lecture02.pdf

Lecture 2 The Turing Test and Searle’s Chinese Room Rob Gaizauskas COM1005 2023-24 Lecture Outline • How can we decide if a machine is intelligent? – Turing Test – Arguments Against the Turing Test – Other tests for Intelligence • Is AI Possible? – Searle’s Chinese Room Argument – Arguments Agai...

Lecture 2 The Turing Test and Searle’s Chinese Room Rob Gaizauskas COM1005 2023-24 Lecture Outline • How can we decide if a machine is intelligent? – Turing Test – Arguments Against the Turing Test – Other tests for Intelligence • Is AI Possible? – Searle’s Chinese Room Argument – Arguments Against Searle COM1005 2023-24 How can we decide if a machine is intelligent? • Suppose someone presents us with a machine they claim is a ``thinking” machine or is ``intelligent”. • What test(s) should we use to determine if this is true? • The most famous test is the Turing Test, named after the British scientist Alan Turing, who proposed it in 1950. – Only 4 electronic computers in existence at the time – Before Dartmouth conference (1956) and the `birth’ of AI COM1005 2023-24 Tests of Intelligence Turing Test • In his paper 1950 “Computing Machinery and Intelligence”, Turing proposed to replace the question “Can machines think?” with a question framed in terms of the imitation game • In the imitation game there are three players: – A man (A) – A woman (B) – An interrogator of either gender (C) • C, the interrogator, is in a separate room and communicates with A and B, who he knows as X and Y, only by asking written (typed) questions or via an intermediary • Object of the game – For C is to determine which of X and Y is A and B, i.e. which is the man and which the woman – For A is to mislead C into making the wrong identification, i.e. convince C he is the woman – For B is to help C make theCOM1005 correct identification 2023-24 Tests of Intelligence Turing Test • Turing’s twist is to suggest A (the one trying to deceive) be replaced by a computer • The question “Can machines think” is now replaced by the question: “Will the interrogator decide wrongly as often when the game is played like this as he does when the game is played between a man and a woman ?” (Turing , 1950) COM1005 2023-24 From: www.alanturing.net ©Copyright B.J. Copeland, July 2000 Tests of Intelligence Turing’s Prediction • Turing did not specify strict criteria for “passing” what has come to be called the “Turing Test” • He did say: “I believe that in about fifty years’ time it will be possible to programme computers, with a storage capacity of about 109, to make them play the imitation game so well that an average interrogator will not have more than 70 per cent chance of making the right identification after five minutes of questioning.” • It’s now 2023 – 73 years after Turing made this prediction. • Are we there yet? COM1005 2023-24 Tests of Intelligence “Eugene Goostman” • Chatbot developed in St Petersburg by 3 Russian/Ukrainian programmers • Pretends to be a 13 year old Ukrainian boy with English as a 2nd language – "not too old to know everything and not too young to know nothing” • On June 7, 2014 (60th anniversary of Turing’s death) Goostman was claimed to have passed the Turing test – Event at Royal Society, London, organized by Kevin Warwick of Reading University – 1/3 of (30) judges fooled after 5 minute conversation • Reaction from AI community mostly negative, arguing – Chatbot used its age/language limitations to trick judges – Turing’s “threshold” meant as a prediction about where research would be in 50 years, not as a sufficient condition for deeming a computer intelligent – Devalues “serious” AI research – nothing new in the approach COM1005 2023-24 Tests of Intelligence The Loebner Competition • Loebner competition ran annually from 1991 to 2019 – now defunct – Funded by Hugh Loebner (d. 2016) an American inventor and social activist – $100,000 prize for first computer to fool the judges – Annual prize ($4000 + bronze medal) for the most convincing (=“least bad”) computer • In 1997 a team from the NLP group in COM, University of Sheffield won (funded by chess player/businessman David Levy) this annual prize – From 2014-2019 the competition was incorporated into an AISB (British AI Society) Exhibition event – since stopped due to lack of funding – No computer was ever able to fool the judges and win the main prize • More details at: https://aisb.org.uk/category/loebner-prize/ and https://en.wikipedia.org/wiki/Loebner_Prize COM1005 2023-24 Tests of Intelligence Daniel Dennett on the Turing Test and the Loebner Competition • Entertaining Interview with Daniel Dennett on the Turing Test • Dennett is a well-known American philosopher and cognitive scientist, who has written extensively on issues around our ideas of mind, brain and consciousness. – See https://en.wikipedia.org/wiki/Daniel_Dennett • Makes clear Turing was not proposing his “test” as a “platform for serious scientific research” but a “conversation stopper” / “thought experiment” COM1005 2023-24 Tests of Intelligence But What about ChatGPT? COM1005 2023-24 Tests of Intelligence But What about ChatGPT? • In late May, 2023 a team of scientists at Israeli-based AI21 Labs published a study called: Human Or Not? A Gamified Approach To The Turing Test – Set up an on-line game where participants are randomly matched with a partner who may be a computer or a human – 1.5 million participants within 1 month – Engage in conversation for two minutes then guess: human or not? – Created multiple bots using various large language models (LLMs) – Bots prompted with a back story to give them an identity + instructions about the game + news about recent events • Results: – 10 million guesses in 1st month • Remember Turing’s prediction … Probability of Correct Guess Overall 68% When partner is a bot 60% WhenCOM1005 partner2023-24 is a human 73% Lecture Outline • How can we decide if a machine is intelligent? – Turing Test – Arguments Against the Turing Test • Objections anticipated by Turing • Criticisms raised by Copeland – Other tests for Intelligence • Is AI Possible? – Searle’s Chinese Room Argument – Arguments Against Searle COM1005 2023-24 Tests of Intelligence Objections Anticipated by Turing In his 1950 paper Turing anticipated 9 objections to his position: 1. 2. 3. 4. 5. 6. 7. 8. 9. The theological objection * The ‘heads in the sand’ objection* The mathematical objection The argument from consciousness * Arguments from various disabilities Lady Lovelace’s objection * Argument from continuity in the nervous system The argument from informality of behaviour The argument from extra-sensory perception NB: Turing’ paper is mandatory reading. You may skip sections 4 and 5. COM1005 2023-24 Tests of Intelligence The theological objection “Thinking is a function of man’s immortal soul. God has given an immortal soul to every man and woman, but not to any other animal or to machines. Hence no animal or machine can think” • Turing’s dismissal: Why not believe that God could give a soul to a machine if he wished? • He adds: “I should find the argument more convincing if animals were classed with men, for there is a greater difference, to my mind, between the typical animate and the inanimate than there is between man and the other animals.” COM1005 2023-24 Tests of Intelligence The heads in the sand objection “The consequences of machines thinking would be too dreadful. Let us hope and believe that they cannot do so.” • Turing’s dismissal: – “I do not think that this argument is sufficiently substantial to require refutation. Consolation would be more appropriate” – objection related to theological objection in that people like to believe they are somehow superior to the rest of creation COM1005 2023-24 Tests of Intelligence Argument from consciousness “This argument is very well expressed in Professor Jefferson's Lister Oration for 1949, from which I quote. ‘Not until a machine can write a sonnet or compose a concerto because of thoughts and emotions felt, and not by the chance fall of symbols, could we agree that machine equals brain – that is not only write it but know that it had written it. No mechanism could feel (and not merely artificially signal, an easy contrivance) pleasure at its successes, grief when its valves fuse, be warmed by flattery, be made miserable by its mistakes, be charmed by sex, be angry or depressed when it cannot get what it wants’” COM1005 2023-24 Tests of Intelligence Turing’s dismissal of argument from consciousness • How could we tell that a machine that had passed the Turing test was not conscious? • The argument seems to suggest the only way we could be sure that a machine thinks is to be that machine and to feel oneself thinking. • But, by the same argument, the only way to be sure that someone else thinks is to be that person. • How do we know that anyone else is conscious/thinks? – to doubt that they do leads to philosophical position known as solipsism. • But we assume other people are conscious. Similarly, we could assume that a machine that passes the Turing test is effectively conscious. COM1005 2023-24 Tests of Intelligence Lady Lovelace’s objection • Charles Babbage (1792-1871) – Specified a general purpose calculator called the Analytical Engine – Was in fact Turing-complete (i.e. could compute every function a general purpose modern computer could) – entirely mechanical, and never actually built. • In her 1842 memoir of Babbage, Lady Lovelace wrote: “The Analytical Engine has no pretensions to originate anything. It can do whatever we know how to order it to perform” (her emphasis) • Turing suggests variants of this objection – a machine can ‘never do anything really new' – a machine can never ‘take us by surprise’ • A computer cannot be creative, it cannot originate anything COM1005 2023-24 Tests of Intelligence Turing’s dismissal of Lady Lovelace’s objection • Points out that all creative acts could be seen as following from “seeds” planted in us by teaching, i.e., that humans never really do anything new either • Further, computers can and do surprise their programmers – producing answers that were not expected. – Data may have originally been given to the computer, but then it may have been able to work out its consequences and implications – Consequences may include much not foreseen by the programmer COM1005 2023-24 Lecture Outline • How can we decide if a machine is intelligent? – Turing Test – Arguments Against the Turing Test • Objections anticipated by Turing • Criticisms raised by Copeland – Other tests for Intelligence • Is AI Possible? – Searle’s Chinese Room Argument – Arguments Against Searle COM1005 2023-24 Tests of Intelligence Criticisms of Turing’s Position Copeland (1993) raises 4 objections to the Turing Test 1. The chimpanzee objection: Chimps can think but could not pass the Turing Test (same for dolphins, dogs, pre-linguistic humans). Therefore, test is too strong. Follows that the Turing Test cannot replace (i.e., exactly substitute for) the question “Can a computer think?”) But, while not a necessary condition for thinking, could still be sufficient – – – • i.e. , “if X passes the test then X thinks” might be true even if “if X thinks then X must be able to pass the test” may not COM1005 2023-24 Tests of Intelligence Criticisms of Turing’s Position 2. The sense organs objection: – Turing Test only focusses on computer’s ability to use words – no test of computer’s ability to relate words to things in the world. – A computer that passes the TT but cannot supply the word “cup” when presented with a cup cannot be said to understand the words it uses. – Therefore, test too weak and should be strengthened by insisting computer be given artificial sense organs. – However • can probe a speaker’s understanding of many words without investigating their sensory interaction with environment • Nothing precludes computers with sense organs entering the competition • Turing noted that giving a computer sense organs and then ‘subjecting’ it to ``an appropriate course of education” might be the best of way to prepare it for the test COM1005 2023-24 Tests of Intelligence Criticisms of Turing’s Position 3. The simulation objection: “A simulated X is not an X”, e.g., simulated diamonds are not diamonds A computer that passes the Turing test has shown that is a good simulation of a thinking thing – but that is not the same as being a thinking thing – – • In the imitation game if a man convinces the interrogator he is a woman, i.e., simulates a woman, doesn’t follow that he is a woman COM1005 2023-24 Tests of Intelligence Criticisms of Turing’s Position 3. The simulation objection (cont): But note – • • Some simulated X’s are X’s (e.g. a simulated voice is still a voice) Need to distinguish 2 types of simulation – – – – Simulation1 lacks essential features of what is simulated, e.g. simulated death, leather Simulation2 is just like what is simulated but is produced in a different way, e.g. artificial coal, simulated voice Objection only holds if computer simulation of thinking is always a simulation1 and never good enough to be a simulation2 But question is whether a computer simulation of thinking could be a simulation2 – simulation objection prejudges this COM1005 2023-24 Tests of Intelligence Criticisms of Turing’s Position 4. The black box objection: – Turing test treats the computer as a black box and looks only at output. A weakness? – Turing thought this was fine – treat our fellow human beings in the same way – Copeland thinks not: our judgement that our fellows think based not only on observation of their behaviour (outputs) but on the fact of our biological similarity COM1005 2023-24 Tests of Intelligence Criticisms of Turing’s Position 4. The black box objection (cont): – Suggests modifying Turing Test to insist • • it pass the interrogation test (outward criterion) it pass a test whereby we examine the program to see how output is produced and be convinced it is not simply a gigantic lookup table/pattern matching chatbot (design criterion) – To meet design criterion program would have to either 1. 2. Work like a human (too anthropomorphic?) Be modular/extensible so it could, e.g. be added to a robot with motor and sensory systems COM1005 2023-24 Tests of Intelligence Criticisms of Turing’s Position • Copeland goes on to suggest maybe Turing Test approach should be abandoned – Simply don’t know how to assess the application of concepts like “thinking” to radically new sorts of entities like programmable computers – Suggests key features of creatures we are happy to apply the term “thinking” to are those whose action-directing inner processes are massively adaptable • E.g. can form plans, analyse situations, deliberate, reason, exploit analogies, revise beliefs in the light of experience, etc. in the real world – For such creatures it is reasonable to explain or predict their behaviour in terms of what the agent thinks/believes/wants, i.e. using what philosophers call intentional language – Thus, if we can build robots to which we are happy to apply the framework of intentional explanation/prediction, why not say they can think? COM1005 2023-24 Lecture Outline • How can we decide if a machine is intelligent? – Turing Test – Arguments Against the Turing Test – Other tests for Intelligence • Is AI Possible? – Searle’s Chinese Room Argument – Arguments Against Searle COM1005 2023-24 Other tests for Intelligence • Researchers have begun using tests designed for humans – high school subject exams, bar exams, medical exams, reading comprehension, mathematics and coding tests, uni entrance exams, … – GPT4 does very well on most of them • However, LLMs do less well on reasoning tasks, such as a new kind and white discs on a 8 AN ABSTRACT-THINKING THAT DEFEATS MACHINES of logic test for AI systemsTEST created by François Chollet, called the LLM examine whether surface statistics of Abstraction and Reasoning Corpus (ARC) text, or if they migh Artificial-intelligence systems have so far been unable to achieve human-level performance on the ConceptARC test. This logic puzzle asks solvers to show how grid patterns will change after the solver has seen multiple demonstrations of an underlying abstract concept. Here are two sample tasks based on the concept of ‘sameness’ — between shapes in Task A and between orientations in Task B. See go.nature.com/43v6fzk for the answers. representations of th – Challenge: detect patterns in visual examples + apply them to new examples When they trained TASK A Demonstrations: Test: ? ? COM1005 2023-24 ? of moves made by play at spitting out accura legal moves. The rese had evidence that the of the state of the boa this representation t than just coming up w Bowman acknowle capabilities of LLMs and more limited tha that they are there, an model size, which ind Other tests for Intelligence • Defenders of LLMs say they have not been trained with ARC-like examples and will improve at ARC tasks if they are • Most researchers agree: “the best way to test LLMs for abstract reasoning abilities and other signs of intelligence remains an open, unsolved problem” – See ChatGPT broke the Turing test — the race is on for new ways to assess AI, Nature, July 2023. • And, of course, intelligence may not be the only trait we want to assess in AI’s … – Moral decision making – Empathy (cf. Voigt-Kampff Test) – … COM1005 2023-24 Lecture Outline • How can we decide if a machine is intelligent? – Turing Test – Arguments Against the Turing Test – Other tests for Intelligence • Is AI Possible? – Searle’s Chinese Room Argument – Arguments Against Searle COM1005 2023-24 Is AI Possible? • Depends on what we mean by AI … – Technology demonstrating some intelligent capabilities at human or super-human level already available • The question of whether the Strong AI hypothesis is true has led to extensive debate in the philosophical community. • Central to this debate has been John Searle’s Chinese Room thought experiment which Searle claims demonstrates the Strong AI hypothesis cannot be true … COM1005 2023-24 Is AI Possible? Searle’s Chinese Room Argument • Searle asks us to imagine: – A man inside a closed room with two slots in the walls – one marked input, the other output – The man is seated at a table with a large rulebook and lots of blank paper and pencils – Through the input slot someone pushes questions written on paper in Chinese (we assume the man inside does not know Chinese) – Following instructions in the rule book the man • looks up the characters on the input slip • performs operations on them (e.g. translating them to binary strings) • finally arrives at a new Chinese string which he writes on a slip of paper and pushes through the output slot COM1005 2023-24 Is AI Possible? Searle’s Chinese Room Argument From: http://www.alexandria.nu/ai/blog/attachments/00000050_ChineseRoom.jpg • Note: This scenario is NOT about translating between Chinese and some other human language (i.e. machine translation) – It’s about taking action based on processing a string of symbols in a language one does not understand COM1005 2023-24 Is AI Possible? Searle’s Chinese Room Argument From: http://www.alexandria.nu/ai/blog/attachments/00000050_ChineseRoom.jpg • Suppose the Chinese room performs so well it passes the Turing Test • Clearly the man inside does not understand Chinese • But a computer answering questions does nothing more than the man in the room would do • Therefore, a computer passing the Turing Test cannot be said to be thinking/understanding and thus the Strong AI hypothesis is false COM1005 2023-24 Is AI Possible? Searle’s Chinese Room Argument • Note Searle claims the question (whether computers can think) is NOT an empirical question – I.e. is not one that can be settled by experimentation • So, no future Turing Test has any relevance for the question “Can machines think?” – He claims his Chinese room argument settles the argument about whether computers could ever genuinely think/understand • Essentially, no amount of syntactic manipulation can lead to semantics (meaning) COM1005 2023-24 Is AI Possible? Searle’s Chinese Room Argument • Is Searle’s argument valid? • Many people have argued that it is not – See, e.g. Copeland (1993) COM1005 2023-24 Is AI Possible? Copeland’s Rebuttal • Copeland begins by noting it’s wrong to infer that because A: the man inside the room (call him Joe) does not understand Chinese therefore B: the Chinese room (slots, rule books, paper, man) does not understand Chinese) • I.e. the system as a whole might understand Chinese even while Joe – the man inside – does not – Compare: does any part of my brain “understand” English? Or only “me” taken as an integrated, biological system COM1005 2023-24 Is AI Possible? Copeland’s Rebuttal (cont) • Searle replies to this point by asking who can believe the whole system understands Chinese, i.e. who could believe ¬B (not B – the negation of B) • Copeland’s response is to observe that he was not claiming this system could, but only that Searle’s argument (because A therefore B) was invalid, not that ¬B was true • Copeland doesn’t think ¬B is true – but this is because the Chinese room is not a plausible system for passing the Turing Test, not because we can establish a priori that computers cannot think • Argument explored in more detail in Copeland (1993) Chp 6 COM1005 2023-24 Summary • First and most famous test for whether computers can think is the Turing Test – Based on the Imitation Game – game relying on conversation to identify gender of unseen participants • Turing anticipated and refuted many potential objections to his proposed test – Copeland summarises other objections and discusses whether the TT could be adapted to take them into account • Searle’s Chinese Room is an attempt to show by analysis alone that thinking machines are an impossibility • Searle’s arguments are rejected/ignored by the AI community, who continue to push at the limits of what capabilities we can bestow upon computers COM1005 2023-24 Mandatory/Recommended Reading • Turing, A.M. (1950). Computing machinery and intelligence. Mind, 59, 433-460. (available in Reading List on BB) You may omit sections 4 and 5, which introduce digital computers to a 1950’s audience! • Wikipedia: Eugene Goostman. http://en.wikipedia.org/wiki/Eugene_Goostman • Copeland, J. (1993). Artificial Intelligence: A Philosophical Introduction. Blackwells. Section 3.4-3.6, pp. 44-57. (available in Reading List on BB) • Biever, Celeste (2023) The Easy Intelligence Tests that AI Chatbots Fail. Nature 619 , 686-689. • Searle, John. R. (1980) Minds, brains, and programs. Behavioral and Brain Sciences 3 (3): 417-457. • Russell, Stuart and Norvig, Peter (2022) Artificial Intelligence: A Modern Introduction (4th ed). Pearson. Chapter 28.1-28.2. (available in Reading List on BB) COM1005 2023-24 References Biever, Celeste (2023) The Easy Intelligence Tests that AI Chatbots Fail. Nature 619 , 686-689. Boden, Margaret (1977) Artificial Intelligence and Natural Man. Basic Books. Copeland, Jack (1993) Artificial Intelligence: A Philosophical Introduction. Blackwells. Jannai, Daniel, Amos Meron, Barak Lenz, Yoav Levine and Yoav Shoham (2023) Human or Not? A Gamified Approach to the Turing Test. https://arxiv.org/abs/2305.20010 (visited 22/09/23). McCorduck, Pamela (2004) Machines Who Think. A K Peters/CRC Press. Rich, Elaine and Knight, Kevin (1991) Artificial Intelligence (2nd ed). McGraw Hill. Russell, Stuart and Norvig, Peter (2010) Artificial Intelligence: A Modern Introduction (3rd ed). Pearson. Searle, John. R. (1980) Minds, brains, and programs. Behavioral and Brain Sciences 3 (3): 417-457. Searle, John (1999) Mind, language and society, New York, NY: Basic Books. Turing, Alan M. (1950) Computing Machinery and Intelligence. Mind LIX(236) , 433-460. Whitby, Blay (2008): Artificial Intelligence: A Beginner's Guide. Oneworld Publications. Wikipedia: Artificial Intelligence. http://en.wikipedia.org/wiki/Artificial_intelligence (visited 23/09/23). Wikipedia: Turing Test. http://en.wikipedia.org/wiki/Turing_test (visited 23/09/23). Wikipedia: Eugene Goostman. http://en.wikipedia.org/wiki/Eugene_Goostman (visited 23/09/23) Wikipedia: Chinese Room. http://en.wikipedia.org/wiki/Chinese_room (visited 23/09/23). COM1005 2023-24