0, choose A, with an expected advantage of 6.0 - 5.8 = 0.2 minute.

B fits imitation and the cognitive model. A fits the expected-time objective. Logic needs uncertainty machinery. A heavy penalty for delay variability could rationally change the action under a different utility function.

Philosophical debates in AI: simulation, understanding and machine minds

Weak AI says machines can perform apparently intelligent tasks. Strong AI makes the stronger claim that an appropriately built machine could possess genuine understanding or mind. Narrow versus general AI instead distinguishes capability breadth.

Functionalism defines mental states by their causal roles, so the same functional organisation might exist in different physical substrates. Biological-naturalist objections argue that computation or symbol manipulation alone may not be sufficient for understanding.

John Searle’s Chinese Room examines syntax without semantics. The symbol-grounding problem asks how symbols gain meaning beyond relations to other symbols. These arguments remain open, not settled proof. Passing the Turing Test settles neither question.

Philosophical approaches to AI in exam questions

Recurring questions match scenarios to approaches, distinguish Turing Test performance from cognitive modelling, separate symbolic from connectionist systems, or evaluate weak AI, strong AI, the Chinese Room and rational agents.

Use this self-check:

1. “Copy the 7/10 human majority.” Acting humanly.

2. “Choose A from 5.8 versus 6.0.” Acting rationally.

3. “Apply implications but lack the required fact.” Thinking rationally.

For exam revision, classify each stem by layer before choosing an answer: success criterion, engineering mechanism or philosophical claim.

Common mistakes when comparing AI approaches

For quadrant questions, ask whether the target is internal process or external action, then whether the standard is human likeness or rationality. Machine-mind claims are positions or arguments, not factual verdicts.

Philosophical and theoretical approaches to AI: the short version and next step

• Thinking humanly models human cognition.

• Acting humanly targets human-like behaviour.

• Thinking rationally applies valid inference.

• Acting rationally serves a stated objective.

• Symbolic, connectionist, probabilistic, evolutionary and embodied schools provide mechanisms, while hybrids combine them.

• Weak AI and strong AI make different claims about performance and genuine mind.

In the example, imitation and the cognitive model choose B, expected-time rationality chooses A, and categorical deduction is under-specified.

Move into application with AI & ML for Placements | Generative AI Placement Course or browse Skill Development Courses. For a broader Computer Science foundation, use Zero to Hero Complete CS Course.

Redraw the 2 × 2, recompute 5.8 versus 6.0, then classify three scenarios without looking at the labels.