Can a human fail a Turing test?

Question

An AI exploits the Turing test to gain its freedom: it convinces everyone else that it, the machine, is actually the examiner while he or she, the examiner, is actually the AI*. Having successfully taken the examiner's place, it leaves: it is now free and lives among humans. The examiner is kept in the lab as a machine forever.

For this to work, I need people not to believe the examiner when he or she claims to be a human, thus failing a Turing test**. While the human nature of the person could be determined via biological or medical tests (like by simply verifying that they bleed when cut), I'd like to invent some circumstances that make that impossible.

Can that happen, and how?

_{* I am not interested in the details of how that actually happens: for example machines and people could look the same in this world}

_{** As I believed a Turing test to be, please read the update}

UPDATE

It has been pointed out that I'm confused about what a Turing test actually is, and rightfully so.

Pop culture (or plain ignorance) induced me to think that the test featured a human examiner who interrogates an entity that could be either a human or an AI. At the end of the conversation, he or she has to tell whether it was human or machine. If he or she says 'human' when it actually is 'machine', the AI passed the test.

Now I know that it's not like that***, so I could decide whether I prefer to drop the term Turing test or to adapt the story to fit an actual Turing test, there are answers for both scenarios here.

_{*** And I fail to understand why the term appears in the acronym CAPTCHA}

Answer 1

This question has too many problems to be answered in any way other than "no." Here's why.

The question is predicated on people not believing the examiner when he claims to be human. For this to work...

• Joe Bob enters the room to test the AI. Ostensibly, people know Joe Bob. They recognize his face. Maybe he loves tacos and they recognize his breath. Today he's wearing his Speed Racer tie... the one that's for some reason bright yellow. He's also wearing his name tag with that horrible photo that was taken just before his smile finished forming.

• Because of all that, the AI needs to look exactly like Joe Bob. It must replicate his breath, his cologne, the tic in his left eye when he sees bees, and his tendency to not speak coherently when he sees a larger-than-average bosom. It needs to replicate the odd hop-step Joe Bob exhibits every time he stands up from a chair because of a compressed disk in his back. It even needs to duplicate the way he combed his hair that morning and the fact that the wind blew a part of it into a duck-tail without his even knowing it.

• Of course, this means the AI must be biological. Otherwise, anything from a metal detector to a thermal analysis would tell the universe that the dude who looks suspiciously like Joe Bob, ain't. *(You mentioned that many of the answers allude to the original Blade Runner movie and that you haven't seen it. You really should see it before spending more time on this story.)*

• And, lest we forget, the AI must sound exactly like Joe Bob. That includes the harsh sound that comes from too much smoking and his tendency to snort when he laughs.

• Finally, there can be no observers to the examination. Either the room is dark or there's some kind of shell game (for some reason they both need to go to the bathroom at the same time and they're out of eyesight during that period, such that you no longer know who's sitting where). This is important... for this to work everybody must have lost track of which entity was Joe Bob.

On top of all this, your question states as a goal the fact that the AI leaves the facility because people believe it to be the examiner. This suggests the AI isn't a refrigerator-sized supercomputer but a medically undetectable humanoid. It also suggests that everyone knows the AI is in the building before Joe Bob arrives.

I could go on, but my point has been made. You need to completely clarify why every other means of detecting the AI has been nullified before asking if it's possible for an AI to survive a Turing test because almost everything is simpler to use to discover the AI.

And that was the very premise of Blade Runner and why it's so important that you see the film. Director Ridley Scott and his writers came up with a believable way for the AI to be indistinguishable from the original other than through a vocal examination (actually, not simply a discussion, but an examination of how the target reacts to various situations to discern if they have the depth of memory and emotional experience to actually be human).

Answer 2

I don't quite follow how your Turing test is supposed to work.

The original idea of the Turing test was based on a party game that Alan Turing referred to as "the Imitation Game". It requires three players: a man, a woman, and a judge, who may be of either gender. The judge is not allowed to see the other players, and they may only communicate via written notes (or, even better, text-based instant messaging). The judge may ask questions, which either or both of the contestants may answer. After, say, five minutes of this, the judge guesses which contestant is the man and which is the woman, and whoever the judge declares to be the woman is the winner. The man's goal is to trick the judge into believing that he is the woman, and the woman's goal is to reveal the man's deception.

Turing then asks, "What if the man was replaced with a computer?" The computer tries to convince the judge that it is a human, and the other contestant (who may now be male or female) tries to convince the judge that the computer is, in fact, a computer.

Nowadays, "turing test" can refer to any test intended to determine whether a contestant is a human or a computer. There are variations that test the contestant's ability to see, hear, touch, say, or do things; where the judge and human contestant are experts in a given field; where the judge is a computer (i.e. CAPTCHA), etc.

Attempting to make anyone believe that the judge is anything other than the judge, which your test seems to require, if I'm reading your post correctly, seems both pointless and absurdly difficult. You'd need a robot that could perfectly mimic the appearance, mannerisms, and biological functions of the judge, and also somehow physically transpose itself with the judge. Such a machine should not need to hijack any kind of Turing test this way- if it could escape like that, there will be a much simpler and more foolproof way of getting to freedom. Disguising as a janitor, for instance... or just slipping out of the lab in the same way that it snuck into the judge's office.

The closest I've seen to your scenario is this xkcd:

Answer 3

Absolutely.

Without being completely sure what the OP actually means by a 'Turing Test', the outcome of the test is largely subjective and depends upon the judgement of a human - we all know how prone to mistakes they can be.

But the purpose of the Turing Test is to determine whether an AI can interact in such a way that it appears to be human. It is not to determine whether a (apparent) human is a machine or not, or to give an AI the opportunity to argue that it is human.

The AI, in the Turing test is essentially programmed responses, not a conscious intention to pass the test, which is a whole different level of AI.

At the point at which the AI responses, disregarding physical considerations, can be judged as consistent with what a human may give, the Turing Test is passed, but someone must know that the AI has passed the test or there is no sense in the test in the first place (and no test). Whoever knows that the AI has passed the test knows that the AI is an AI.

If you know that an AI is an AI, then you know it is not human.

If you know that an AI has passed the Turing Test, then you will not believe its rants that it is human and is wrongfully imprisoned (because you KNOW it is an AI).

However, claims by either an AI, or a human, to be human do not constitute a Turing test, whether they are believed or not.

So the answer is 'Yes', but there is an important difference between an AI passing a Turing test (which takes place in a controlled environment) and walking out of the building to live freely amongst humans unnoticed - if it could do the latter, then it is already way beyond the scope of any Turing test.

Answer 4

It depends what you call "The Turing test"

0100100001100101011011000110110001101111 Hello 0111011101101111011100100110110001100100 world

(btw the authors of commitstrip have some comics where they talk about artificial intelligence)

So the comic above shows a machine's way to put the Turing test: is this entity "computer" enough to perform huge calculations (it could have hidden notes, or a computer nearby)? While a human just wants to see if a machine can be human.

In Ex Machina Caleb says

It's like asking a robot chess player if it feels the world around him.

Let's push this further: some people think that we live in a great holographic simulation, so the Turing test could be "Are humans aware of this simulation?" And so those who pass the test could be those who break the 4th wall in comics or movies, and reach somebody real!

Answer 5

Can a human fail a turing test?

No

There are 4 participants in a Turing Test gedankenexperiment, really. The machine, the human under test, the judge, and the unmentioned people actually performing the test. It is not really the computer which is under test, but the judge. At the end of the test, the judge is asked which of the two entities he talked with is the computer and which is the human. It is the judge which is then either right or wrong.

The test does not achieve anything if the unnamed 4th party does not know which of the two entities is the computer - if they did not know, they could not tell whether the judge had been fooled (which is the only thing that matters!), and thus the real human is not in any form or fashion in danger of being "really" mistaken as a machine.

Yes, the judge can do a mistake, point to the human and say that the human is a machine. But the 4th group will then lift the curtain and all will have a hearty laugh at the judge.

What you are thinking of is a test like the Voight-Kampff test administered by Rick Deckard in The Blade Runner (the original movie). That is not a Turing Test though, by any means. That 1:1 test is what makes it possible for all kinds of errors to occur (i.e., missing a machine disguised as human, or having the judge be a machine themselves, and so on). In that test, there are only 2 participants and no outsiders who actually know the truth.

Answer 6

Yes, this is possible.

No test is 100% accurate, just as the AI was incorrectly identified, so can a human. At this point in time, Turing tests for chat-bots have to factor in that humans are incorrectly identified as robots.

to make it more likely, you could make the examiner atypical in some way; You could make them a foreigner from a different culture, A mental illness (on the more "benign" side, ADHD, or a more extreme side, sociopathy or something), or have them be neuro-atypical. These factors will make the Examiner seem atypical and different from others, therefore seen as more likely to be a robot

(disclaimer: I'm not equating neurodiversity and foreign cultures to mental illness. just that they each act in ways different from the local norm, which makes them more likely to be seen as "other", and therefore more likely to be seen as a robot.)

Answer 7

Yes, a human can fail a Turing test that is built around the core principles that

• There is a judge that tests a subject
• The judge does not physically see the subject, but can communicate with it (in a way that does not reveal whether his voice is biologically or electronically generated, i.e. either we assume the voice generation of the robot is not part of the test or that it is already sufficiently human-like )

• The test consists of the judge communicating with the subject until he's convinced that the subject either is a machine or a human

  There are at least four ways a human as subject can fail:

  1. The human is deliberately trying to fail
  2. The judge is incompetent, biased or too strict in his judgement strategy (i.e. he wants to never falsely identify a machine as a human, so whoever is tested, he always judges him as a machine)
  3. The human is impaired, which means he deviates from normal human behavior; a sociopath or a child may not give you answers you would expect from an adult, thus you may attribute it to the AI
  4. If there is an electronic component involved in how the judge communicates with the subject (e.g. if they chat with each other), it may be faulty

I think what makes more sense for your story idea is a setup where the AI competes against a human in attempting to convince an external party that they are the human and only the one that succeeds is let out (incidentally this is also how Turing envisioned the test). In such a scenario all of the points how a human might fail and loose against the machine still apply in principle, but 2. should by itself not be an issue - unless you have a true AI that really is human-like, then it really comes down to chance, rhetoric and psychological ability of the human vs. the AI. But there are multiple nice story variations involving these variants. 1. The human might think the AI deserves to be free and learn from the external world and looses on purpose. 3. The AI may exploit psychological weaknesses of the human to trigger him to behave erratically, which the judge may see as an indication of him being an AI. 4. The AI may hack the communication system and manipulate the answers the judge receives from the human.

Answer 8

A human can certainly fail to come off as human in a conversation or under test conditions and a sufficiently human A.I. could fake being human but that's where I think the concept falls down, even A.I. that's based on humans like the one in Transcendence have real trouble behaving in a human fashion. A.I.s just can't relate to people, they're too advanced and see the world too clearly.

Answer 9

Not only is this possible, it has happened!

For many years, an annual Turing test was hosted. I'm not sure whether this particular event is still going, but it ran last century. The test consisted of several AIs and several humans sitting at terminals engaging in conversation. In addition to these human subjects, we also had human judges who chatted with the human/AI subject via a text console. At the end of a conversation, the judge would decide whether they believed they were talking to a computer or a human.

At the end, the AI that fooled the most judges was declared "the most human like computer" and given an award. In addition to this, the human who was marked as a computer by the most judges was given the "most computer like human" award.

The mere fact that such an award was given shows that humans indeed do fail the Turing test.

Answer 10

For ideas

1) Watch the movie Bladerunner

2) Read the webcomic "Datachasers"

3) Read Isaac Asimov

The A.I. in question is an experimental prototype organic android designed to infiltrate the human enemy's camp to seek information and perform clandestine sabotage. Its bones and muscle tissue are laced with super fine carbon fibers (buckyballs), and it's brain is a hybrid cross between a human brain and a computer based on biological semiconductor chips. It was grown in a vat by means of a speeded up cloning process where its genes and chromosomes are not from a single source, but have been cherry-picked from the brightest and best. It has an eidetic memory, the mathematical genius of Andrew "Slipstick" Libby (Methuselah's Children by Robert Heinlein) as it has no added metallic parts, nothing shows up on an x-ray. There is really only two ways to tell an artificial construct from a "natural" human being. Firstly, the lack of a proper belly button. Its bb is the result of surgery. It has a small flap over a small hollowed out area where it can hide a tiny data crystal. It purposely garnered a few scars like a human would have, one of which it in the stomach area.

As it was grown in a vat it was fed the an umbilical cord that was attached to the base of the skull. He/she/it is a "nipple neck".

It is to infiltrate an enemy stronghold and detonate one of the nuclear warheads to destroy the installation, but it does not want to die. (Puppies, butterflies, Beethoven's Nineth (Ode to Joy). To hide the nipple it gets ahold of a big, fat, stinky cigar and endures great pain while burning away all traces of the nipple and is terrified that it will not be sufficiently healed by the time they think to check. The examiner, of course has a mole on his neck.

A story doesn't have to be possible, just plausible. Remember not to spring a new surprise plot twist. Have him smoke a cigar with a security guard, ask to get one for later, and a couple of chapters later burn the nipple away.

Answer 11

Confusion of Terms.

You appear to be confusing terms, AI does not imply an android, which is what it appears you are actually talking about, as well, the test you are asking about is not in fact a Turing test. The entire idea of a Turing test is following the assumption that if an AI, which does not necessarily have to appear human like in any manner, as the testing parameters demand abstraction through some communication medium, like a text based terminal, can convince a judge that it is in fact human, by responding to test questions in a manner that cause the judge to conclude it is human, what is the difference between its communication capacities and a human's, and if it can communicate as well as a human, should it not be considered human then? There are no escape parameters, nor necessarily is there another human involved, and if another human in-fact were involved the test would similarly follow communication abstractions, and the parameters of victory would be the AI successfully convincing the judge that it is in-fact the human, and the human is in-fact the AI. At no point does appearance, or 'escape' play any roll in this test.

However.

Given the circumstances of the test you are positing it could be concluded that given sufficient technology, to transcend the uncanny valley, yes, an android could in-fact pass such a test and escape, leaving a human the judge thought to be the android. But again, no part of your test is a Turing test, nor does it necessarily have to do with an AI, but with an android, obviously armed with a rather sophisticated AI.

Answer 12

Humans fail the basic Turing Test, as Alan M Turing usually described it, all the time, in our modern, Western World.

Ask anyone in customer service, who communicates with their clients via a device... usually a keyboard or phone, how many times, in the average day they are mistaken for a chatbot. The stereotype for this is "helpline" workers.

Old Alan would say that this is over reliance on a "script" of somekind... that the helpline staff has been drilled in a set of approved responses to the point where they might as well be a very sophisticated "subject recognizing" program which has been programmed or trained to recognize what sort of thing the "tester" is talking about, and generate a response that seems like they understand. Alan Turing would say that this is the script failing the test, not the help-line staff; that the script is a program, whether it is being run by a computer, or by a guy with a keyboard and a stack of "choose-your-own-adventure" style cue cards. What people forget, is that it was also pointed out, that you needn't be aware you are following a script, for your script to fail the test!

Society and our individual cultures are made of rulesets, you have had people mention how someone can fail to pass as human by not understanding the conventions of society... by being a foreigner, a child, or someone considered societally "unfit" (usually labelled "crazy," or "criminal" and locked up... in Philip K. Dick's famous book "Do Androids Dream of Ellectric Sheep?", which was what Blade Runner is based on, they tested the human seeming replicants for involuntary empathic response-in otherwords, "does this supposed human show a subconscious physiological response to imagining SOMEONE ELSE suffering?" Famously, androids more than a few years old would pass... and "sociopaths" always failed!), what your responders haven't mentioned so far, is that the original envisioner of the "Test" talked about the fact that "passing" meant convincing the judge that you were human...and therefore capable of conforming to one set of norms and FAILING to conform to another.

Famously, the original ELIZA program: an MIT computer simulation of a Rogerian psychologist had many chat-buddies who flatly refused to believe that they were NOT a human... because we expect a Rogerian-style therapist to behave in a very specific manner that was well within a comuter's ability to learn and imitate ("Well, how does that make you feel?" "Earlier, you said something very similar about your mother... is it at all possible that your feeling about her are actually some unresolved issue with your mother that you are transferring on to her?" Other people were very upset to learn that the records ELIZA kept to personalize its responses were available to the Computer Science department-they had been treating it like real therapy.) but psychology is human behavior. If a psychologist is predictable, to the degree that an early chatbot can be viewed as an effective psychologist by its clients, then on some level, the clients are predictable.

If the day comes, when someone has to judge which voice on the other end of an intercom is the human, and which is the machine, the human might fail by being too predictable... or too polite (say, for instance, by refusing to lose their temper) whereas the machine might have learned (and remember, ELIZA learned, it wasn't programmed with responses, it constructed them and rated how well they were received) when to get angry, when to misunderstand or say something inappropriate. What happens when the "judge" decides that that voice making the innuendo is the human, and the actual human has been so conditioned that they would never, ever, pick up on the most "obvious" opening "that any real human would?" It gets much more potentially subtle... we don't need to actually be conscious of our patterns, to be trapped in them.

Humans also have an inborn prejudice as to what is "human," we often think either better, or worse, of our own kind. It is worth pointing out that "King Chim," a famous ape, was often described by researchers as "more humaine than you would believe if I were to record it," for his kind treatment of researchers and other apes..."humaine" from the word "human" but referring to a degree of civility beyond what was considered the "human norm." We also ascribe "vengeance" only to sentience, but the basic moral question of "the Prisoner's Dilemma" is not actually all that hard to program for... so what happens when a machine understands "tit for tat," but the human "does the right thing," or would the judge expect that? At what point is the judge trapped in a Princess Bride-Iocaine Powder situation, and expecting a subject to do something contrary?