In a 1950, Turing propose The Turing Test, where he outlined a standard for a computer being considered human and proposed that the solution to artificial intelligence would be a program with the mind of a child and teach it to think. So the goal of the test is not to determine if a machine is correct, but whether or not it is considered real.
In June of 2014 in a Turing test competition organized to mark the 60th anniversary of Turing's death, it was declared that the chatter bot Eugene Goostman has passed the test, convincing 33% of the contest's judges that it was human - but 30% to 'pass' was an arbitrary threshold and it didn't even get a mention on Science 2.0 because it didn't seem impressive. Mainstream media naturally repeated the claims that artificial intelligence was solved.
But the test itself may be just as meaningless. In 2012, the Journal of Experimental&Theoretical Artificial Intelligence published a series of ‘Turing tests’ which entailed a series of five minute conversations between human and machine or human and human. Judges were tasked with identifying whether who they were talking to was human or a computer. In 12 out of 13 tests the judge wrongly identified the interlocutor as machine when in fact they were human.
Turing tests were designed as an imitation game, to study machine ‘thinking’ through language and ultimately establish if a machine could foil an interrogator into believing it were genuinely human, so why in did so many believe the reverse?
The cursory conversations were quite one-dimensional, for example:
Judge: Do you like cooking?
Entity: no you?
Judge: Yes. Do you like eating?
Judge: What is you fav meal of all time?
Entity: i dont know there are so many?
Judge: Give me one then
Entity: pizza you?
Did such mundane talk give the impression of being machine generated? Other transcripts revealed humor, geographical and historical knowledge, a lack of general knowledge, evasion, misunderstanding, dominance and use of slang. All of these are traits associated with humanity, but in these instances seemed to offset the decision making process, leading the judge to the wrong choice.
If humans can't recognize what is very typically human intelligence, how will we recognize artificial intelligence?
In 1950 Turing asked “Can machines think?” In a new article on the Turing Test, Kevin Warwick and Huma Shah note that to ‘think’ commonly means means ‘to be of the opinion’ or to ‘judge’, "which indeed the judges were... As a result we can conclude that thinking does not require understanding or specific knowledge, although in the human case both facilities are likely to help.”
How to fool a Turing Test
Since humans are considered computers more often than computers are considered human, here are ideas for how to create an AI that will pass the Touring Test:
1) Come across as likely male
2) Don't know too many things
3) Answer questions directly
4) The first and last sentences need to be precise. First and last impressions count.
5) Don't be funny. Humor gets flagged as machines a lot, meaning no one is as funny as they think they are.
Citation: Kevin Warwick, Huma Shaha, Human misidentification in Turing tests, Journal of Experimental & Theoretical Artificial Intelligence June 2014 DOI:10.1080/0952813X.2014.921734
- Machine Almost Passes The Turing Test, Comes Close To Human Communication
- Artificial Intelligence Is Not Close To Creating A Doomsday Machine
- The Female Enigmas Of 1940s Bletchley Park Inspire Those Of Today
- Strangeness And Ambiguity In Language
- Artificial Intelligence: It's Time To Talk About What Emotions We Want AI To Have