The Turing Test Philosophy of AI: The Turing Test
• Human (tester) communicates with a human and a machine via a typing interface • Conversation is totally unconstrained: Any subject, any duration, any language (including slang), lying allowed, etc. • Tester must determine which is the machine. If no better than chance, must grant that machine is intelligent (acc to Turing).
What’s the point?
A mini-Turing Test
• To cut through philosophical discussions of “what is intelligence?” and “can a machine think?” with a simple test. • A common misconception is that the Turing Test is too easy – in fact, very very difficult, and no program has ever passed it.
• Sam, as Alice (human) and Alice (chatbot) • Sam (human) will do all the typing (so that you can’t tell by typing speed) • Neither has to tell the truth! • Later, check out alicebot.org to play with Alice.
The Theological Objection
Objections to the Turing Test
• Thinking is a function of the soul • Machines don’t have a soul, therefore they can’t think
1
The “heads in the sand” objection • Thinking machines would be terrible, so let’s not think about it.
The mathematical objection Gödel: All complete axiomatic formulations of number theory include undecidable propositions. (e.g. “G is not a theorem of formal system X”) Lucas (and Penrose): Machines are formal systems, so there will be a formula the machine will be unable to produce as true, although a mind can see that it is true. And so the machine will not be an adequate model of the mind. "Lucas cannot consistently assert this formula". ?
The argument from consciousness
Argument from various disabilities
Two parts: • Self-awareness • Qualia: ‘really feeling’ some sensation or emotion. Turing’s response: • Can’t know about consciousness unless you are the thinker (even with other humans) • If there are any external manifestations, they will show up in the Turing Test – and if there aren’t, who cares? [paraphrased ☺]
• Turing lists a number: machines can’t… be kind… have a sense of humor… make mistakes…fall in love…use language…be creative… • Turing’s response is more or less that these are areas for research, but that he doesn’t see any particular reason why they can’t do these things.
Lady Lovelace’s Objection
Argument from continuity
• That is, machines can only do what they’re programmed to do. Turing responds that machines surprise him all the time… Also, what if they learn? Evolve? Change in such a way that their behavior is surprising even to their programmers?
• Nervous system is not a discrete state machine. • Turing argues that a discrete system can mimic a continuous system well enough to pass the TT. • Penrose returns to this argument – and so will we, next week.
2
Argument from Informality of behavior • Humans don’t strictly follow rules; computers do. Turing responds: • Humans *do* follow the laws of physics, at least, and probably higher level laws of behavior. • Machines can break ‘rules of conduct’ as easily as humans can.
Other objections • Ignores other kinds of intelligence (e.g. a dolphin couldn’t pass, nor a very clever Mars rover) • Overemphasizes linguistic fluency • Why should an intelligent computer pretend to be a human?
Argument from ESP • If humans have ESP and machines don’t, then the TT could distinguish. Turing gives up on this one!
Constrained Turing Tests • Not used to establish intelligence – but are used to support claims that program is ‘human-like’ in some specific way • Can be limited by time (e.g. < 5 min interaction), subject matter (e.g. must talk about sports), or medium (e.g. art, music, etc. • Often used to evaluate domain-specific AI apps. • Check out The Loebner Prize… (www.loebner.org)
Joke Analysis and Production Engine (JAPE)
A super-mini TT for humor
JAPE is a program that generates jokes from scratch (see “An implemented model of punning riddles” on my home page if you are interested). Look at the following jokes, and decide which ones are JAPE-generated, and which ones come a published joke book. Discuss online!
3
Spot the JAPE joke I • What’s the difference between money and a bottom? One you spare and bank, the other you bare and spank. • What do you give a hurt lemon? Lemon aid. • What do you call a sour assistant? A lemon aide. • What do you call Martian beer? An ale-ien.
Spot the JAPE joke II • What kind of pig can you ignore at a party? A wild bore. • What animal runs round the the forest making the other animals yawn? A wild bore. • What do you get when you cross jewelry and a bobcat? Cuff lynx. • What do you get when you cross the Atlantic with the Titanic? About halfway.
What do you think? • Is the Turing Test a reasonable way to establish whether or not a machine is intelligent? Why or why not? • What would you propose instead?
The Physical Symbol Systems (PSS) Hypothesis “A physical symbol system has the necessary and sufficient means for intelligent action.” – Newell & Simon • Belief in the PSSH is also referred to as “strong AI”.
The Physical Symbol System Hypothesis
What makes up a Physical Symbol System? • Symbols: physical patterns that can be part of expressions. • Expressions: instances of symbols related in some physical way. • Processes: act on expressions – creation, modification, reproduction and destruction. So, a Physical Symbol System is a “machine that produces through time an evolving collection of symbol structures.” [Newell & Simon] Must exist in a greater world of external objects.
4
The Turing Machine
The Turing Machine
• A read/write head that scans a (possibly infinite) tape divided into squares, inscribed with 0 or 1. The head scans a square, erases it, prints a 0 or 1, moves back or forward, and goes into a new state. • Behavior completely determined by:
• An abstract representation of a computing device • All Turing Machines are PSSs • All modern digital computers are Turing Machines So, the PSSH says that a digital computer is sufficient for intelligence (but not that digital computers are necessarily intelligent).
– the state the machine is in – the number on the square it is scanning, and – a finite table of instructions, specifying, for each state and binary input, what the machine should write, which direction it should move in, and which state it should go into.
What’s so special about a digital computer? • A digital computer is a Universal Turing Machine • So, given the right program, it can mimic the behavior of any discrete-state machine • So, we don’t need to worry about the physical instantiation of the computer, just its software.
Summary: the PSSH/Strong AI • “running the right sort of program necessarily results in a mind” [Russell and Norvig, p 959]
Weak vs Strong AI Weak AI: not committed to the PSSH… • Maybe physical instantiation matters? • Maybe sub-symbolic representations necessary? (e.g. neural networks?) • Maybe probabilistic/quantum effects necessary? …but still holds that AI is possible.
Searle’s Chinese Room
5
Objections to the PSSH: Searle’s Chinese Room
What is intentionality?
• Take a machine that passes the Chinese imitation game, and reimplement it as a giant library, with a non-Chinese-speaking human as the `processor’. (compare with Turing Machine) • The inputs and outputs would still pass the Turing Test – but where is the understanding? Who or what understands Chinese?
• A quality of mental states: their aboutness. • Intentional states include believing, knowing, desiring, fearing… • In particular, Searle claims that the Chinese Room thought experiment shows that machines cannot have the intentional state of “understanding”. • Conceptually linked to ideas about consciousness and qualia.
The systems reply
Searle’s response
• The human in the Chinese room is only one part of the system – it is the whole system (including the room itself) that does the understanding. • Compare with the human brain: Do neurons ‘understand’?
The robot reply • Add a camera, and manipulators, and relate the formal symbols to objects in the real world (grounding). Searle: • Still no intentional states – where would they be? • Robot response acknowledges need for more than formal symbol manipulation. What do you think?
• Let man memorize all the rules, and carry them out in his head. Now the whole system is held within his brain, but there is still no understanding. • If you ask the man in Chinese if he understands Chinese, he will say “yes” (in Chinese), but if you ask in English, he will say “no”…
The brain simulator reply • The program simulates the neurons in a Chinese speaker’s brain. Searle: • Replace the books with water pipes, and have the man open and close pipes. Man still doesn’t understand, nor do pipes.
6
The combination (grounding + brain simulator) reply
The other minds reply
Searle’s response here is interesting: It would be “rational and indeed irresistible to accept the hypothesis that the robot had intentionality, as long as we knew nothing more about it”. However, “…if we knew independently how to account for its behavior…we would not attribute intentionality to it, especially if we knew it had a formal program”. (1980, p421) If we came to understand how the human brain worked, in its entirety, would we still attribute intentionality?
• We don’t know how anyone actually thinks aside from behavioral tests, so those will have to do. Searle’s response: • Not about tests, but about what mental states really are • “…it couldn't be just computational processes and their output because the computational processes and their output can exist without the cognitive state.” • i.e. the zombie argument
Many mansions reply
Hofstadter’s reply
• If programming is not enough to produce cognitive states, then perhaps there are other ways Searle: • Strong AI says: "mental processes are computational processes over formally defined elements." • If it’s not Strong AI, then Chinese Room argument does not apply.
Where does intentionality come from, according to Searle? • Implementation of the right kind of programming in the right kind of stuff "specific biochemistry" (1980, p. 424) • Simulation of that biochemistry will not do – must have the real meat. • Not easily distinguished from simple dualism (mind and matter are different ‘substances’)
• Searle’s argument is based on the “intuitively obvious” lack of understanding in the Chinese Room. • If the system is a) in a body, with sensors and actuators, b) functioning in real time, and c) behaving as if it had intentional states, would those intuitions still hold?
Consciousness and intelligence • Is intelligence sufficient to produce consciousness? • Is consciousness necessary for intelligence? Or, rephrased, are “zombies” and “zimbos” (Dennett) possible?
7
Zombies and Zimbos • Zombie: Behaviorally indistinguishable from a human, but no consciousness. • Zimboes: A type of zombie, which has explicit representations of its own mental states, and rep’ns of those rep’ns, and so on up. Behaviorally indistinguishable from Zombies. Dennett (1998) mocks the idea that either are possible: “Zimboes thinkz they are conscious, thinkz they have qualia, thinkz they suffer pains--they are just "wrong" (according to this lamentable tradition), in ways that neither they nor we could ever discover!”
Philosophy vs. science What level of similarity is required to produce intelligent behavior? • Symbolic models of intelligent reasoning • Neural models of the brain • Biochemical models… • Quantum models… • Or other: evolutionary models, etc. Does intelligence really require a living human brain in a human body?
The COG project (MIT AI lab) • An attempt to build a robot (body, actuators, sensors), with a “neural network” type ‘brain’, that can learn in the real world. • Claim: Because it is grounded in its environment, its symbols have real meaning – so capable of real thought. • Implementation of the “combination” reply to Searle. • Doesn’t (yet?) work well enough to be evidence in either direction.
Bibliography • • • • • •
John Searle (1980). "Minds, Brains, and Programs." Behavioral and Brain Sciences, 417-424. Daniel Dennett (1998). Brainchildren, Essays on Designing Minds, MIT Press and Penguin. A. M. Turing (1950) Computing Machinery and Intelligence. Mind 49: 433-460. A. Newell and H. A. Simon, `Computer science as empirical enquiry: symbols and search', Communications of the ACM, 19(3), 113--126, (1976). Russell & Norvig.Artificial Intelligence: A Modern Approach (2nd edition). Prentice-Hall, 2002. There is an EXCELLENT bibliography of Turing/chatbot related material at http://cogsci.ucsd.edu/~asaygin/tt/ttest.html
8