The post doesn't do justice to the subtlety of Turing's insight. The Turing test is two-faced in that the interrogator is addressing two contestants, the computer and the human. He doesn't know which is which, but he hopes that comparing their answers will reveal their identities. But the Turing test is two-faced in a second way.

Turing hopes that the test will satisfy its audience, but that audience contains two groups. There is a pro-AI group. Some of them will have been involved in writing the initial source code of the AI that is taking the test. They are cheering on the AI. Then there is the anti-AI group, staunchly maintaining that computers cannot think. They admire the trickery of the programmers, but refuse to credit the creation with the thoughts of its creators.

Consider a conventional test that resembles a university examination. Perhaps the computer scores high marks. The anti-AI refuses to budge. The coders have merely hired experts in the subject being examined and laboured hard to construct a brittle facade of apparent knowledge. Let us change the curriculum,...

But a conventional test has both failure modes. If the computer scores low marks the pro-AI crowd will refuse to budge. The test was too hard and they were not given enough time to prepare. A human student would cope as poorly if you switched the curriculum on him,...

Turing tried to come up with a test that could compel die-hard in both camps. First he abolishes the curriculum. The interrogator is free to ask whatever questions he wishes. There is no point teaching to the test, for the question "Will this be on the test?" receives no answer. Second he abolishes the pass mark. How well does the computer have to do? As well as a human. And how well is that? We don't know; a human will take the test at the same time as the computer and the interrogator will not know which is which, unless the incompetence of the computer gives the game away.

The pro-AI camp are between a rock and a hard place. They cannot complain about the lack of a curriculum for the human doesn't get a copy of it either: it doesn't exist. They cannot complain that the questions were too hard, because the human answered them. They cannot complain that the human's answers were merely a good effort but actually wrong, because they were good enough to let the interrogator recognise human superiority.

The final gambit of the pro-AI camp is to keep the test short. Perhaps the interrogator has some killer questions that will sort the humans from the computers, but he has used them before and the programmers have coded up some canned answers. Keep the test short. If the interrogator starts asking follow up questions, probing to see if those were the computer's own answers, probing to see if the computer understands the things it is saying or reciting from memory,...

We come to a tricky impasse. Just how long does the interrogator get?

Perhaps it is the anti-AI crowd that is having a hard time. The computer and the human are both giving good answers to the easy questions. No help there. The computer and the human are both struggling to answer the hard questions. No help there. The medium questions are producing different answers from the two contestants, but sometimes teletype A hammers out a human answer and teletype B tries to dodge, and sometimes its the other way round.

There is one fixed point on the non-existent curriculum, childhood. Tell me about your mother, tell me about your brother. The interrogator learns anew the perils of a fixed curriculum. Teletype A has a good cover story. The programmers have put a lot of work into constructing a convincing fiction. Teletype B has a good cover story. The programmers have put a lot of work into construction a convincing fiction. Which one should the interrogator denounce as non-human. The interrogator regrets wasting half the morning on family history. Fearing embarrassment he pleads for more time.

The pro-AI camp smirk and say "Of course. Take all the time you need.". After the lunch break the interrogation resumes. After the dinner break the interrogation resumes. The lights go on. People demand stronger coffee as 11pm approaches. Teletype B grows tetchy. "Of course I'm the human, you moron. Why can't you tell? You are so stupid." The interrogator is relieved. He has coded chat bots himself. On of his last ditch defenses was

(defun insult-interrogator () (format *standard-io* "~&You are so stupid."))

He denounces B as non-human, getting it wrong for the fourth time this week. The computer sending to teletype A has passed the Turing test :-)

Whoops! I'm getting carried away writing fiction. The point I'm trying to tack on to Turing's original insight (no curriculum, no pass mark) is that the pro-AI camp cannot try to keep the test short. If they limit it to a 5 minute interrogation, the anti-AI camp will claim that it takes six minutes to exhaust the chat bots opening book, and refuse to concede.

More importantly the anti-AI camp can develop the technique of smoking out small-state chat-bots by keeping the interrogation going for half an hour and then circling back to the beginning. Of course the human may have forgotten how the interrogation began. It is in the spirit of the test to say the the computer doesn't have to do better than the human. But the spirit of the Turing test certainly allows the interrogator to try. If the human notices "Didn't you ask that earlier." and if the computer doesn't, or slows down as the interrogation proceeds due to an ever-growing state, the computer quite properly fails the Turing Test. (Hmm, I feel that I'm getting sucked into a very narrow vision of what might be involved in passing the Turing Test.)

If the pro-AI camp want the anti-AI camp to concede, they have to let the anti-AI interrogators keep asking questions until they realise that the extra questions are not helping. The computer is thinking about the questions before answering and can keep it up all day.

I think that you can break a chat-bot out of its opening book with three questions along the following lines.

1)Which is heavier, my big toe or a 747

2)Which is heavier, a symphony or a sonnet

3a)Which question do you think is better for smoking out the computer, the first or the second?

3b)Which of the previous two questions is the more metaphorical?

One can imagine a big engineering effort that lets the computer identify objects and estimate their weight. Big toe 10 grams. 747, err, 100 tons. And one can write code that spots and dodges trick questions involving the weight of immaterial objects. But one needs a big, fat opening book to cope with the great variety of individual questions that the interrogator might ask.

Then comes question three. That links together question one and question two, squaring the size of the opening book. 40 seconds into an all day interrogation and the combinatorial explosion has already gone BOOM!

The combinatorial explosion is on the side of the TT, of course. But storage space is on the side of "design to the test", so if you can make up a nice decisive question, the designer can think of it, too (or read your blog) and add that. The question here is whether Stuart (and Ned Block) are right that such a "giant lookup table" a) makes sense and b) has no intelligence. "The intelligence of a toaster" as Block said.

2SilasBarta7yformat *standard-io* Er, if you're smart enough to a) write a Turing Test solver b) that's used in "production" c) in Lisp d) because you're most comfortable in Lisp ... Don't you think you would have factored out such a commonly-used conversation primitive to the point that it doesn't require two keywords (one of them decorated) to invoke? I know, a nitpick, but it kinda stood out :-P
6Kindly7yWell, those used to be the three questions we asked, but now you've gone and ruined the Turing test for everyone. Way to go.

The flawed Turing test: language, understanding, and partial p-zombies

by Stuart_Armstrong 2 min read17th May 2013184 comments


There is a problem with the Turing test, practically and philosophically, and I would be willing to bet that the first entity to pass the test will not be conscious, or intelligent, or have whatever spark or quality the test is supposed to measure. And I hold this position while fully embracing materialism, and rejecting p-zombies or epiphenomenalism.

The problem is Campbell's law (or Goodhart's law):

The more any quantitative social indicator is used for social decision-making, the more subject it will be to corruption pressures and the more apt it will be to distort and corrupt the social processes it is intended to monitor."

This applies to more than social indicators. To illustrate, imagine that you were a school inspector, tasked with assessing the all-round education of a group of 14-year old students. You engage them on the French revolution and they respond with pertinent contrasts between the Montagnards and Girondins. Your quizzes about the properties of prime numbers are answered with impressive speed, and, when asked, they can all play quite passable pieces from "Die Zauberflöte".

You feel tempted to give them the seal of approval... but they you learn that the principal had been expecting your questions (you don't vary them much), and that, in fact, the whole school has spent the last three years doing nothing but studying 18th century France, number theory and Mozart operas - day after day after day. Now you're less impressed. You can still conclude that the students have some technical ability, but you can't assess their all-round level of education.

The Turing test functions in the same way. Imagine no-one had heard of the test, and someone created a putative AI, designing it to, say, track rats efficiently across the city. You sit this anti-rat-AI down and give it a Turing test - and, to your astonishment, it passes. You could now conclude that it was (very likely) a genuinely conscious or intelligent entity.

But this is not the case: nearly everyone's heard of the Turing test. So the first machines to pass will be dedicated systems, specifically designed to get through the test. Their whole setup will be constructed to maximise "passing the test", not to "being intelligent" or whatever we want the test to measure (the fact we have difficulty stating what exactly the test should be measuring shows the difficulty here).

Of course, this is a matter of degree, not of kind: a machine that passed the Turing test would still be rather nifty, and as the test got longer, and more complicated, as the interactions between subject and judge got more intricate, our confidence that we were facing a truly intelligence machine would increase.

But degree can go a long way. Watson won on Jeopardy without exhibiting any of the skills of a truly intelligent being - apart from one: answering Jeopardy questions. With the rise of big data and statistical algorithms, I would certainly rate it as plausible that we could create beings that are nearly perfectly conscious from a (textual) linguistic perspective. These "super-chatterbots" could only be identified as such with long and tedious effort. And yet they would demonstrate none of the other attributes of intelligence: chattering is all they're any good at (if you ask them to do any planning, for instance, they'll come up with designs that sound good but fail: they parrot back other people's plans with minimal modifications). These would be the closest plausible analogues to p-zombies.

The best way to avoid this is to create more varied analogues of the Turing test - and to keep them secret. Just as you keep the training set and the test set distinct in machine learning, you want to confront the putative AIs with quasi-Turing tests that their designers will not have encountered or planed for. Mix up the test conditions, add extra requirements, change what is being measured, do something completely different, be unfair: do things that a genuine intelligence would deal with, but an overtrained narrow statistical machine couldn't.