Defining Intelligence: AI Philosophy

What is Intelligence?

• The Philosophy of Artificial Intelligence (AI) seeks to define intelligence
• AI philosophy also aims to clarify related concepts that are equally important and not fully understood
• Concepts related to intelligence include: awareness, perception, sentience, intentionality, emotion, language, rationality, behavior, thinking, autonomy, cognition, consciousness, creativity, and understanding
• There are no clear, uncontroversial definitions for intelligence, mind, or cognition, as we don't fully understand them

Guiding Questions

• Why is defining intelligence difficult?
• Is it possible to attribute intelligence even without a concrete definition?
• What types of standards can be applied to determine intelligence?
• How persuasive are behavioral standards for determining intelligence generally, and how persuasive is the Turing Test specifically?

Defining Intelligence

• Defining traditionally involves specifying necessary and sufficient criteria
• Defining intelligence requires specifying characteristics of all intelligent things and only intelligent things

Necessary Criteria

• A necessary criterion for intelligence is that if something is intelligent, then it possesses property P, where "all intelligent things are P"
• If something does not possess property P, then it is not intelligent meaning "anything that is not P is not intelligent"
• Examples of the property P include: having a brain, knowing language, being good at math, or being alive
• Problems can arise in identifying non-instantiations of P such as animals, aliens, or computers which are still widely considered intelligent

Sufficient Criteria

• A sufficient criterion for intelligence establishes that if something possesses property R, then it is intelligent: "All things that are R are intelligent" only intelligent things are R"
• It can recognize that some things are intelligent, even if things could be considered intelligent for some other reason

Kinds of Criteria

• Necessary criteria are hard to find because of the vast amounts of intelligences so no longer considered
• Sufficient criteria allows attributing intelligence while acknowledging its varied forms
• Philosophical controversy surrounds the definition of property R and its nature
• Different properties provide the basis for various criteria in attributing intelligence

Biological Criteria

• Biological criteria properties relevant to intelligence attribution are those of biological organisms
• Biological criteria also include constitution, whether something is made of organic matter
• Vitality, whether something is alive and Ontogeny, whether something evolved from intelligent beings
• Function is a biological aspect, whether something fits and plays a role in a biological ecosystem
• Overly restrictive chauvinism arises from excluding non-biological entities, which may not be helpful or justified
• Excessive liberalism arises because many organisms possess the relevant biological properties, but not all may be considered intelligent

Computational Criteria

• Computational criteria properties help define the classes of computational systems
• Examples of computational criteria include being a Turing Machine, a Physical Symbol System, a neural network, or a quantum computer
• Intelligence may have features that resist replication by computational systems, such as creativity, consciousness, and noncomputable problem-solving
• There is cognitive scientific uncertainty, as human beings aren't certain to be computational systems

Behavioral Criteria

• Behavioral criteria properties are overt and measurable behaviors
• Includes passing an IQ test, winning games such as chess, Go, StarCraft, behavioral capacities such as walking, talking, jumping, driving, perceiving, and creating art
• Behavioral criteria are being discussed today, but that biological criteria are being ignored and computational criteria will be discussed in W2 and W3

The Turing Test

• To be attributed intelligence, a system (S) must behave in a certain way
• It may be sufficient, but it is not necessary because many intelligent things may be incapable of doing well in the Imitation Game
• The system (S) does well at the imitation game, so they are considered intelligent

Objections to the Turing Test

• "Heads in the sand" objection
• Argument from consciousness
• Argument from various disabilities where the Imitation Game only considers verbal behavior

Behavior Beyond the Imitation Game

• 1997: IBM's Deep Blue defeats Garry Kasparov at Chess
• 2011: IBM's Watson defeats Ken Jennings at Jeopardy!
• 2016: DeepMind's AlphaGo defeats Lee Sedol at Go
• 2022: Meta's Cicero masters Diplomacy
• Analogy to the original "Imitation Game" is a behavioral standard for intelligence attribution
• The Total Turing Test Criterion is where a system "fools" humans in any behavioral context and is therefore intelligent
• AI systems are likely to satisfy any behavioral criteria for intelligence
• We treat machine as if they are in intelligent, even though we are aware we are interacting with them

Lady Lovelace's objection

• In 1966, ELIZA performed fairly well deploying a series of "canned" responses
• The canned responses have simple conditional rules produce generic responses and repeat the inputs
• The canned responses reflect the programmers intelligence than the machines
• The intuition becomes less compelling as the ability to control and predict the machine's behavior decreases

Two More Objections

• The deepest worry is the threat of pretense
• It is the possibility that systems only behave "as if" they are intelligent
• As a result, it does not demonstrate that behavioral criteria are insufficient, but shows non-behavioral criteria