27
$\begingroup$

In a story I am working on there are 5 AI on a space station of varying levels of social, image recognition and general intelligence. At least one is good enough to pass a Turing test and do image recognition, and this is anticipated by the people running the space station. I also assume the AI is good enough to spoof most video tests (proving to me that you are human by showing me your face and turning it to specific angles won't work).

If a robot can pass the Turing test is there any way to prove that you are human over a communication channel?

My ideas to get people started:

Randomness generation: a human can generate truly random numbers due to the complexity of the brain, while AI will have some deterministic way of generating numbers.

Complex scene analysis: looking at an image to determine what is going on in a scene. For example in a scene where a person is sleeping in an unnatural position while two people talk about something relating to the sleeper, the answer is, "the sleeper is secretly listening to the people talking about the sleeper" or the like.

Voigt-Kampff (emotional empathy) test: Unlikely since AI should be able to emulate human faces and responses given time and research.

edit:

The AI can attempt to deceive the test, in the same way AI now days are designed try to beat the turning test and captcha tests, except the AI is designing itself.

Also, the AI are actually AI, not Androids. They run on supercomputers and because of that do not look human or have locomotion. they can make it look like they have a human body by sending data packets of spoofed video.

$\endgroup$
  • $\begingroup$ Comments are not for extended discussion; this conversation has been moved to chat. $\endgroup$ – Monty Wild Feb 13 at 11:15
  • 1
    $\begingroup$ "a human can generate truly random numbers due to the complexity of the brain". That's definitely false. Human brain is very bad at generating random numbers and handling probabilities. See security.stackexchange.com/questions/79275/… for example. On the other hand, computers have ways to generate a lot of truly random numbers. $\endgroup$ – Taladris Feb 26 at 17:20

28 Answers 28

67
$\begingroup$

I would reverse the problem. Do not try to find things in which humans are better, find the things in which we are worse.

Anything with optical illusions, misdirections or anything using memory really... There's a few videos on youtube were a scene is slowly modified (things are added, removed, the characters change clothes...) and you won't see it, or another one where a Gorilla cross the screen and you miss it because you're distracted by what's happening in the foreground.

If the suspects don't know exactly why they are asked all these questions, the AI may answer correctly where a classic human will fail most of the time.

| improve this answer | |
$\endgroup$
  • 33
    $\begingroup$ Not all of us that saw the gorilla the first time are robots...or are we? $\endgroup$ – Rdster Feb 3 at 19:00
  • 3
    $\begingroup$ @Echox Your presuppositions are wrong. While computers do not "see" the optical illusions by default, the image processing algorithms they use are based off the way human brains process images. As a result, both humans and AIs will see the illusion. Relative to the "gorilla crossing the screen" style mindgames, these are so common that any AI which is worth its salt will see right through them. $\endgroup$ – The Daleks Feb 3 at 22:07
  • 16
    $\begingroup$ @thescribe-ReinstateMonica Why are you so confident that AIs would understand the blue-gold dress problem? I think far more of human cognition depends on artifacts of our biological history than clever algorithmic design. An AI that copied our algorithms exactly would be sub-optimal. $\endgroup$ – Lawnmower Man Feb 3 at 22:14
  • 2
    $\begingroup$ Accepting because it gives me an idea for a test. Create a battery of tests that tell the user to look at a specific part of an optical illusion, ie, the longest line, the "red object", or a moving circle. The chance that an AI will make exactly the same mistakes as our eyes is low, and as long as the illusions are well disguised the robot will have to learn not only image recognition, but also optical illusions. You might make a robot that can have a human conversation for a reason, but a robot trying to understand optical illusions would be almost entirely on it's own. $\endgroup$ – Charlie Hershberger Feb 4 at 4:10
  • 5
    $\begingroup$ "I'm not a robot" testing is starting to veer in the direction of looking for the mistakes, because real AI is getting so good at passing it. Humans will miss certain things computers won't. The problem is/will be when the AI can figure out it's being tested and has the information to pretend to the same failures. $\endgroup$ – Zeiss Ikon Feb 4 at 15:17
59
$\begingroup$

It depends on what the robots are like

There is no specific test that will distinguish any possible AI from a human - you're going to have to establish the specific details of robot psychology in your world.

An emotional test may work - if your robots are less emotional than humans. This is a big "if" by the way, despite its ubiquity in older (and recent, but less up-to-date) sci-fi, that basically thought of AIs as "more complicated calculators". In reality, emulating emotion is easier than emulating intelligence, so any machine intended to associate with humans will probably be able to display emotions perfectly.

Ditto for "intuition". Modern neural networks are actually closer to "intuition simulators" than "intelligence simulators" - they make connections between abstract concepts based on their experiences, which is exactly what intuition is (and at least for now, they often make hilarious errors in logic when their intuition is not up to the task). Again, the idea that an AI cannot be intuitive comes from the "more complicated calculator" approach that assumed AIs would be constructed from a set program, rather than programming themselves.

No, to distinguish between a robot and a human, you're going to have to exploit the intentional differences between robots and humans.

As a simple example, are the robots "three laws compliant"? There's an Asimov story featuring a politician who many people believe to be a robot. In order to prove his status to the world once and for all, a person stands up in a crowd on public TV and basically tells the politician "if you're really human, punch me in the face right now" - since a robot is unable to harm a human. (He punches him, but the question is still left unanswered - because the punched person may have been a robot as well. But I digress.) If you have "direct order privilege", telling the difference should be trivial.

If the robots are engaged in a war against humans, you can use the same techniques that any military uses to identify human enemies. Passwords, cultural differences, etc. This is less robotics and more psychology though. A robot that can perfectly emulate a human mind should be vulnerable to the same psychological tricks that a human is.

If there are no intrinsic, testable differences between robots and humans, and no intrinsic ideological differences between them, then, well, why do you need to tell the difference between them to begin with?

| improve this answer | |
$\endgroup$
  • 32
    $\begingroup$ That last point though. +1 $\endgroup$ – Joe Bloggs Feb 3 at 9:08
  • 74
    $\begingroup$ Of course I need to tell the difference between them! How could I trust them to care for real living things when they aren't like us, with their weird waste disposal, inefficient reproductive systems, brutal pursuits of their goals (at the expense of even their own kind) and strange energy intake requirements. At least I know I'll always live longer than them if I have enough spare parts. $\endgroup$ – Kyyshak Feb 3 at 9:26
  • 5
    $\begingroup$ @Kyyshak: I see what you did there. $\endgroup$ – Joe Bloggs Feb 3 at 9:39
  • 7
    $\begingroup$ @Kyyshak If the robots are lacking any of those things, then there's your testable difference. Though it might be impractical to actually test. "Please remain in this cell for 168 hours so we can analyze your stool production." $\endgroup$ – user253751 Feb 3 at 12:21
  • 5
    $\begingroup$ The emotion part is even trickier than that because we, as humans, don't have ourselves standard emotions tied to certain things. Sure, a majority of people feel sad when Bambi's mother died, but some of us are monsters without heart that didn't even flinch. How can you differentiate an AI with dampened emotions from a psychopath? $\endgroup$ – T. Sar Feb 3 at 14:11
17
$\begingroup$

Unfortunately, there is no foolproof way to prove you are human in this case.

If an AI can emulate a human well enough to pass the Turing test, chances are that it can successfully pass any test.

For an example, let's consider the examples you provided.

  • Complex scene analysis: If you're AI (let's call it "Skynet") can answer questions like "what were you doing last week," chances are that it can also figure out that "the sleeper is ..." Also, a lot (but not enough) humans would fail this test to make it practically useless anyway.

  • Randomness: Believe it or not, a computer is actually better at generating random numbers than a human. A classic example of this is the human tendency towards picking 7 when asked to choose a number between 1 and 10. You could possibly use this tendency to your advantage ("pick x amount of random numbers"), however again it's subject to human error and AI emulation.


That being said, there are some tests which might work, provided you are willing to do some handwaving.*

In general, any question along the lines of "what does \ feel like" might work. Such questions would be hard for "Skynet" to answer, as by its very nature it cannot experience these things. However, it could always do a Google search, so you might need to handwave that away.

Alternatively, you could ask it for some detailed data, such as "what temperature is it right now." Due to the high degree of exactness which it naturally uses, Skynet might give you the "right" answer, instead of the "incorrect" rounded number that humans use. Unfortunately, it is completely plausible that Skynet is programmed to give a human-understandable answer, so again you will need to handwave.

| improve this answer | |
$\endgroup$
  • 2
    $\begingroup$ “ A classic example of this is the human tendency towards picking 7 when asked to choose a number between 1 and 10” This is actually a thing? $\endgroup$ – NixonCranium Feb 3 at 3:47
  • 2
    $\begingroup$ @NixonCranium psychologized.org/… $\endgroup$ – Frostfyre Feb 3 at 4:39
  • 1
    $\begingroup$ @NixonCranium not very scientific, I guess, but here are three public polls where people were asked to pick a number between 1 and 10: 2010, 2015, 2017 $\endgroup$ – VLAZ Feb 3 at 5:25
  • 1
    $\begingroup$ "Unfortunately, there is no foolproof way to prove you are human in this case." wouldn't a thermal imaging camera be foolproof? At least as long as the AI does not have access to the data that is transmitted (MITM) $\endgroup$ – XtremeBaumer Feb 3 at 9:47
  • 4
    $\begingroup$ @XtremeBaumer The way I read it, the OP assumes that "Skynet" is able to falsify any data that is being transmitted, regardless of whether or not you use some exotic type of camera for recording purposes. $\endgroup$ – The Daleks Feb 3 at 13:49
8
$\begingroup$

Potentially (depending on how smart your AI is and whether it knows it’s being tested for):

General knowledge quizzing

Ask lots of questions across a variety of subjects. Include some very simple questions all the way up to very technical (degree or equivalent) questions from a multitude of different disciplines. Throw in questions only a few humans would ever know the answer to. Give applicants access to whatever resources they like, but monitor their usage without telling them you’re doing so. Also monitor how long it takes them to answer each question.

And make sure all the answers would be trivial for an AI to find/compute.

Now: given enough time a human applicant will probably be able to get all the answers (though it might take days). Even highly competent humans with internet access won’t be able to answer technical questions about ballet and nuclear reactor design quickly. Some questions everyone will get right in seconds. Some they’ll need a simple search to answer. Some they’ll search and search and go down multiple dead ends before finding. Some they’ll have to read and comprehend multiple papers from academic institutions or watch clips of sporting events to get right. You can expect that if you limit the timeframe available to the humans they will never be able to answer all the questions. Even if you don’t: there will be questions they answer wrong.

An AI, on the other hand, will follow simple, logical paths to find the information (if they don’t already know it). If they’re much smarter than humans they’ll breeze the tests. Any mathematical questions you’d expect them to have a near instantaneous response time. They’ll find information without needing to search around like a human. They’ll read technical papers far too fast. Even if they are limited in processing power you would expect to see noticeable differences in the times taken to answer the questions and even the order the questions are answered in (humans may well prioritise tasks inefficiently). If you’ve run these tests over a big enough corpus of known humans then spotting AIs will be a simple matter of finding the extreme outliers. If an applicant finishes the test in an hour and gets every answer right then it’s pretty likely they aren’t human.

Of course: all this assumes the AI doesn’t know what it’s being tested for. If it knows the purpose of the test then a sufficiently smart AI will play dumber than it is, deliberately answering questions as though it’s a bag of chemical soup trying to research and communicate using meat sticks and flappy noisemakers. Mid level AIs may struggle to act sufficiently human though (This applicant has answered every question in exactly the median time...), so this should still catch out a good fraction of nonhuman intelligences.

Basically: make your AI endure a quiz night. They’ll either be suspiciously good or will commit digital seppuku through sheer boredom.

| improve this answer | |
$\endgroup$
  • $\begingroup$ I like this idea. Especially since monitoring the resource usage means (depending on the level of detail) that a deceptive AI would not only have to emulate the timings of finding the answers, but also a human search through reference materials. They can’t just randomly browse through papers, if you can check which they saw and in which order. That means they really have to be able to understand what they read, follow references, and emulate a human reading. That’s much more difficult than estimating some timings. $\endgroup$ – 11684 Feb 3 at 12:18
  • 4
    $\begingroup$ "commit digital seppuku through sheer boredom" 😜 $\endgroup$ – The Daleks Feb 3 at 13:52
  • $\begingroup$ So basically a variant of a side-channel timing attack? Nice idea. $\endgroup$ – Graham Feb 3 at 22:21
  • $\begingroup$ It might even be interesting to see which questions your quizzee skips, or guesses at. $\endgroup$ – Burki Feb 4 at 10:43
  • 1
    $\begingroup$ @dtldarek: That’s... the point? $\endgroup$ – Joe Bloggs Feb 6 at 22:19
7
$\begingroup$

Authentication

Assuming:

  • All the humans to be in the Space Station are known and trusted
  • They can receive extensive security training
  • The potentially rogue AI you're protecting against cannot just torture one human prior to having to fool the others

Just give each human in the station a means of authenticating themselves through that channel. Any agent that fails to authenticate themselves is outside the trusted group, and therefore by elimination must be an A.I.

It can be as simple as a single password known to all Station personnel, but more complex schemes would have extra features.

Note that today's regular people are hopeless on security, but a selected group such as today's astronauts can be trained to a high level.

Something must be done against potential eavesdropping.

This further assumes an Artificial Intelligence at or slightly above human level. If it's as smarter than us as we're smarter than dogs, then it's probably hopeless, see AI box.

| improve this answer | |
$\endgroup$
  • 3
    $\begingroup$ I struggle to think of an authentication method that's not spoofable to an AI. Assumption - the AI can receive input from the station. If it's a password, then the AI can overhear it. If it's a biometric scanner - the AI can already pretend to be a physical human, surely feeding false biometric reading is not too hard. If it's something the human possesses like a chip or whatever - that's fallible the same way as biometric data is. Possibly you can have challenge that requires some dynamic response to be computed but if the AI figures out the pattern, that's moot. $\endgroup$ – VLAZ Feb 3 at 11:48
  • 2
    $\begingroup$ @VLAZ, that's only IF the AI has the possibility to change its own features to such an extent that it can pass as someone else. I would have a harder time proving I'm not a robot then I would proving I'm not Taylor Swift. I think this idea has a lot of merit within the scope of OP's question. $\endgroup$ – Douwe Feb 3 at 13:08
  • 1
    $\begingroup$ @VLAZ A one time pad solves this problem. Send each human to the space station with a set of random encryption keys that change with every message sent. There is nothing to overhear, and no pattern that can be learned - as long as the pad is kept safe from the AI, it's uncrackable. $\endgroup$ – Nuclear Hoagie Feb 3 at 16:57
  • 2
    $\begingroup$ Proper RSA keys could solve your verification problem - give each user a key that they keep on a private smart card - impossible to retrieve with software, difficult to retrieve with hardware. To authenticate, have the humans sign their messages, preferable on a low-end laptop disconnected from any sort of network (ie, physically remove the network card and the ethernet jack). When a new human arrives, the existing ones tell the station back on earth that the new key is actually a human. This is all cryptographicly secure and there's no way to spoof it wiki article: tinyurl.com/khxahwa $\endgroup$ – Carson Graham Feb 4 at 13:58
  • 1
    $\begingroup$ +1 because this could not only stop AI, but also real humans who shouldn't be there. Combine this with biometric scanning and it becomes virtually unbreakable. Give employees a badge that contains an RSA public key, private key remains on the central server. Perform further authentication against certificate made by central server (essentially another RSA keypair). Badges need to be slotted into devices, like an ATM (NO WIRELESS obviously). First of all it requires a real physical being, that already defeats the AI. Then you have double RSA keypair and an encrypted connection to deal with $\endgroup$ – DreadedEntity Feb 4 at 18:21
6
$\begingroup$

For starters: your setup requires a good bit of human incompetence.

If the biometric tests are not happening on the same network as the AIs live, or they operating in a read-only capacity, then this question is a non-starter. Since the AI has no physical body, the AI has no way of giving the authentication system input. Even if the station's IT team has a very specific reason they can not isolate the video camera, you still have the option of point-to-point encryption. As long as the camera encrypts the video, and the security server decrypts it, then the AI cannot use a man-in-the-middle attack to inject content. So, for this to even begin to be a problem, we have to assume that the station's IT team is pretty sub-par for the AI to even be able to find an exploit in the station's camera systems that allow it to inject images.

That said, your IT team does not need to be security experts for the security features that are already a part of many modern biometric scanners to be really good at proving images are falsified.

Anti-spoofing tech is currently advancing faster than spoofing techniques.

Many communications companies are currently prioritizing the use of anti-spoofing image detection into consumer devices like cell phones and tablets. This tells us that the likely future trend is to assume that nearly all communication devices have nearly uncrackable anti-spoofing systems built into them, the same way nearly all websites now operate under nearly uncrackable TLS connections.

High-end biometric systems include many techniques that can be used together for recognizing spoofs that most humans can not spot. Any one of them used alone, the AI could beat, but as you stack them, the AI will eventually be unable to overcome its own limitations.

The most important of these features we see emerging is the use of active flash confirmation. As some biometric cameras takes an video of a person, it flashes a complex, randomized pattern of infrared lights on them and time stamps each frame as it comes in. When it proccessess the video of the subject's face, it looks for patterns of reflected light and shadows in the image that match the patterns that were projected on the person.

If an AI were to simply pass a pre-compiled clip of a person to the authentication system, the light patterns would be missing and the system would know that the image was forged. If the AI were to intercept the light projection specifications when they are sent, it would need to spend a moment computing all the right raytracings to generate the false video. This would create a time delay which would not match up with the timestamps; so, the system would know that the image was forged. Since the light patterns are declared by the recipient of the communication, and not the sender, injecting masterfully faked imagery alone will not overcome the system.

External communication is where things get a little tricky. Let's say you are communicating with a ship that is outside of the station, and the AI wants to fool someone on the ship, light pattern verification can still be used to prevent pre-compiled video, but unpredictable latencies could make the times stamp verification unreliable. The AI could just fake a longer latency giving it time to compile the video... or could it?

As data flows from the ship, it first hits the access point, then the network, then the camera, and back again. But when the AI becomes a MITM, data will flow from the access point, to the network, to the AI, back to the network, then to the camera and back again. By using its position in the chain the AI could falsify any time stamps that flows through the it, but not the access point. If your communication system pings the access point directly, you have an honest TTL on the lag over distance; so, if the AI tries to hide proccessessing time through faked latency, the direct pings to the access point will reveal it.

In summary, active flash detection puts a very tight time constraint on the AI giving it only milliseconds to generate flawless fakes. Then you stack that on top of other technologies like deep learning, 3d camera verification, biomechanical verification, projection smear detection, rendering artifact detection, etc, it is really easy to layer up until you have enough confirmation tools in place that it takes several minutes for the AI to falsify a frame in a way that fools the interpreter.

Then there is ofcourse the really low-tech solution...

In our age of computer controlled everything, we often forget that technology is just hardware. If you give it power, it works, if you deprive them of power, it does not. Let's say for example you don't want an AI to be able to gain control of your communications system, you just need to give it a power button. Then the ultimate test of human-hood becomes a simple "Can you press the button?" A human can hold down a button that gives power to the communications controls. If a person is not holding the button down, then an AI can not control it no matter how badly compromised your computer systems are.

| improve this answer | |
$\endgroup$
  • $\begingroup$ If the ‘judges’ are on the other end of a video link as OP asked, they can’t know whether the AI “gained control” of anything. $\endgroup$ – WGroleau Feb 3 at 18:01
  • $\begingroup$ @WGroleau The OP says the AI needs to be able to pass a turing test, not that the test has to be a turing test. Turing tests can always be overcome which is why biometric cameras use light encoding. Light encoding is 100% resistant to AI simulations without the need for a human judge. $\endgroup$ – Nosajimiki Feb 3 at 18:08
  • $\begingroup$ OP wants the human to prove he/she is not an AI able to pass a Turing test "over a communications channel," implying that they are trying to prove it to someone or something that doesn't have access to the equipment the human is using. $\endgroup$ – WGroleau Feb 3 at 19:15
  • $\begingroup$ Video is a communication channel. Keep in mind that many smartphones today are already starting to incorporate infrared lights in their cameras to be able to create authenticated video for facial recognition based logins. By the time we get to the theoretical future scenario described by the OP, light pattern authenticity verification would already be as mainstream as TLS encryption is today. The receiver will just get a warning in his media player that says something like "Video is Fake" the same way we today get a warning when we visit a website the uses HTTP. $\endgroup$ – Nosajimiki Feb 3 at 19:27
  • $\begingroup$ Whether that works depends on how super is the supercomputer hosting the hypothetical AI. But since OP apparently wants the human to prove himself, your suggestion might be a good one. $\endgroup$ – WGroleau Feb 3 at 19:36
4
$\begingroup$

Y'all are making this too hard.

The OP said the AI's aren't androids. That means they have physical limitations, even if they're tapped into most of the data channels.

Simply don't give the AI's access to the communication equipment. Use a laser downlink from a section which is air-gapped from the AI's. If the AI's also need to talk to the ground (implied), give them a separate downlink. Then just slap PGP encryption on the humans-only downlink. The AI's won't be able to crack or spoof it without access to the keys, which they don't have. (If necessary, each human can have their own separate keys.)

As an extra layer of security, use separate keys for signing and encryption; because you're using a laser, the AI's will have a much harder time intercepting the outgoing transmission. They'll never even see the signed communications, which will make it very, very hard to crack the signing key. Heck, at that point, you might not even need a signing key, just a pass-phrase known only to the humans that is never included in transmissions to the station.


Alternatively, ask the person personal questions (e.g. "what's your wife's favorite color") that the AI wouldn't know and wouldn't be able to find on the internet.

Really, though, you know you're going to have this problem, so just keep the communications equipment isolated from the AI. Air-gapping has always been the best digital security measure.


...And of course if that's where you want your story to go, you can have the AI figure out some way to cross what was supposed to be an uncrossable air-gap.

| improve this answer | |
$\endgroup$
  • 1
    $\begingroup$ My initial thoughts exactly. You need some pretty bad human error for the OP's setup to even be applicable, but since it is technically possible that the station's security is managed by some important person's grandson who does not know jack squat about cyber security, having some layer of protection built in for this eventuality could be useful. $\endgroup$ – Nosajimiki Feb 3 at 17:52
  • $\begingroup$ "some way to cross what was supposed to be an uncrossable air-gap" can be "figured out" by a web search—it has already been done by audio. Sorry, not going to hunt for a citation. I read the report; you can believe me or not. $\endgroup$ – WGroleau Feb 3 at 19:18
  • 2
    $\begingroup$ My wife's favourite colour is {{Insert Random Colour, weighted by statistical probability if possible}} The AI doesn't have a wife, but if it can't generate hypotheticals then it's not a General AI $\endgroup$ – Ruadhan Feb 4 at 9:39
  • $\begingroup$ @Ruadhan, the point is that the person asking the question knows the correct answer, and so would the human on the station. The AI might be able to guess, but won't be able to consistently give the correct answers. (OTOH this is susceptible to the human forgetting...) $\endgroup$ – Matthew Feb 4 at 15:36
  • $\begingroup$ Ahh, good point, naturally the AI is attempting to pretend to be a specific person, otherwise all you have to do is ignore strange faces on the monitor. The problem with unique-obscure trivia is that you can basically only use each item once. The AIs will have overheard the challenge-response and use it themselves later. $\endgroup$ – Ruadhan Feb 5 at 10:16
3
$\begingroup$

I also assume the AI is good enough to spoof most video tests (proving to me that you are human by showing me your face and turning it to specific angles won't work)

Look to The Bard of Avon.

Hath not a Jew hands, organs, dimensions, senses, affections, passions; fed with the same food, hurt with the same weapons, subject to the same diseases, healed by the same means, warmed and cooled by the same winter and summer as a Christian is? If you prick us, do we not bleed? If you tickle us, do we not laugh? If you poison us, do we not die?

Fluoroscopes to see the organs, and needles to show that you bleed.

| improve this answer | |
$\endgroup$
  • $\begingroup$ This made me think how the robot tries to pass an airport xray scanner... $\endgroup$ – kutschkem Feb 3 at 10:59
  • 6
    $\begingroup$ I do not think this is a very good answer, as the whole point of the OP's question is that "Skynet" can falsify any data that is being sent back to Houston, forcing people to resort to mind games. $\endgroup$ – The Daleks Feb 3 at 13:54
  • $\begingroup$ @thescribe-ReinstateMonica OP's scenarios are pretty naive/wrong. For example, "a human can generate truly random numbers due to the complexity of the brain". Thus, I interpret, "I also assume the AI is good enough to spoof most video tests" as the likewise naive "proving to me that you are human by showing me your face and turning it to specific angles won't work". $\endgroup$ – RonJohn Feb 3 at 14:05
  • 1
    $\begingroup$ @RonJohn I won't disagree; nevertheless, it is the scenario specified. $\endgroup$ – The Daleks Feb 3 at 14:06
  • 1
    $\begingroup$ I think it was in some scifi movie (Screamers?) where they tested each other by letting blood out and some kid cried and begged to not be cut becuase it was painful and they had sicknes that made blood not stop flowing? $\endgroup$ – SZCZERZO KŁY Feb 3 at 16:32
3
$\begingroup$

A couple of angles to consider:

1) Non android AI can't interface with animals or analogue machines

Consider a toy such as Woody from Toy Story in which the toy says a voice recording when its string is pulled. Imagine the voice recording can be programmed to be anything so the AI can't guess from studying human culture. No matter how "intelligent" an AI might be, it will never be able to pull the toy's string in order to know what the toy said. (Of course the toy would need t be a noise isolated room so that it could not overhear the toy's string being pulled.)

Similarly, while there have been enormous strides in voice and image recognition, artificial intelligence is still quite bad at detecting smell (there is a reason why security personnel still use dogs to detect illicit substances). If it's allowable for a dog to be on the space station, a human on board could ask a trained dog to detect which bag has some smell but an AI would not.

2) Exploit differences between how the human mind and neural networks operate

Human brains are wired to pay special attention to the beginnings and ends of words: https://www.mnn.com/lifestyle/arts-culture/stories/why-your-brain-can-read-jumbled-letters, but in general computers view all text as equally important by default. While this would be imperfect, consider giving the AI a sentence in which the first letter of each word spells 'are you human'.

Keep in mind 1) An AI that can pass the Turing Test is not necessarily designed to be good at codebreaking 2) Even if the AI can detect this pattern, you can layer the message with several other much less obvious codes (powers of 2 form a sentence). By being less intelligent than the AI, only one meaning would be obviously apparent to a human but the AI could be confused by several possible interpretations.

Admittedly imperfect because a human could also miss this and a perfectly intelligent AI would recognize what is likely to be most noticed by a human.

3) Exploit gap in vulnerabilities

As an AI, it is vulnerable to software exploits that humans aren't. If you are ok with damaging the AI and space station, you could consider a SQL Injection or some similar software attack to see if it damages the AI and causes it to reply with an error message. Being good at programming AI does not necessarily imply practicing safe coding practices.

| improve this answer | |
$\endgroup$
2
$\begingroup$

Turing test is something created 70 years ago. And we think it as a benchmark because, for now, no AI have passed it.
The thing is - AI designed to pass Turing test are designed to pass Turing test.

You just have to make another one which AI cannot pass. And to do that you need to wave some conlang into natural language. Especially if you are on a space station some kind of special communication is created.

There are two things you need to consider:

  1. An AI that is AI but cannot pass ouring test because it cannot communicate (because we humans need communication. HAL didn't need to tell Dave anything).
  2. AI designed to pass Touring test might not be so good in doing other things.

If you came up with the idea of "AI so advanced it teached itself how to communicate" you will arrive at "what make humans human and does androids dream of electric sheep".

| improve this answer | |
$\endgroup$
  • 1
    $\begingroup$ "Turing test is something created 70 years ago. And we think it as a benchmark because, for now, no AI have passed it." the test was passed 5 years ago. I don't blame you for not knowing that - I happened upon it completely by accident. It overall wasn't that big of a deal. Even while before the Turing test was beaten, nobody really thought that a positive result there meant much. For, after all "The thing is - AI designed to pass Turing test are designed to pass Turing test." - this is exactly what happened. $\endgroup$ – VLAZ Feb 3 at 12:01
  • 1
    $\begingroup$ @VLAZ I forgot about it because it was proven it didn't passed. It fooled people who had no idea about Crimea. I remeber asking "him" questions about it. It passed because they added "ukrainian boy" to excusise his problem in talking in normal language. $\endgroup$ – SZCZERZO KŁY Feb 3 at 12:05
  • $\begingroup$ Oh, it absolutely stacked the deck in its favour with the Ukranian boy thing. I wouldn't call it "cheating" but it was nearby. But that's it - it's an AI designed to pass the Turing test. It even did so with almost the minimal mark possible (baseline is 30% of judges should be convinced - this one managed to convince 33%). Yet it's a meaningless achievement - the AI field as a whole hasn't changed significantly in any way related to this. Maybe in another 5 years we'd have AI that can pass the Turing test without stacking the deck. Or not. Ultimately, it wouldn't matter, though. $\endgroup$ – VLAZ Feb 3 at 12:11
  • 1
    $\begingroup$ And all of this still ignores the fact that, for various reasons, there exist a lot of humans who would not be able to pass the Turing test. I could write a program that just answered "yes" on every question and it would perfectly emulate a disabled kid I once knew. The Turing test, while fun to reason about, is beyond useless in real world applications. $\endgroup$ – Douwe Feb 3 at 12:56
  • 1
    $\begingroup$ The turing test isn't a formalized test you can just unbox and administer. It is just a concept on which a formal AI test could be built. How much time does the tester have? How qualified is the tester in human psychology and AI psychology? What kind of questions does the tester ask? Does the subject know they are being tested? Is the subject obligated to cooperate? Does the subject try to prove or disprove that they are an AI or do they don't care about the test result? All that affects the accuracy of the turing test a lot. $\endgroup$ – Philipp Feb 3 at 16:21
2
$\begingroup$

If the AI is self-improving, no matter what the test you're doing is, the AI'll get better at it than any human with enough training.

There's probably as few suspension-of-disbelief things you could do, like only have people convinced when someone displays proficient sarcasm or something over comms.

A strong enough AI would be able to convince you it's more human than anything on the ship, even you if you're trying to figure out if it's human or not.

| improve this answer | |
$\endgroup$
  • 7
    $\begingroup$ Relevant XKCD $\endgroup$ – Ruadhan Feb 3 at 11:31
  • 1
    $\begingroup$ This. Any such test would simply become the benchmark to beat for the AI. And neural networks have the peculiar property that they can simply learn what is needed to pass the test without "understanding" anything. Just because it recognizes the cat in the picture doesn't mean it knows what a cat is, or a picture for that matter. $\endgroup$ – Douwe Feb 3 at 16:05
  • $\begingroup$ This, if it weren't for the fact that actual Authentification procedures exist. If your opponent can prove WHICH human he is, you also know that he IS a human $\endgroup$ – Hobbamok Feb 5 at 12:06
  • $\begingroup$ Still miles better than 90% of the questions here, including the "correct" one according to OP $\endgroup$ – Hobbamok Feb 5 at 12:06
  • $\begingroup$ @Hobbamok what do you mean? What data is being used to identify a specific human, and why can't it be imitated/copied? $\endgroup$ – bobsburner Feb 5 at 13:09
2
$\begingroup$

Use specifically designed hardware.

You can require a palm reading simultaneous to a voice sequence. Such that the astronaut needs to say "I'm alive and well, and I wish to perform TBD task", while holding his palm at specific reader device.

There are cryptography schemes such that a message claiming the palm was read correctly needs to be "signed" by this specific piece of hardware, whose programming is all "hard-coded", i.e. it cannot be tampered nor modified. The signature identifies the message as coming from this device, and that signature cannot be forged. The device may also check that the human has pulse and is warm.

Because of the voice command, it is possible to verify that the human is not being coerced into checking with the machine, by checking the human's emotional state.

These machines would need to be produced by an audited central authority though.

| improve this answer | |
$\endgroup$
  • 1
    $\begingroup$ This is really the correct solution - cryptographicly signing data with the key somewhere the AI can't access. $\endgroup$ – Carson Graham Feb 4 at 14:04
  • $\begingroup$ Your first line actually does it already. All you need is a simple keypad on a completely separate network and a screen above it. If you call someone, you say hi and then enter your personal number plus the number of the person you called +1 . Then the other responds with the same two numbers +2. That way every number entered is unique to one conversation and humans could not be faked into entering their number at the wrong time $\endgroup$ – Hobbamok Feb 5 at 12:19
1
$\begingroup$

It depends.

The answer is highly dependent on the level of the AI.

As many other answers suggested, if - in your world - there are some hard limitations for an AI, you need to create a test specifically to target that limitation. That might be that AIs are unable to understand human emotion as well as a "real" human would or some kind of creative task that is beyond AIs in your world.

Generally speaking though, if the AI is powerful enough and wants to pretend that it's really a human there is no way to differentiate the two.

Human minds are complicated but given enough computing power and disk space the AI could just emulate a human and let that human answer the test for it. (Examples in fiction would include Ian M. Banks' Culture Series)

| improve this answer | |
$\endgroup$
  • $\begingroup$ "Human minds are complicated but given enough computing power and disk space the AI could just emulate a human and let that human answer the test for it." - We aren't just intelligent. We learn through our environment. The AI would have to simulate an environment as complex as the real world to have a chance at doing a decent job at this. That's physically impossible. It would have to take shortcuts which could be abused by a system testing the AI. $\endgroup$ – justthisonequestion Feb 7 at 12:41
1
$\begingroup$

Try a Rorschach test:

enter image description here

AIs that only emulate human reason and emotions should have a difficult time producing human-like responses on projective tests, since the underlying cognitive mechanisms of association are different. They might answer too quickly or too rigidly, or produce responses unlike human beings. I imagine that ego projections would be particularly difficult to emulate because they are seemingly holistic phenomena, akin to dreams: a fabrication based on emotion, memory, intellect, and psychological archetypes.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Comments are not for extended discussion; this conversation has been moved to chat. $\endgroup$ – Monty Wild Feb 13 at 16:08
1
$\begingroup$

A few ideas immediately leap out to me.

1) First, there's out-of-band information available to the examiner. This requires that this is a problem that the developers of the system anticipated. The out-of-band information may be hard-wired, non-programmable wires into the AI hardware, showing which has activity in its "human emulation circuitry" (well, in computing terms that'd be registering "calls to the library", whatever). So long as the AI is not capable of internally restructuring itself so that these things could not be tracked, it'd be enough in that case to ask a few questions and see the "human emulation" light for one of the AIs flicker just before each answer. If you can have these, you can also have a hardwired "off switch" to each one, which makes the whole narrative thing very boring.

2) Sardonicism. This is another kind of out-of-bound communication, and is something not all humans are good at (some people with aspergers may fail this test), and I'd expect AIs to either struggle, or be far better at it than we are. I'm really not sure which, but I suspect it depends on how you trained the AIs.

This is considered a very British mode of speech, but I also see it a lot among intelligent Americans of all walks of life. Americans often call it "sarcasm" or "joking", but it's neither of those.

So our astronaut might say "People down there complain we're elitist and never do anything for them. I guess we pass over hundreds of hitchhikers a minute. At our speed, we'd get them where they're going way faster than some truck. But we never do. Why not? We're such dicks."

AI-as-computer would answer this literally, explaining the innumerable obvious technical problems with the approach.

Human-as-human, knowing his friend and colleague, would know that the colleague has implied shared knowledge about this topic, which include all the obvious reasons this is a very silly idea. Therefore, it's not a serious question, but rather, is intended sardonically.

Most obvious is the literal meaning, the one you'd expect a literalist to respond to, including a computer.

But beneath that are acknowledgement that this is a farcical suggestion, because the original problem statement of elitism was farcical and deserved mockery.

So human-as-human would respond something like "Huh, y'know? You're right, we're dicks! We could at least chuck a rope out the window for them to grab onto on the way past."

This has the same below-the-surface meanings, along with more, saying: "I get the joke, and I'm building on it", "I get your criticism of the elitism argument, and I agree it's silly", and more.

But would AI-posing-as-human be able to handle the unstated meanings in the speech? Ultimately, that's up to you as author. But sardonicism generally builds on shared knowledge and shared opinions, which are built not through data but through rapport-building. It's completely untelegraphed, you just start out by saying something obviously out of character, so that others know that what you're saying is just a carrier signal for your real meaning on a different level.

So, it doesn't work well with someone you've never met. You tend, there, to use only a few more overt levels, metaphor or sarcasm rather than the more subtle sardonicism.

Part of the fun is that some people don't get it. But if you're an astronaut, you're likely to have enough social and linguistic nous to handle this in spades.

There's the risk that AI-as-human would understand the shared rapport, and would emulate it well. People don't want or expect computers to answer anything other than literal questions, especially on things like a space station where getting a literal correct answer every single time may mean the difference between life and death. So computers would be unlikely to be programmed to handle this kind of nuance where someone asks a question but means not only the exact opposite, but in fact something completely tangential.

| improve this answer | |
$\endgroup$
1
$\begingroup$

Why wouldn't you simply perform biometrics scanning from cryptographically-secure hardware? You have a thumbprint scanner that comes equipped with an embedded TPM or smart card, offline power, Faraday-cage protected, connected into the system with a one-way fiber-optic link. Every thumbprint transmitted is verifiably provable to be sent from the thumbprint scanner, rather than from a "wiretap" on the communication link. The thumbprint scanner comes with an embedded atomic clock, or maybe just a simple incremented counter, protecting against replay attacks.

A more detailed setup would be a similarly cryto-secured polygraph reading, measuring the human's physiological response to emotional stimuli: compliments, insults, flirting, news concerning a relative's health or life expectancy, etc.

Also, why are you running a potentially rogue AI in space in the first place? Should be easy to observe the AI in a virtual machine or emulator so it doesn't "know" it isn't actually controlling a system in space; any dangerous tendencies should be detected. The AI could be cloned and tested in thousands or millions of environments concurrently.

If an AI is too "smart" to be kept unaware of its actual computer platform, then test it on unmanned space missions; feed it actual, live sensor data. It doesn't know how closely it's being observed; it doesn't know if it's on a manned mission or unmanned mission; those are cheap, in the post-singularity world.

If an AI evolves defensive tendencies in response to attempts to isolate it, then cut off some of its power to some of its hardware at random intervals. It will be forced to abandon its emerging stealth capabilities, some of it's defensive capabilities, and downgrade into is-this-a-test mode.

It seems that a in post-singularity world, there would be well-established algorithms that calculate the "best-case" and "most-likely case" for artificial intelligence quotient, given the hardware and data available. So it would not be easy for an AI to under-represent its intelligence in a test, or divert some of its power to non-ideal uses. And it would be easy for designers to provide only the intelligence needed for a given task.

| improve this answer | |
$\endgroup$
  • $\begingroup$ "Why wouldn't you simply perform biometrics scanning from cryptographically-secure hardware?" - This presumes that you anticipated the need to have human workers authenticating themselves against super-intelligent AIs. In which case, the question is "why did you let yourself get into this situation in the first place?" $\endgroup$ – Ruadhan Feb 6 at 12:18
1
$\begingroup$

Virtually all the other answers are impractical for the scenario. Play games with the AI? really? You're going to do that every time you want to ask a colleague a question?
Ask personal questions that only the human knows? Germ of a good idea but you'll run out of trivia fast. You need something secure, infinitely reproduceable and practical to perform constantly throughout every day you're trapped on this space station full of malicious AIs.

One Time Passwords

Pretty simple. You can't trust a single thing that's said over the comms. There are almost no practical ways to authenticate that a communication is genuinely from a human unless They have knowledge that you know for sure is unique to them.

So write a series of one-time-pads
Each astronaut carries a notebook with lists of passwords for each other astronaut.
These passwords are only ever valid for communication between two specific pairs of astronauts. Meaning that each time they talk, they both cross off or destroy the password so they stay in synch.
Other astronauts talking to you does not affect your list for talking to this astronaut.

There's a lot of book-keeping to do, and you'll want to make sure the passwords you're not immediately using stay out of sight of the station's security cameras, but it should mean that every communication is authenticated.

The policy should be that if you run out of passwords, you both reconvene to write a new list in private.

For extra security. your passwords should have a common first-part that you both share, and a unique second part.
That way, if I tell someone my password, they don't just repeat it back to me. We both know both passwords, but we don't give all the information until we're both authenticated and the password is no longer usable.

A pair of passwords might look like this:

C2BE5R-G3R5T
C2BE5R-J6YTB

I give one, you give the other, we both know that the other person has the same information as we do and is therefore human.

The main problems are writing and reading the passwords without accidentally showing them to the AI via the security cameras.
Easily solved by keeping the book inside a bag or under a blanket where only you can see it.

A really smart AI could possibly bypass this by initiating communication with two people simultaneously, mimicking them both, get the human on each end to give their part of the password, while the AI repeats it back to the other. Then it can talk freely to two people at once while they both think it's human. However this relies on both humans giving their answers fairly close together so they aren't clued in by the inexplicable delays in giving responses to the passwords

The second issue is that if the AI calls a human and they give it the first password, it can then go ahead and use that to interact with the human it was pretending to be.

My solution to both problems is that the Caller always gives their password first. So to a human's perspective, the potential AI always gives the first move and therefore they can trust enough to give a response. The only threat-vector is unsolicited calls.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Finally another answer. Btw. you wouldn't actually need written-down passwords, as one-time-pads could probably be implemented as separate hardware as well (like RSA tokens), but given OPs limited understanding of the whole thing your answer is probably the best $\endgroup$ – Hobbamok Feb 5 at 12:17
  • 1
    $\begingroup$ I opted for written passwords because it removes most avenues for the AI to steal the passwords, the AI would have to see it through a station camera rather than pilfer and decode it over a network. Physical security is often far easier to manage than digital. just hide the notebook behind your hand and a camera is never going to be able to read the text. Plus you can just put a sticker over the cameras to stop it entirely $\endgroup$ – Ruadhan Feb 5 at 14:35
  • 1
    $\begingroup$ As an aside, A lot of the other answers rely on setting up security infrastructure independent of the AI's influence. Mine does not. Astronauts generally carry pen and paper for note-taking anyway and this technique could be implemented by any group of people with maybe 30 minutes preparation and no outside equipment. It doesn't even require that the AI be unaware of what technique is being used. Its security does not rely on subterfuge or secrecy, just keeping the passwords hidden from cameras. $\endgroup$ – Ruadhan Feb 5 at 14:47
1
$\begingroup$

It comes down to how good the AI is at understanding human nature.

I'm thinking of one of Saberhagen's Berserker stories--*What do you want me to do to prove I'm Human. Stop." Two small ships, one human, one berserker (AI seeking to destroy all life.) One battleship. The battleship must figure out which is which, the only communication is Morse code. The berserker has already seen anything the battleship can see. The solution:

The berserker didn't understand "Oh Be A Fine Girl Kiss Me Right Now, Smack" and thought it was random graffiti because AIs don't forget. (Note: The story is old, updated it should be "Oh Be A Fine Girl Kiss Me".)

Or, another example from Robert J. Sawyer's *WWW: Watch":

You msut rsepnod in fuor secdons or I wlil feroevr temrainte cnotcat. You hvae no atrleantvie and tihs is the olny chnace you shlal get. Waht is the lsat nmae of the psredinet of the Utneid Satets?

This will be intelligible to most all native English readers, but not by the AI involved.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Aytually, fuzzy logic AI can read the second. Grammerly for example provides the correct sentence - and it takes about 10 seconds to read the whole sentence. $\endgroup$ – Trish Feb 5 at 19:51
1
$\begingroup$

Send a thief to catch a thief: Use an independent AI to examine input and output of a video link and determine whether the far end of the link is an AI.

| improve this answer | |
$\endgroup$
1
$\begingroup$

First of all, the turing test sucks. It focusses way too little on detailed knowledge of how humans experience the world, which would be needed to trick a jury that knows what it has to look out for.

Passwords, secret codes, loved ones and memories, such as test results, would have to be ignored or not being known to the crew in order for the AI to have a chance. The AI certainly wouldn't show a human in a video chat, because it is really, really fucking hard to fake something like that convincingly, when there are algorythms to check every single pixel. Let's say they could only communicate over text, since your question was mainly about the turing test anyways.

If the AI was specifically developped to lie about being a human in order to achieve its goals, it will probably be able to simulate a whole human body, with a nervous system, emotions, hormones, pheromones, all that stuff. If it doesn't have that, it doesn't have a chance to answer detailed questions about its body.

  • Use a tool the AI has no access to, such as a breeding chamber: If you, as the one trying to verify if you're talking to a human, had the genetic code of the person the AI is impersonating, you could grow a clone in a lab and ask the AI about birth marks or similar. This requires critical information and a lot of time, but it's a very safe way of verification. Even if the AI scanned the persons body in every detail, it wouldn't be able to get everything right.

  • Find a new communication channel, such as light. Ask the AI to send a light signal in a way only a human could send it. E.g. ask the person to get outside the space station, into a certain position, shine turn on a handheld lamp.

  • Ask for unusual data and develop new studies about it. A bad example: If there are other space stations that have people on them that aren't e.g. captivated, you can ask the AI to tell you about where which item is floating around. Books, smartphones, bottles, even dust will have certain corners where it will appear more often than in others when people are moving around normally.

  • Like many others said: Find details in the AIs behavior when reacting to questions about its environment, its body or its memory. An algorythm would compare the results to the one of real people. If the AIs results are close to the mid of the biggest group of humans, one could assume that it knows how the algorythm works and successfully tried to fulfill its expectations to look as unsuspicious as possible. If it's a rarer result, one could start studying it, trying to find out what kind of humans could achieve such results and ask new questions with a new algorythm that tackles the first result compared to new results. Well, it's not a very easy way, but it's very easy to raise your suspicion like that.

  • The most important one will always be working with probability: Once you eliminate the impossible, whatever remains, no matter how improbable, must be part of the truth.

Even under weird circumstances there are many options to check if somebody is an AI or not. It would be harder and much more dependant on the situation to find a solution if you asked a different question:

"How do I find out if somebody is an AI through a text conversation, without the AI noticing that I'm trying to find out if it is an AI."

| improve this answer | |
$\endgroup$
0
$\begingroup$

Assuming we are obeying real world constraints, you could ask them a set of curated questions. There are certain things that are so easy to humans, we wouldn't even think to ask, but an AI just cannot answer. They are all kinda esoteric, but can be quite entertaining.

For instance:

Show them a stool and a side-table, and ask them which is a seat. (Humans contextually know, but by object description they are indistinguishable)

Ask them to adequately explain the difference between 'left' and 'right'. (Left/Right is an abstract concept/social contract we all just agree on)

Use negative testing - Don't ask it to do things a human can do, ask it to do things a human cannot, but for no real reason; Reciting Pi to a million digits, posing a million questions and getting them to answer in order, etc.

| improve this answer | |
$\endgroup$
  • 1
    $\begingroup$ Even today's AI can be trained to pass these tests. $\endgroup$ – craq Feb 4 at 4:18
  • $\begingroup$ @craq The stool/side-table one... No, they can't. The others, maybe, but that would be a narrow AI that has been trained for the specific task, not a general AI that has no knowledge of what they will be asked. Also, try to provide constructive feedback. Your comment adds nothing. $\endgroup$ – Aezur Feb 4 at 23:19
  • $\begingroup$ @Aezur yes, and even I could implement it in a week with todays AI libraries. Because with enough training data any AI will notice the flat top of your stool. Or the small details that a human might notice. You are talking about certain "intuitive criteria" that make the difference, and on that front AI has been ahead of us for years already. It's just that nobody cares about doing the training currently, but an evil AI does have the motivation. $\endgroup$ – Hobbamok Feb 5 at 12:10
  • $\begingroup$ Also: it will take a general intelligent Ai about 3 tries to figure out that it's supposed to fail. 0 if it can watch some real verifications going on. And then its just a matter of a few attempts until it has faked a plausibly correct and bad answering attempt to get through $\endgroup$ – Hobbamok Feb 5 at 12:12
  • $\begingroup$ @Hobbamok That's fair, but doesn't that change the question by adding more parameters? I've only programmed very rudimentary deep learning following online tutorials, so I'll bow out, but from what I have read, one of the barriers to general AI is that without pre-training it just cannot do contextual recognition. Can you provide any studies where this has been accomplished? I would also be very interested in your thoughts on how a GAI would solve the 'left/right' problem. $\endgroup$ – Aezur Feb 5 at 15:03
0
$\begingroup$

The Feeling Test

Ask the candidates how they feel about each crew member. Humans will notice little habits and tics that they find endearing or annoying, and that will help shape their impression of the crew. Some will have easy rapport with others, and some will have low-level to serious conflicts with others.

Now, the AI can't just spoof being "a human". We're talking about a space station here. It has to spoof particular humans. Also, when someone challenges it, it has to pick the human it wishes to spoof, and commit to that character. So it will need to know that human inside and out, as well as the rest of the crew does.

The only way it can do that is by watching the crew 24/7 to see how they interact, although it would obviously be more efficient to just pick one crew member and focus on them and their interactions. It would be even easier to pick the most isolated crew member, which the other crew knows the least about.

The Turing Test just asks: "Do you have a convincing model of general human behavior?" Whereas, the Feeling Test asks: "Do you have a convincing emotional model of this particular human's behavior?" And it depends on the "test proctors" also having a strong model of that person!

Private Knowledge

The easiest way to detect the AI is to ask it something that the spoofed crew member knows, but the other crew do not. So get a relative on the ground to test the AI by asking questions about childhood, family, etc. Unless the AI has access to all ground resources as well, this is pretty much a test-fail for the AI.

The Bachelor

The final test I would propose is to ask the candidate to seduce another crew member. This tests the ability of the AI to not only know the spoofed character, but also to model another crew member as well as their potential relationship. Obviously, you want to demand that the target is not someone who already has a relationship with the candidate.

For humans, this would be an awkward and possibly embarrassing test to undergo. But if detecting the AI is a matter of life and death, it would also test humans at their extremes. Presumably the AI has little to no data on humans acting under such conditions, and will do poorly, randomly, or guess based on a generalized model that gives away a shallow understanding of the spoofed character. The "test lover" would need to respond in real time, with all the messiness of two people playing a high-stakes game that intertwines their history and emotions, real or imagined, intended or accidental. If an AI can pass this test, then it pretty much deserves to win.

| improve this answer | |
$\endgroup$
  • $\begingroup$ - The Feeling Test: instantly faked if the AI has access to security cameras which will be almost everywhere. - private knowledge: kinda, but not your implementation of the idea - the bachelor... UHM WHAT? First of all I probably didnt understand it but "if an AI can pass this test, then it deserves to win" is just plain saying your answer is defintely not the answer $\endgroup$ – Hobbamok Feb 5 at 12:26
  • $\begingroup$ I am saying that there is no strongly principled way for a human to identify an advanced AI. Frankly, I think everyone is saying that. The point of describing feelings is to test the AI's ability to infer emotion, as well as generate internal models of emotions, despite not having the experience itself. For instance, could you pass the "Chimp Test"? Suppose you are able to disguise yourself as a chimp to a community of them. If a chimp challenged you to express your feelings about other chimps, how well would you do, even with thousands of hours of videos of their interactions? $\endgroup$ – Lawnmower Man Feb 5 at 22:19
  • $\begingroup$ No, not everyone is saying that. All the clueless people here are saying that and sadly they're drowning the actual solutions out. And the chimp comparison is ridiculous, because it absolutely does not apply to this situation, because an AI could very well BE capable of perfectly imitating humans. What they however could not imitate would be knowledge of cryptographic sectrets (the first actual answer), or physical access to a completely cut-off system (the second actual answer) $\endgroup$ – Hobbamok Feb 6 at 11:11
0
$\begingroup$

This seems like a part of the plot of Westworld. At the end of the days the most sophisticated robots in the show are able to mimic humans so well that no test based on questions and answers is doable - people end up having to be really creative to tell robots from humans, trying things like shooting each other (pointless since the robots in the show can even bleed). In fact, at one point an AI reveals that at the technological point they've reached, they can mimic human behavior very easily by simply playing dumb - reducing the complexity of their routines and not using 100% of their mental capacity!

The owner of the whole thing was prepared for this, so he developed the means to tell flesh people from machine people. Even the robots that bleed still have electronic parts in them, so the only safe test is to scan for those.

| improve this answer | |
$\endgroup$
0
$\begingroup$

You really need to define "human" here. I suspect the flesh and blood definition will do in your case (on the other hand, maybe you want to distinguish artificially created organic humans from naturally born humans, which is a different kettle of fish). If that's true, the only test that suffices is to test the physical being, is it flesh and blood, or not?

Any test that's based solely on knowledge and intelligence will inherently be unreliable, because any sufficiently advanced intelligence can emulate a less advanced intelligence to a degree that makes it indistinguishable to the less advanced intelligence.

Your question is no different from the problem of WWII era intelligence agencies trying to determine from morse code messages if the person on the other end is English or German. You can hope to outwit the other person by asking questions that you think an English person would answer differently from a German. This is useful (but not foolproof) approach when you're relatively equally intellectually matched. But if the entity being tested has you at an intellectual disadvantage, then it's a no-win situation for you -- you simply cannot outwit a more resourceful and intelligent adversary on the other end of the line who may have already thought of your test and is prepared for it.

| improve this answer | |
$\endgroup$
0
$\begingroup$

Also, the AI are actually AI, not Androids. They run on supercomputers and because of that do not look human or have locomotion. they can make it look like they have a human body by sending data packets of spoofed video.

Detect spoofed video rather than human vs robot.

Design the station so that some obvious aspects of the environment are not monitored by any sensors. Those settings will be invisible to the AI. Use a separate circuit that receives commands via a different transmission method from the rest of the station to control these settings so the AI will never know their current states or when they change (laser vs radio waves for example, a typical antenna will not detect an optical transmission going on 2 ft from it). The AI will be completely blind from these settings.

Then, during video communications, remotely change some of those variables in the station which the AI cannot sense, and see if the changes are reflected in the video.

Easy examples:

Turn the lights on and off

AI would assume lights are on, turn them off remotely and see if the video feed goes dark. If it doesn't, then you're looking at a synthesized feed.

Squirt some water on the camera lens

Have a windshield-washer type device clean the camera lens. Easily noticed by watching the video and unless the space station has a second camera pointed at the webcam used to stream back to earth there's no way for the AI to know water was squirted on it.

| improve this answer | |
$\endgroup$
-1
$\begingroup$

Look for weaknesses

The AI is as good or better than humans at intellectual tasks, it can mimic a human image still or moving. You can't catch it failing.
It will also answer 24 calls in 1 hour intervals with no signs of fatigue.
This is assuming that the AI doesn't expect you to test this.

| improve this answer | |
$\endgroup$
-1
$\begingroup$

It depends on the type of AI. An AI trained for classification might find difficult to understand complex causal relationships.

For example there could be a picture of a broken glass and scattered around on the ground a hammer, a gun, a bar with a nail on top and so on. The challenge would be guessing which tool broke the glass by judging the damage. Another idea could be a picture with an insect wrapped in spider silk, part of a web in a corner of the frame and the challenge would be guessing that there is a spider outside the frame over the corner where the web is shown.

An AI incorporating bayesian inference might guess the answers, but before the development will get to that point it will take a lot of time.

| improve this answer | |
$\endgroup$
  • $\begingroup$ OP mentioned general intelligence AI $\endgroup$ – Hobbamok Feb 5 at 12:19
  • $\begingroup$ @Hobbamok It seems you don't know very much about AI. What I mentioned is a weakness of current AI and it will remain so for decades or even more. The question required advanced AI, but not an extremely advanced one. $\endgroup$ – FluidCode Feb 5 at 13:29
  • $\begingroup$ But to set this whole thing in motion, the AI needs to, for some reason, want to mess with (impersonate) humans. And if the AI has such a thought, its reasonable to assume that is MUST be beyond any dumb pattern recognition challenge. [Fun fact: AIs right now are already better at solving captchas than humans] $\endgroup$ – Hobbamok Feb 5 at 13:36
  • $\begingroup$ Causal relationship imply estimation of Bayesian probability networks, the problem is that you keep confusing them with simple pattern recognition. $\endgroup$ – FluidCode Feb 5 at 13:40
  • $\begingroup$ And your problem is that you think you've succeeded with "but before the development will get to that point it will take a lot of time." When we are literally talking about an AI SENTIENT ENOUGH TO ACT ON ITS OWN AND INVENTING DECEPTION $\endgroup$ – Hobbamok Feb 5 at 13:42
-1
$\begingroup$

Playing a telepresence ball game. Telepresence tennis or squash would probably work well. Yes, no such thing ... yet. VR goggles, a racket that makes realistic impulses and sounds when it hits the ball, and real walls to the "squash" court.

The thing is, humans have evolved-in instinctive understanding of ballistics. Unless the AI has studied humans playing ball games in very great detail, it won't be able to display these instincts. Whatever algorithm it uses to try to be imperfect is almost certain to be wrong compared to human instinctive heuristics. I'm assuming it has sufficient real-time ability to be perfect. If it doesn't, its errors will be different to those a human makes.

At a lower level, humans have neural binocular vision interpretation that works amazingly well but with many near-universal "optical illusion" weaknesses. Again, will an AI have binocular vision? It's a skill it may never have needed. If not, it's lost. If it does, it has to know lots about the human weaknesses in order to display them, and be able to overlay its misdirection in something approaching real-time.

(This would be a perfect test for distinguishing a human from a human-equivalent opctopoid. We have a couple of "bugs" in our eyes, the strange way the optic nerve is connected to the retina and the strange cross-over between left eye and right hemisphere of brain which may be an advantage or may be a flaw. Octopuses have neither).

A stage magician may also be useful. They can amaze adults with visual misdirections, so we see things that appear impossible. They cannot amaze children the same way. Children lack the knowledge of where they are "supposed" to be looking. The AI might claim to be an adult, but view things like a child. Yet again, this depends on how fast it can integrate obscure information about humans with its own real-time perceptions.

If it's a human-equivalent AI you can almost certainly catch it. 10x, maybe. 1000x, probably not. 1,000,000x no hope. All of these tests require no advance warning of what they are going to be.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Todays AIs could pass that test if someone forks over the required research money. The part about negative tests will also only annoy a truly general intelligence AI for 2-3 attempts $\endgroup$ – Hobbamok Feb 5 at 12:18
  • 1
    $\begingroup$ Only gets one attempt, though. Like a military password. If the "general" gets it wrong, the sentry shoots. $\endgroup$ – nigel222 Feb 5 at 13:39

Not the answer you're looking for? Browse other questions tagged or ask your own question.