For those (readers) who have seen the blockbuster movie titled I, Robot (2004 film) there is a humanoid robot who prefers to be called Sonny: Clearly this implies that Sonny is different from the others. Anyway, for those who have not seen it Sonny is being investigated by Detective Spooner (who is a human) for causing the death of its creator/designer.

Things get complicated when people assume different identities. We tend to become hostile to those who appear different from us in aspects like language, nationality, race etc.

Case study A for human: A female child grows up thinking that she is a male trapped inside a girl's body.

Case study B for robot: A robot suddenly thinks that it is a human whose conscious is trapped inside a doll.

Q1: Is it possible for a robot to have identity crisis like the description in case study B?

Q2: Suppose we isolate the robot on the moon so that there is no interaction with any live human beings, is it possible that it will soon question its ability to explore the environment and even its own identity?

  • 4
    $\begingroup$ There's no rule that says it can't. Since we've not yet created any Strong AI's you're free to come up with your own ideas about how their minds might world. Nobody can say that an AI couldn't suffer some of the same mental problems as a human. $\endgroup$ – Murphy Oct 7 '15 at 11:11

Most ideas for Hard AI (human level learning AI) involve some form of genetic programing or other 'growth' algorithm. These algorithms could lead to identity crises. But first let me explain a little about AIs in general, keep in mind I'm over simplifying some details here:

Soft AI

Right now our 'AI' that we write is mostly soft AI, and it is not really intelligent. It simply has a list of rules that says things like "if x then do Y unless Z", with a bunch of probabilities, random number generation, and game theory thrown in. Still, the point is that right now with soft AI we can only do things we explicitly programed. If I never programed a rule for what to do if situation A comes up my AI will never properly handle that situation, no matter how often it comes up.

The limit here is obvious, this AI can't really learn or grow. I can write rules that allow some limited form of quasi learning, like "see how many times X happens if you do Y and Z, and do either Y or Z depending on whichever caused X to happen more often", which allows my AI to 'learn' that either Y or Z is more effective if I want X; but this is a very limited form of 'learning' that had to be programed in. Humans simply can not program in the kind of rules required for truly intelligent hard AI, there are too many situations to write even if we could think of them all.


Hard AI is more like the AI you read about in science fiction, intelligence that thinks and learns like humans. In sure actual intelligence, not a bunch of rules faking it.

As I said the sort of rules based approach we currently use for Soft AI don't seem capable of producing this. Instead were looking at tricks like genetic programing and 'growing' of an AI for this....

Genetic Programing

Genetic programing is a simple idea in general, though it should really be called evolutionary programing, since it's inspired by evolution that drove our genetics. If you know how evolution works the basic idea is we have blue prints (genes) that occasionally get random changes (mutations). Most of these mutations were bad and killed those who got them, but a very small number were good and lead to stronger humans which passed on there genes. This basic concept evolved extremely complex beings like humans, something humans could never design on our own.

We use a similar idea in programing. Start out with a very basic 'program' and intentionally add random 'mutations' to that program, by changing parts of the code. Then run the resulting code through a suite of tests to decide rather or not the new code is closer to our desired goal then the old. Toss out code that doesn't work closer to a goal, keep 'mutations' that do until survival of the fittest code slowly leads to 'evolving' a program that exactly meets our desired goal.

The important idea here is that we DO NOT write the program itself. Instead we define the end state that we want the program to have and then set it lose to 'evolve' towards that goal. We don't know how it will do it. For instance a genetic programing attempt to build a radio once created one of the smallest radios...which worked by exploiting a flaw in the physical hardware to create signals in a manner that was never intended by the hardware itself! That particular example proved less-then-useful since we couldn't mass produce the flaw, but it did show that very unique solutions can come out of genetic programing that are different from how humans would ever design something.

It's believed that Hard AI will need approaches like genetic programing to work, approaches which do not attempt to define exactly how the AI should be created, but instead only encourage the AI to develop towards a final goal on its own. Our genetic programing approaches are still far to basic to produce a hard AI, and the final hard AI will likely use approaches complex enough to not be considered genetic programming, but they will likely build off the same concept. For instance one idea is to actually grow a silicon brain which models human brains and then develop an AI by effectively teaching a new brain the way we teach a newborn; instead of starting with a final AI we simply help the brain develop intelligence.

The Final Answer

This is all relevant because a side effect of genetic programing is that we don't actually know or control specifically how the AI develops its intelligence. Just like the example with the radio that worked only based off of a defect in it's hardware we may grow an AI that functions only through bizarre, and not always 100% desirable, means; but since we can't build an AI from scratch we need to settle for the best we can grow even if it has flaws.

TO give an analogy look at humans. Were a wonder of evolutionary engineering, but we have many 'bugs' due to evolution. Allergies are our own bodies sickening us with disproportionate defenses to harmless allergens, our brains have numerous logical flaws that cause us to misunderstand the world and make numerous mistakes (check Wikipedia for list of logical fallacies some time), and our pursuing short term pleasure can lead to long term suffering if not kept in check (addiction for example). We don't want these traits, but for various evolutionary reasons they were evolved for. Were stuck with them now. In much the same way an AI that we helped 'evolve' may have it's own quirks and limitations we would not have wanted, but are still an unavoidable side effect of their growth.

Therefore it's quite reasonable to presume that the same limits of our own brains would 'evolve' into an AI we grow; for exactly the same reasons these limits grew into our own brain. Confusion, pursuit of short term goals to the exception of long term, some 'logical fallacies' (the exact form may be different) could all evolve into a brain we grow. Thus any human conditions, including identity crises or even specifically gender dysphoria like you mention, are entirely possible with grown AIs; particularly in our earlier versions of hard AI (we may later learn ways to encourage evolution of AI hat avoids a specific undesirable side effect, but doing so risks evolving some other bizarre side issues).

It is possible each grown AI will have a 'personality' that results from their 'evolving' through slightly different steps as well, so an AI may very well assign itself a gender because it's evolution predisposed it towards certain traits that are associated with specific gender so it chooses to identify as that gender.

One key thing though, AI is still software. If we have to grow each AI from scratch (instead of growing a few and then mass producing copies of those we grew) then we would likely have some way of testing rather an AI was functional after it finished growing. Like with the example of the smallest radio which could never be mass produced because it depended on a defect in the hardware it's possible that some AI will grow to meet our stated standards, but prove to not actually be functional or useful (imagine an AI which grew to be every bit as smart as we wanted, but for some reason was suicidally depressed and only wished to end it's own life, perhaps a bit excessive of an example, but you get the idea). An AI with an identity crises will likely be seen simply as an AI that was 'grown' wrong, and best replaced with a newly grown AI...

  • 1
    $\begingroup$ +1 awesome comprehensive answer - starting with the definition of what is intelligence which we still don't' have a handle on. $\endgroup$ – Jim2B Oct 7 '15 at 16:26

Possible? If intelligence is based on human brain models, then it might be a real possibility. More generally, whether such beings have problems at all depend on how well we understand the brain architecture. If we have perfect understanding, then they will be built without such flaws. If we understand the low-level network connections but are iffy on how that comes together to "think" (like we are now) then the constructed brains might have all sorts of problems.

If that's what you need for your story, then to make it plausible for the reader look to reverse-engineered human brains and purposeful human-like programming.

As for Q2, it depends on the capability of the robot. If you're talking about my vacuum cleaner, the question is absurd. If you take an uploaded human mind in a new explorer body, then expect him to feel literally like a guy who woke up in a mechanical body.


Q1: Is it possible for a robot to have identity crisis like the description in case study B?

A1: Yes, if the robot has been programmed to have a sense of identity and the sense of identity is mutable. As long as the internal idea of the robot's identity matches the robot's observed state then there's no pressure to adapt to a new identity. However, if the robot were to observe in itself some very human attributes then its concept of personal identity may grow to include "I am human". If this change persists long enough then the robot will have to deal with the dissonance between "I have a robot body but I think like a human".

Q2: Suppose we isolate the robot on the moon so that there is no interaction with any live human beings, is it possible that it will soon question its ability to explore the environment and even its own identity?

A2: Again, this comes down to programming. If the robot is programmed with a mutable sense of identity and also given a desire to explore then changes to identity are possible. A robot programmed to explore an environment will do so until it can't anymore. Is interaction with a living human being required for this robot to go exploring?


There would need to be a really good reason to give a robot a mutable sense of identity as so much can go wrong...look at all the angst in teens as they make the transition from an identity as a child to an identity as an adult. Do you really need/want angsty teen robots?


I think both of your questions are reasonably easy to answer if you first answer what it means for a robot to have identity in the first place. Unfortunately, that particular question is a remarkably difficult philosophical question for humans, much less our creations.

The route I would take is the one proclaimed by Douglas Hofstadter in his book, I Am a Strange Loop. To take the quote off of the wikipedia page:

In the end, we are self-perceiving, self-inventing, locked-in mirages that are little miracles of self-reference. — Douglas Hofstadter, I Am a Strange Loop p.363

Now ignoring the human philosophy of whether this is accurate or not, Hofstadter found it meaningful enough that it could be accurate for us, and it is mathematically consistent, so it seems like as reasonable of an approach as any for identity in robots.

I would consider the two key phrasings in that quote to be "self-perceiving" and "self-inventing." We've already got self-inventing going pretty strong: nearly every major AI algorithm has some sort of self-inventing behavior like genetic algorithms. Self-perceiving is more difficult, especially when we're still building the concept of a robot's self in the first place.

Let's weaken self-perceiving a bit, since that is a hard one. Let's start by trying to perceive the effects of one's self on the environment. This is easier because it fits more in line with what we usually think of with sensors. Consider a robot that is supposed to "help society," like so many good robots are. At some point the robot will either need to become aware that its presence changes human behaviors (to something different than that behavior would be away from the robot), or find frustrating issues that they cannot seem to resolve. They need to realize a key limitation of their physical form: their sensors are not perfect. The mere act of trying to measure something changes it because they are strapped to the robot, and the robot's presence changes the people (this works for non-localized systems like traffic cams too, but its a lot more complicated).

The robot will need to come to grips with its own effects, in its own way. At some point, one of the self-inventing bits will figure out something that gets "close" to that idea. At that point, the robot could consider that self-invented bit its "core," a self-image of the greater robot self.

Now it can have identity crises, for it now can have an identity in the first place.

And, if we're lucky, the robot will not build itself around an image that it believes is immutable and perfect (most modern optimizers are built to assume their fitness function is perfect). In such cases, the robot will undergo the transformation seen in countless robot stories into a dangerous machine, until it either realizes the image is mutable and imperfect, or the hero of the story reaches out and breaks that image, destroying the identity complete.

All humans learn how expensive it is to believe in a thing that is false. We learn to adapt. Hopefully the machines will learn to as well.

The weakness of thinking machines is that they actually believe all the information they receive, and react accordingly.
-Dune: The Machine Crusade


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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