Turing test

from Wikipedia, the free encyclopedia
In the usual form of the Turing test, C tries to decide whether A or B is a computer or a person. A and B try to convince C that they are human themselves.

With what was later known as the Turing test , Alan Turing formulated an idea in 1950 as to how one could determine whether a computer , i.e. a machine, had a thinking ability equivalent to a human being . This test was initially just a theoretical sketch. It was only formulated more precisely and concretely later (i.e. after Turing's suicide in 1954; see also Dartmouth Conference , 1956) after artificial intelligence had become an independent academic subject as a sub-area of computer science . Since then, this test has been on everyone's lips in the discussion about artificial intelligence and has repeatedly served to revive the myth of the thinking machine for the computer age .

Test procedure

In the course of this test, a human questioner uses a keyboard and a screen, without visual or auditory contact, to have a conversation with two people he does not know. One of the interlocutors is a person, the other a machine. If the questioner cannot say which of the two is the machine after the intensive questioning, the machine has passed the Turing test and the machine is assumed to have a thinking ability that is equal to humans.

criticism

A number of arguments have been put forward that consider the Turing test to be unsuitable for determining intelligence :

  • The Turing test only checks for functionality, not for the presence of intentionality or consciousness . This argument was worked out by, among others, John Searle in his thought experiment of the Chinese room . Turing was already aware of this problem when formulating his test, but was of the opinion that it could also be used as evidence of consciousness. Searle, however, rejects this.

Forecasts

Turing assumed that by the year 2000 it would be possible to program computers in such a way that the average user had a maximum of 70 percent chance of successfully identifying humans and machines after having "spoken" to them for five minutes. Many see the fact that this prediction has not yet been fulfilled as evidence of the underestimation of the complexity of natural intelligence .

Turing tests carried out and similar tests

Programs like ELIZA briefly appeared human to test subjects without being able to formally pass the Turing test. In their response strategy they only seemingly responded to their counterpart; the subjects were not aware that they might be dealing with non-human interlocutors.

In October 2008, an experiment at the University of Reading , in which six computer programs took part, narrowly missed the 30 percent mark. The best program managed to fool 25 percent of the human subjects.

On September 3, 2011, the AI ​​web application Cleverbot and real people took part in a test based on the Turing test at the 2011 technical festival at the Indian institute IIT Guwahati . The results were announced on September 4th. 59% of 1334 people thought Cleverbot was a human. The human competitors, on the other hand, scored 63%. However, these people were not allowed to question Cleverbot themselves, they were merely spectators. Because the test subjects could not question the machine themselves, this is not yet a valid Turing test.

Whether the chatbot Eugene Goostman passed the Turing test in 2014 is controversial.

In the summer of 2017, researchers at the University of Chicago presented an AI that can independently write reviews. These machine-generated reviews, along with human-written reviews, were presented to 600 subjects for evaluation. On average, they rated the reviews created by the AI ​​in the blind test as useful as the reviews written by humans. In this test arrangement, the Turing test is passed because it was no longer possible for people to recognize which reviews were made by machine and which were by humans.

In July 2017, researchers at Rutgers University presented an AI that produces images that appear artistic. The AI ​​has been trained with many paintings by famous painters from different eras. In a blind test, the images created by the AI were mixed with images by contemporary artists that had been exhibited at Art Basel and presented to 18 test persons for assessment. The test subjects were asked to assess whether the images were created by humans or a computer. The test subjects rated the AI ​​images as better overall, i.e. more man-made than the paintings created by the artists for Art Basel. When compared to large works of abstract expressionism , the AI ​​works performed worse than the man-made works.

In May 2018, Google presented its "Duplex" system at the developer conference. The AI ​​made a call to a hair salon, restaurant, etc. to make an appointment. The goal of Google is to make the AI ​​language appear so natural that the other person no longer recognizes that the caller is a machine. For this purpose, the KI u. a. Pauses in thought, deliberate inaccuracies and sounds like "aha" and "hmm" etc. added, which should make the AI ​​sound human. Commentators felt the result u. a. as terrifyingly convincing. So far, the system only works in English. Strictly speaking, the demonstration is not a real Turing test, as the conditions for such a test were not given. For example, all viewers knew beforehand which side of the conversation the human was on and which the AI ​​was on. In addition, the conversation in this test was only focused on making an appointment, as soon as the human had steered the conversation in a different direction, the AI ​​would most likely have failed. Nonetheless, the system is considered a great achievement in the field of artificial intelligence.

Practical meaning

In order to ward off spam , it is necessary to distinguish between automated input and input from humans. The CAPTCHA method often used for this derives its name from the Turing test ( Completely Automated Public Turing test to tell Computers and Humans Apart ). Another name for this method is Human Interaction Proof (HIP).

Advanced concepts

In order to address the fundamental shortcomings of the Turing test, alternative, more comprehensive concepts have been proposed, e.g. B.

  • Lovelace test - an AI must demonstrate creativity and produce original services.
  • Metzinger test - an AI must intervene in the discussion about artificial consciousness with its own arguments and argue convincingly for its own theory of consciousness.

Loebner Prize

The Loebner Prize has been advertised since 1991 and is to be awarded to the computer program that is the first to pass an extended Turing test, in which multimedia content such as music, speech, images and videos must also be processed. The award is named after Hugh G. Loebner and is endowed with 100,000 US dollars and a gold medal, a silver medal and 25,000 dollars are awarded for passing the written Turing test. So far, however, no computer program has been able to meet the necessary requirements. Furthermore, a Loebner Prize is awarded annually to the computer program that comes closest to a human conversation. This is endowed with 4,000 US dollars and a bronze medal.

Cultural references

  • Philip K. Dick used androids of electric sheep? (1982 filmed under the title Blade Runner ) the so-called Voigt-Kampff test , a variant of the Turing test. In 1992 (or in later editions 2021) artificial people, so-called replicants, who physically resemble humans, will be subjected to an empathy test, which will examine their emotional reaction through lengthy interviews and determine whether they are human or replicant.
  • In the 2K BotPrize , human testers check bots for human reactions in the computer game Unreal Tournament 2004 . The aim is to program a bot in such a way that it can no longer be distinguished from a human player.
  • In Ian McDonald's science fiction novel River of Gods (2004; German: Cyberabad , 2012) a world is designed in which artificial intelligences are forbidden from a more highly developed level comparable to humans. The Turing test for testing the level of AI is rejected here with the argument that a sufficiently highly developed intelligence could also provoke one's own failure in the test.
  • In the book / audio book of German author , songwriter , small artist and comedian Marc-Uwe Kling " Quality Country intentional failure in the Turing test is also treated:" An AI is intelligent enough to pass the Turing Test, could also be smart enough him not to pass ".
  • In Alex Garland's directorial debut Ex Machina , the programmer Caleb is supposed to test the artificial intelligence Ava using a Turing test in a modified form, because he already knows that Ava is a robot.
  • In Neuromancer trilogy of William Gibson artificial intelligences play a central role of the Turing Police are monitored and switched off as soon as they are intelligently and develop their own ambitions.
  • In the TV series Westworld (since 2016), based on the 1973 film of the same name , wealthy vacationers spend time in an amusement park created especially for them and populated by androids, so-called "hosts". It is mentioned several times that the hosts already passed the Turing test after the first year of development.

literature

  • Alan M. Turing: Computing Machinery and Intelligence . In: Min . tape LIX , no. 236 , 1950, ISSN  0026-4423 , pp. 433–460 , doi : 10.1093 / mind / LIX.236.433 ( loebner.net - basic article of artificial intelligence, suggests the “Turing test” for checking the thinking ability of a machine).
  • Alan M. Turing: Computing Machinery and Intelligence . In: Robert Epstein, Gary Roberts, Grace Beber (Eds.): Parsing the Turing Test . Philosophical and Methodological Issues in the Quest for the Thinking Computer. Springer, 2008, ISBN 978-1-4020-6710-5 , pp. 23–65 , doi : 10.1007 / 978-1-4020-6710-5_3 (first edition: 1950, with comments in the text by Kenneth Ford, Clark Glymour, Pat Hayes, Stevan Harnad and Ayse Pinar and a separate commentary by John Lucas).

English editions

  • Darrel C. Ince (Ed.): Mechanical intelligence (=  Collected Works of AM Turing . Band 1 ). North Holland, Amsterdam 1992, ISBN 978-0-444-88058-1 .
  • Jack Copeland (Ed.): The Essential Turing . Seminal Writings in Computing, Logic, Philosophy, Artificial Intelligence, and Artificial Life plus The Secrets of Enigma. Oxford University Press, Oxford 2004, ISBN 0-19-825080-0 .
  • S. Barry Cooper, Jan van Leeuwen (Eds.): Alan Turing: His Work and Impact . Elsevier, New York 2013, ISBN 978-0-12-386980-7 (Contains almost all of the 'Collected Works', with extensive commentary, but without concordance).

German edition and translations

  • Alan M. Turing: Can a machine think? In: Hans Magnus Enzensberger (Hrsg.): Neue Mathematik (=  course book ). tape 8 . Suhrkamp, ​​Frankfurt a. M. 1967 (Original title: Computing Machinery and Intelligence . 1950. Translated by P. Gänßler).
  • Friedrich Kittler , Bernhard Dotzler (Ed.): Intelligence Service. Fonts . Brinkmann et al. Bose, Berlin 1987, ISBN 3-922660-22-3 (Contains very informative introduction. Extensive selection).
  • Alan M. Turing: Can a machine think? In: Walther Ch. Zimmerli, Stefan Wolf (Ed.): Artificial Intelligence. Philosophical problems . Reclam, Stuttgart 1994, ISBN 3-15-008922-0 (Original title: Computing Machinery and Intelligence . 1950. Translated by P. Gänßler).

further reading

  • Brian Christian: The Most Human Human . What Artificial Intelligence Teaches Us About Beeing Alive. Anchor Books, New York 2011, ISBN 978-0-307-47670-8 (experience report and evaluation of a philosopher who took the Turing test as a person).
  • Donald Davidson : Turing's Test . In: Problems of Rationality . Suhrkamp, ​​Frankfurt am Main 2004, ISBN 3-518-58471-5 (Original title: Turings Test . 1990. Translated by Joachim Schulte, first published in K. Said (Ed.): Modeling the Mind , Oxford University Press).
  • Graham Oppy and David Dowe: The Turing Test . In: Edward N. Zalta (Ed.): The Stanford Encyclopedia of Philosophy . Spring 2011 ed. 2011 ( plato.stanford.edu - first entry summer 2003).
  • Ayse Pinar Saygin et al .: Turing Test: 50 Years Later . In: Minds and Machines . tape 10 , 2000, ISSN  0924-6495 , p. 463-518 ( crl.ucsd.edu [PDF] Review (status: 2000)).
  • Stuart M. Shieber (Ed.): The Turing Test: Verbal Behavior as the Hallmark of Intelligence . MIT Press, Cambridge MA 2004, ISBN 0-262-69293-7 (collection of articles on the prehistory and debate of the Turing Test with introduction and summaries).

Web links

Commons : Turing test  - collection of images, videos and audio files

Individual evidence

  1. ↑ Alan M. Turing: 4 . In: Computing machinery and intelligence  (= 59). Mind (journal), 1950.
  2. ^ John Searle: The Rediscovery of the Mind . MIT Press, Cambridge MA 1992.
  3. ↑ Computers still not quite clever enough to fool humans, Turing Test shows . In: The Daily Telegraph , October 12, 2008.
  4. ↑ Software tricks people into thinking it is human . New Scientist; Retrieved March 14, 2012.
  5. ↑ Martin Holland: "Eugene" and the allegedly passed Turing Test: It's not that easy ... heise online, June 10, 2014
  6. ↑ "Researchers taught AI to write totally believable fake reviews, and the implications are terrifying" In: businessinsider.de , 29 August 2017.
  7. ^ "Humans Prefer Computer-Generated Paintings to Those at Art Basel" In: hyperallergic.com , July 31, 2017.
  8. ↑ Christian Gall: Can computers also produce art? In: Augsburger Allgemeine. Retrieved February 29, 2020 .
  9. ↑ Google Duplex is scary good , spiegel.de from May 9, 2018
  10. ↑ Google Duplex: 8 Questions & Answers on Voice AI , turn-on.de from June 1, 2018
  11. ^ Artem Oppermann: Did Google Duplex beat the Turing Test? Yes and No. September 26, 2019, accessed October 25, 2019 .
  12. ↑ Selmer Bringsjord en , Paul Bello, David Ferrucci en : Creativity, the Turing Test, and the (Better) Lovelace Test (=  Studies in Cognitive Systems . No. 30 ). Springer Netherlands, 2003, ISBN 978-94-010-0105-2 , pp. 215-239 , doi : 10.1007 / 978-94-010-0105-2_12 (25 p., Kryten.mm.rpi.edu [PDF; 2.1 MB ]).
  13. ↑ Thomas Metzinger : Postbiotic Consciousness: How to Build an Artificial Subject - and why we shouldn't do it. In: Paderborn Podium / 20 Years HNF. Heinz Nixdorf MuseumsForum , October 24, 2001, accessed on December 1, 2016 .
  14. ↑ Lars Sobiraj: Unreal Tournament 2004: When Bots Play More Humanly Than Humans ( Memento from January 25, 2013 in the web archive archive.today ), gulli.com
  15. ↑ David Cornish: Unreal Tournament bots appear more human than humans . Ars Technica, September 29, 2012; Retrieved October 1, 2012.
  16. ↑ botprize.org
  17. ↑ buchwurm.org: William Gibson - Neuromancer