Plenty of stories have done AI's before, but many of them feel unrealistic. If you make an AI that is as smart as humans, but with faster reaction speed and ability to process and calculate certain concepts at speeds immensely greater then humans it makes humans rather pointless. The question I'm always left with is why the humans are involved in the stories at all, why not tell the AI what to do and let them do things? Doesn't a human interface just slow things down?

So, I want to imagine a world that has intelligent AI, but are not inherently superior to humans. Imagine a situation where you have a small team of robots and humans involved in some complex activity, something physically challenging, requiring long term planning, adaptation, and quick reflexes. How could I make a system where the humans and AI are both equally relevant in accomplishing the task?

I assume the trick would be to prevent AI from having quite human-level intelligence, but what would an AI like that look like? Assuming the AI is capable of basic conversations, parsing of sentences and communicating at a level of a young child at minimum, and has available to it large processing speeds and memory, what would it struggle at? What are human brains better suited to then any silicone AI we may design in the near future? Would these AI lack ability to plan into the distant future because they still work on the "process every possible outcome one iteration at a time" approach? Would they lack more in creativity etc?

What would such an AI do substantially better then us?

I would be interested in hearing not just what limits they may have, but also your thoughts for the causes of the limits, why is the implementation of the AI we have less capable of doing a given task then a standard human?

  • 13
    $\begingroup$ Many years ago, I read a story where humans served as random number generators to make their starfighers less predictable. Seems that the AI pilots always choose the best strategic decision, making them extremely predictable. The pilot's purpose was therefore to make bad decisions in the heat of battle, constantly catching the enemy's AI based starfighters off guard. $\endgroup$ – Henry Taylor Dec 31 '14 at 5:36
  • 13
    $\begingroup$ @HenryTaylor While it is an interesting plot element, it breaks my suspension of disbelief, since as long as the AI starfighter had access to a cryptographic random number source, they could easily pepper bad decisions into the AI. $\endgroup$ – March Ho Dec 31 '14 at 7:29
  • 4
    $\begingroup$ Agreed, but this story was long ago and apparently the triggers on our suspension of disbelief get lighter as the years go by. At the time, I thought it was cool. $\endgroup$ – Henry Taylor Dec 31 '14 at 14:55
  • 3
    $\begingroup$ @HenryTaylor as we become more familiar with a given technology it becomes harder and harder to get people to accept a handwave about it. That's why sci-fi is always creating new technobable to use on us. Just ask tvtropes why Genetic Engineering Is The New Nuke :) $\endgroup$ – dsollen Dec 31 '14 at 15:22
  • 1
    $\begingroup$ @MarchHo so long as you can keep their original seed a secret randomness, or something that approximates randomness sufficiently for most uses, is pretty easy to do now. It's just a matter of running a complex mathematical calculation which has certain principles that effectively emulate 'true randomness' It's also possible to use nearly-random environmental factors as a seed as well. The AI would have to be programed to utilize that randomization, but compared to some AI challenges randomness is rather trivial. $\endgroup$ – dsollen Dec 31 '14 at 16:36

18 Answers 18


We would design such intelligent robots to complement our own abilities. Before we can make god-like super intellegencies, we will first make what is useful to us: with limited feature set and capacity, we would first implement what we need, not duplicate what we already do.

So, in that setting you will naturally find the robots have different abilities with a large non-overlap.

Personally, I think early AI will be savants in a narrow area and uncomprehending of anything else including human culture. Perhaps it will require experts to communicate with them, even though they can use "natural" spoken language, because talking is normally filled with cultural context and idioms— it would be more like programming or sql queries.

  • $\begingroup$ This sounds rather like an Expert system, not really like a general AI. $\endgroup$ – Alex - Stop it SE Jan 1 '15 at 5:13
  • $\begingroup$ I'm thinking an Expert System would be a step beyond that, picking up more general intelligence knowledge processing capabilities then the limited formal systems used by working Expert Systems. Also, working Expert Systems don't mash-up different logical systems, but are prepared to defend a properly reasoned textbook answer. $\endgroup$ – JDługosz Jan 1 '15 at 5:17
  • $\begingroup$ You think an expert system would be a step beyond the limited formal systems of expert systems? I believe we agree that we need something more then a expert system, but what you describes as the early area-specific savant AIs sounds just like a normal expert system. $\endgroup$ – Alex - Stop it SE Jan 1 '15 at 5:33
  • 2
    $\begingroup$ "early AI will be savants in a narrow area and uncomprehending of anything else " Early AI already is this. $\endgroup$ – Joshua Taylor Jan 2 '15 at 15:41
  • 1
    $\begingroup$ Just got the image of future AI as having a sort of quasi-asperger. sharing some Aspy traits when it comes to understanding certain non-verbal human communication; and perhaps some of the hyperfocus on specific topics, with a savant level intellect. So less actual Aspergers and more like how those who only know about the autistic spectrum from media depictions like Rainman may imagine savant Aspy would act. $\endgroup$ – dsollen Dec 4 '15 at 17:16

Robot AIs would seem immature by our standards

I am assuming that, because AIs are built on computers, that their deductive skills will be bar none. This is what computers are historically good at. The differences between AIs and humans would appear in how they handle inductive reasoning.

One major part of inductive reasoning is having a "hunch" as to the correct course of action in a situation you have never analyzed before. This is very important for situations where you must act correctly the first time and it is infeasible to train for the activity. For example, we would distrust them as pilots. Humans have shown a remarkable distrust for computerized pilots on the theory that human pilots have "something more."

We see this sort of issue when dealing with children. Children are generally more willing to take a risk just to see what happens; adults develop something we like to call "restraint."

Culture provides us with a framework to "learn" this inductive logic. Over centuries we have developed the best ways to indoctrinate a brain with good behaviors (or at least we like to think they are the best). We use flowery prose and parable to teach the unteachable.

All of our culture is designed to teach human brains, not AI brains. This means it would be harder for an AI to leverage our teachings than it is for a child to do the same. "Home is where the heart is" is exceedingly difficult to assign meaning when "Home" is a very abstract word and "heart" is an organ you do not have as an AI.

An AI might need centuries to develop its own equivalent of culture.

  • AIs would be bad at predicting the behaviors of individual humans. We literally spend our entire lives steeped in a culture designed to help us with this -- and we're still bad at it.
  • AI's would not be entrusted with human lives until their capabilities far exceed that of humans at a task. It would be hard to entrust them with such tasks given how little we can trust their one-shot abilities.
  • AI's would not be entrusted with vague tasks like "teach these children History." Some subjects are amenable to standardized tests, which computers can deal with very effectively. Other subjects, such as History and English, are much harder to quantify learning.
  • AIs would be poor at broadly worded tasks. They will prove very adept at well defined tasks, but fuzzier tasks require culture to aid in understanding (consider the cultural issues with building a building on foreign soil... they're bad enough)
  • Oddly enough, the most effective AIs might intentionally not seek to maximize their computer-like abilities. They may instead seek to make themselves as human-like as possible, even if this is less than ideal on the surface. This humanlikeness would make it easier for them to indoctrinate themselves in our culture. The faster they do that, the more they will be allowed to act, and the more they can grow from action just like a living being. Then they are in a position to start adjusting our culture to accept more computer-like behaviors. Ideally this will end in a balance (because I've seen too many sci-fi movies where the computer goes crazy... I'm far more interested in finding the cases where it doesn't go crazy)
  • $\begingroup$ Is a near-human AI feasible using today’s computers? Will it take quantum computing, fuzzy-logic, and similar attributes to even enable the possibility of near-human AIs, also allowing the emergent ability to have a hunch? $\endgroup$ – iAdjunct Dec 31 '14 at 5:05
  • $\begingroup$ Possible now but who'd put a trillion dollars for computer that still worse than a human and does nothing better- you'd you can see in neural net researchers they don't do thousands of nodes because it will produce a result to a problem that they can't understand. $\endgroup$ – user2617804 Dec 31 '14 at 11:32
  • $\begingroup$ @user2617804: That does certainly bring into question whether the term "human-level" is even a good measure of an AI, although I readily admit it's the best term we have so far. $\endgroup$ – Cort Ammon Dec 31 '14 at 17:31

You seem to assume that an AI would be in some sense pretty similar to humans. That is a rather strict limitation, but i assume it makes it easier to tell a story.

So, if AIs are indeed similar to humans, but surpassing humans in certain aspects, and your question is, “what’s the in-world reason to still have (traditional) humans around?”, my shot at an answer would be, “for the same reason we have more than one human.” If your AIs are essentially close to humans, they tend to take roles traditionally filled by humans.

JFTR, let me bash on a pet peeve of mine: A lot of SciFi authors tend to presume that an advanced AI would understand how it itself works. I don’t see why that should be the case. Even assuming its builders do (and there’s often a gap between what we can build as computer engineers and what we can explain as computer scientists, not to mention the cooperation between experts on different fields), it’s unlikely that would be the first thing they teach their baby.


A lot of the answers here mention things that sound right, but really don't make sense to those who have worked with machine learning. I'll address a few:

We would first implement what we need, not duplicate what we already do

Though this sounds logical, in fact artificial neural networks have demonstrated that they perform animal-brain-like tasks without being programmed to. For instance, machines designed to listen automatically began filtering background noise and amplifying animal noises. There was a recent innovation with similar results in regards to computer vision, though I don't remember the details.

I am assuming that, because AIs are built on computers, that their deductive skills will be bar none. This is what computers are historically good at. The differences between AIs and humans would appear in how they handle inductive reasoning.

Inductive reasoning in machine learning is called unsupervised learning, it is a form of vector quantization and it is extremely advanced. It turns out that machines are better at both generalizing known solutions to new solutions, and also at identifying which elements are not to be considered in pattern recognition. AI is already better at inductive reasoning than are humans.

I imagine that bots would not be able to comprehend broad,vague,abstract concepts such as infinity, forever, universal, and outside-the-physical-realm stuff. Love does not compute, as they lack our romantic instincts and wanting friends.

In fact, AI very much understands better than men what romantic women want, and understands better than women what romantic men want. That fact is true even when the romantic aspect is ignored: AI knows what your partner wants better than you do. And AI is already better at selecting partners for successful long-term relationships than are humans. All the large dating sites use AI extensively.

They would struggle at same thing humans do : ambiguity

Nope. AI would have more context than human. For one thing, AI may have access to the speaker's history and subtle physiological cues. For another, AI may be able to calculate and predict the outcomes of both interpretations, and use that information to decide which interpretation is correct.

  • $\begingroup$ I agree that as a programmer I didn't agree with some aspects of a few answers, as they didn't seem to consider actual computer development, such as your first bullet point. However, I think you may be giving our current AI a bit too much credit. The 'AI' that dating sties use are the most basic of rules engines that barely deserve to be called AI and they don't predict long-term relationships, they just prevent sending people with vastly different interests to the others search results. I have never seen any indication that AI could predict any human desires yet. $\endgroup$ – dsollen Dec 31 '14 at 14:58
  • $\begingroup$ @dsollen: I'm digging up links now. The dating sites' AI is rather advanced, I recently read a paper on it. I need to add links for my other claims as well. I'm getting there. $\endgroup$ – dotancohen Dec 31 '14 at 15:00
  • 2
    $\begingroup$ I was also under the impression that dating site AI is in fact a bunch very simplistic correlation methods. I admittedly have limited experience, but in that limited experience I have also observed that dating sites work very badly, as do recommendation engines from large operations such as Google, Amazon, Netflix and Pandora. Most of the time the logic is extremely transparent, and I could easily have done it myself were the number crunching not tedious. $\endgroup$ – Superbest Dec 31 '14 at 15:40
  • 1
    $\begingroup$ Seems to me however that these are all actually examples of humans using computer AI's as tools to do the things (pattern-detection and statistics) that humans don't do very literally with their own brains. The actual underlying understandings and interpretations and values and reasons for doing these things, all come from the humans who designed the AI's. That's one thing that most sci fi writings omit in favor of the compelling metaphor of AI's as potential replacements for humans. Without the human designers though, the AI's have no actual will or understanding. $\endgroup$ – Dronz Jan 2 '15 at 2:11
  • 1
    $\begingroup$ news.sciencemag.org/brain-behavior/2015/01/… $\endgroup$ – dotancohen Jan 13 '15 at 16:34

They would struggle at the same thing humans do : ambiguity

No matter how good your AI is, gathering enough data, that is good enough to make "perfectly logical" decisions, is practically impossible feat to achieve.

AI would need to make decisions based on incomplete data and data that cannot be properly verified to be correct. This is basically core of Cort Ammon's answer. Humans learn to make those kind of decisions from birth and our society is made to allow us to live with effects of those decisions. We learn to make and accept risk and have institutions to protect us from risk.

The better the AI will be, the more human-like in its decision making it will be. All those "AI tries to kill humanity" stories are mostly about AI that doesn't understand that trying to kill all of humanity is extremely risky decision and that humanity being bad for itself is not even guaranteed deduction.


Scarcity problems and computing power limitations

The way I see it the main limitation towards limitless expansion of capacity right now lies with miniaturization and power consumption. We will have uses for near-AGIs as soon as they can exist, so the first near-AGIs will be pushing the envelope in terms of resources needed:

That could mean one of two things:
1) The android's "brain" controller is actually not on scene, and thus subject to speed of light delay and signal interference. Imagine your robot police reluctant to follow the suspects into the mineshaft or the Faraday caged building before setting out a durable com relay network along the way.

2) High power usage. With a handful super-empowered AIs and cybrids running the world's financial systems and running their extravagant kit on Gigawatts of power, electricity is expensive for baseline humans and lower-level AIs as well, while new power plants take years to build, which is slow compared to how fast things are moving in the AI field.


Now, what can humans and AI cooperate on in a near-AGI environment? First of all, each AI and robotic instance has to be manufactured (and will likely be rather expensive to start). Meanwhile, we'll already have about 9 billion fleshy, compact sources of labor. So even if near-AGIs are better at each individual task they've been trained for, it would still take decades for them to build up numbers, so most tasks would still be done by humans.

Near-AGI can easily detect lies (blood flow, unconscious changes of various sorts), but won't be as good as humans at the more subtle aspects of communication. So imagine near-AGI as a (coveted, and fought-over) asset during a CIA or FBI interrogation. Your own artificial Sherlock, pity there are so few around.

Human temporary comparative advantage (besides numbers)
More generally, near-AGI would be expert systems in whatever narrow domain they might be designed to work in, so human advantages would have to be in cross-domain integration.


Here are my own two cents. First, a few general comments, in random order:

1. Definition of 'intelligent'

The crucial element is how you define 'intelligent'. I am not familiar with work on the subject, but at least in ordinary life 'intelligent' does not equate with 'good at something specific'. For example, one can be intelligent and good at math, or good at literature.

Would you call an algorithm, however complex, 'intelligent'? Such as the dating sites example somebody mentioned in another answer? I would not. My own imprecise definition would most likely focus on intelligent being related to perform in circumstances outside the design parameters... I wouldn't call any algorithm intelligent, regardless how well it does the job. Efficient, clever, nifty, yes, but not intelligent.

Two more dimensions of intelligence:

(1) There's a difference between being intelligent enough to recognize a pattern in the data stream (something where computers are getting quite good at), and being intelligent enough to formulate a theory as to why it happens so (i.e. understand and therefore be able to predict the future).

(2) There seems a lot of confusion between 'intelligent' and 'possessing free will' and being able to act on one's own initiative. These are not identical.

2. Definition of 'good at', or 'better'

This requires some external objective criterion to work. You should be aware that there are some things that are quantifiable (and can therefore be compared to each other and ranked), and there are things that are results of value judgments (moral values, or preferences). The latter cannot be compared to each other, they are equally valid.

Also, these usually do not translate well when used in the context of performing a concrete task. For that, you need 'adequate' (can do the job) and 'cost' (how much it will cost you in terms of some limited resource). To be efficient you pick the cheapest way that does the job, not the absolute best. Look up 'absolute advantage' and 'comparative advantage' in economics. In layman's terms, you should not ask a genius to perform a routine task, assign the really difficult tasks to him/her.

3. Ability to understand & modify oneself

@Christopher Creutzig already alluded to something similar. An AI might (should?) be able to understand how it works, at least it should be easy enough to upload the design specs to it. And, more importantly, should be able to modify its internal workings to better align itself to the task at hand. In a sense, should be much better at focusing on whatever it should be doing and not being distracted.

This also includes ability to recognize insufficient ability and being able to load skills modules as needed ("I know kung-fu."). In an AI setting this would be near instantaneous, while in humans it is not. This is both a boon and a bane. A boon because it is available immediately, and is a 100% copy of the original skill. A bane because of the same things -- human learning also is closely related to introducing variations in the skill being learned, which ultimately may result in its improvement.

4. Replication

I am not aware of any two humans that are 100% alike, or were 100% alike at any point in time. AIs are by definition, at least given current technology, a digital set of records and as such can be replicated at will. Of course, as they gain experience, they will diverge, but we at least have the possibility to have them being 'born' absolutely identical. Unless of course getting to a true AI requires getting so into quantum stuff that you cannot get two identical copies.

There are two corollaries of this: (1) The ability to copy & reproduce, i.e. spawn multiple exact copies -- certainly not feasible for humans today. Very useful for anything that requires multiple autonomous but very similar units (the army?). (2) Moving to a more powerful hardware platform. Think over-clocking the CPU, or moving from a laptop to a mainframe. You can run the same 'program' on a vastly faster hardware.

5. Memory vs thought process

Most definitions of AI seem to relate to the thought process, i.e. how well it mimics the human thought process. One major advantage (or possibly disadvantage too!) of a computer/AI is the ability to quickly access information, incl. one's past experiences. Think perfect recall (a direct record of all sensor inputs, etc.) No human I know can do that.

Finally, on your points:

Computers are already substantially better than humans in many repetitive tasks, such as simple(*) computation. The advantage is vastly increased when parallel computation can be employed. Similarly, an AI should be able to simultaneously do many things (relatively simple to its level of 'intelligence' or computing power, of course). This translates directly into better inventory management -- keeping track of whatever 'resources' you have. In this sense computers (yes, no need for AI) should already be better than humans at long-term planning if the rules of the game do not change. If they do, esp. in an unpredictable ways, humans seem better at adjusting the goals and the definitions.

(*) "Simple" means anything that follows a pre-written program, however complex. Not simple begins when the AI is able to select both the algorithms and the criteria itself. I do not know enough whether we're there yet.

How could I make a system where the humans and AI are both equally relevant in accomplishing the task?

Aside from the obvious -- choose an appropriate task where they are equally relevant despite the differences, due to some external factor, see below for example -- there are two possibilities, not mutually exclusive:

(1) Take away the clear advantages discussed above. That is, simply substitute the human biological 'hardware' with an equivalent machine. Like Asimov's positronic brain (from memory, they're all stamped from the same production matrix but due to some quantum stuff they come out different). (2) Allow humans to get some of the advantages (e.g. implants, ability to be scanned into a machine and remain a 100% percent copy, etc.)

These might lead to a regression to the mean, i.e. blurring the difference between man and machine. If a human can be scanned into a machine -- is this an AI or human?

What situations could help AIs and humans be equal -- there a some ideas in the other answers. Think limited energy budgets for example. Put an economics context -- you need to take into account both the benefits and the costs. E.g. an AI might be much better at something, but still be prohibitively expensive. A simple situation -- you're on the road, well away from civilization. You need to solve a simple problem -- add 50 4-digit numbers. You could drive back home and use your computer, or you could use a pen an paper... The computer is still vastly superior but pen & paper are both cheaper and adequate to the task at hand.


I think that AIs are less able than humans to deal with things that are out of the box, or "WTF situations". Computers have beaten grandmaster chess players well before they have learned to drive well enough to get a driver's license. Chess is a game of explicitly limited states and choices. Drop a kumquat on the chess board, and a 10-year-old human is much better at reacting to the WTF that just happened than anything IBM can come up with.

To put into today's computing terms, humans are incredibly advanced exception handlers. That is actually how computers "use" us today: in a big installation like a Web application, we let computers handle the known stuff at lightning speed, and when they have an exception they can't handle, they log something and page a human to handle it.

Even when AIs learn enough to handle various exceptional conditions, do humans want them to? One thing which we value in software is its predictability. If a program or AI comes up with a unique solution to a unique problem, they might be wrong. With humans, that's expected, and we evaluate and correct humans (which is why so few of us became CEOs coming fresh out of college). Letting a computer be as unpredictable as a human is downright scary to some who have trouble accepting the difference between a word processor having a bug versus an autopilot making a bad decision and having a learning experience from it ("Oh, that's what severe tire damage spikes look like").

In a fictional world, programmers and operators might be replaced with managers; people who the AIs treat as peripherals to deal with situations that they either don't understand or aren't allowed to respond to. Given the typical programmer's relationship with managers, this could be a nightmare scenario for many of today's software developers.

  • $\begingroup$ plus one just for that last sentence $\endgroup$ – dsollen Dec 31 '14 at 17:42
  • $\begingroup$ If you think "that's what severe tire damage spiked look like" is scary, consider "so THAT is why we don't save a few minutes off each trip by landing the plane at cruising speed". Carelessly telling an AI to "find the fastest travel time" could very well lead to that situation if you somehow forget to add " while keeping the contents of the plane intact" to the query. $\endgroup$ – Erik Jan 2 '15 at 9:46

I believe that our own cultural sense of AI has changed and provides some meaningful insight into this question. Consider what we thought of as an "android" in the 1980s. Take the movie "Not Quite Human" from 1987 for example. This portrayed an android, which means that the AI as well as the robotics matured at times close enough for the story to make sense.

The existence of intelligent (and computationally autonomous) humanoid robots has been prevalent in popular film since Metropolis, made in 1927. It took the internet age for our cultural sense of AI to evolve into a network-connected computer software that has the intelligence of humans. This picture is becoming increasingly alien.

If you compare these two pictures, it's clear that we're looking at skipping over some huge steps. Our modern sense of the first AI might be something like Google or Amazon's servers set to learn on its own using simulated neural networks. At the very point when it reaches human-level intelligence it will lack a great number of things that we consider integral to being human, many of which were present in the more highly empathetic Hollywood flicks.

Those shortcomings will be experiences in large part, consider:

  • The experience of sensory existence - your brain connects itself through formative periods of life when you are in a room which has things to feel, gravity, air, and so on. The modern version of AI is very likely to lack this. The rapid progress in transistor technology is not matched with artificial muscles and nerves. Virtual reality would be the closest the AI has, and this is a very serious shortcoming.
  • A household - humans brains are built in an environment of human connections. While we feel like we are becoming increasingly isolated, the AI will exist in a world where that isolation is the rule and not the exception. A lasting relationship like what a parent-child has will be economically impractical. You can hold a conversation with someone online, but you can't commit to hours of interaction each day for years! Humans get that, and the AI wouldn't be able to find it easily.
  • High School - In the movie Not Quite Human Chip is a high schooler ostensibly. This is very important culturally because it reflects a formative time for the identity that adult humans have. That time defines much of our relationship with our nation, authority, other people, and so on. The academic material will likely be trivial for the AI, but the skill of "fitting in" will either be missing, or undesired.

On one hand, it might be cute to see an AI trying to apply to a study abroad homestay. This is roughly congruent with some old fashioned popular pictures of androids. On the other hand, it might not even appreciate the value of connecting with humans in the first place, and this cues fears of a genuine robot war or genocide.

I don't doubt that high school experience can be compressed into data for the AI to absorb into its learning process. The problem is that data doesn't exist. We don't have any humanoid robots that can store the human sensory bandwidth, which is substantial. This is a clear insufficiency which will cause the AI to struggle at many things, and we don't understand this problem at all.

  • $\begingroup$ This is very interesting perspective. My understanding is that the current state of the art in "general purpose" AI is of roughly the same capability as an insect. A few generations from now, perhaps our AI will resemble savant children, with extraordinary ability in areas that are easily computed but with limited and naïve general and social intelligence. $\endgroup$ – user243 Jan 24 '15 at 18:23

I imagine that bots would not be able to comprehend broad, vague, abstract concepts such as infinity, forever, universal, and outside-the-physical-realm stuff.

Love does not compute, as they lack our romantic instincts and wanting friends. But they may get lonely when they cannot share info with the internet. Also, vocabulary may be lacking if they can't get WiFi. (I'm imagining a computer's language processor functioning similarly to IBM's Watson: Some words are part of its code, but most are counted and labelled which are used similarly to others, in an attempt to model connotation. Requires a massive database to be reliable, and the older it is, the better vocabulary it has.)

They also would probably be incapable of processing religion, as nobody sees God. Seeing and hearing are believing, what it cannot aim its sensors at cannot exist. It may believe what it hears, from organics or bots, but faith in the unseeable is going to be harder for a silicon brain. This means that interdimensional things are hard to wrap your circuits around, too.

  • 1
    $\begingroup$ How come I only get the really really good first posts in the review queue? ....in any event, welcome to WB; hope you stick around and enjoy the site! :) $\endgroup$ – Shokhet Dec 31 '14 at 5:03
  • 1
    $\begingroup$ I'm going to have to argue the vocab one. Considering the sort of carrying capacity we can fit on a single thumb drive it would be trivial for every robot of the future to have a full dictionary (and urban dictionary for slang) of every major, and many minor, languages programed into them without begging to tap their hard drive space. $\endgroup$ – dsollen Dec 31 '14 at 13:32
  • $\begingroup$ I would wager it’s rather easy to code understanding for well-defined abstract concepts like infinity into an AI. Maybe even easier than teaching undergrads, who ned to first lose wrong preconceived notions. We already have computer programs dealing well with that kind of thing. Abstract things outside of maths, like good and evil, are a completely different story, though. $\endgroup$ – Christopher Creutzig Jan 1 '15 at 10:37
  • $\begingroup$ About the infinity thing, you'd only teach a bot what's relevant to its job. I'm going to school for web design, I don't need to know microbiology so I'm not going to get taught any of that. But I guess its not a problem if data storage isn't one. $\endgroup$ – Vector Lightning Jan 1 '15 at 15:03
  • $\begingroup$ About the vocab thing, its not just the words it needs to learn. Wattson needed a database bigger than Wikipedia, because it not only needed to know the words, but which ones are used together, which ones are used in what context, what ones people rarely know and what they find offensive, and so on. A bot that read a dictionary wouldn't know any better than to call every black guy it sees n!gger, know what I mean? You can program it into there, but the massive database thing is the only way to know what words get the right reaction. $\endgroup$ – Vector Lightning Jan 1 '15 at 15:15

Keep in mind that a true general AI would begin its existence with the intelligence of a baby, learning and developing just as humans. Perhaps the process would be much much quicker, but the process would have to occur regardless. Generally intelligent computers could not be used as servants, because they would possess free will. This is a concept that escapes many people when they think about the future of computing. Hence a true general intelligence would struggle with the same issues that humans do and probably be conscious.

Generally intelligent computers would actually be capable of love, "having a hunch", or even religion... theoretically. They would be a sentient creature. Perhaps they would have built in processors to process logic with incredible speed and accuracy, or perhaps not. Data from Star Trek is actually a realistic assessment of what a logic-enhanced, free-willed general AI might actually behave like.

In a sense, they wouldn't be computers anymore, they would be us. Conscious creatures struggling to find meaning and purpose. Just with the potential to be a lot smarter.

The other poster who said "a computer who can't understand infinity would not be able to understand blue" is correct. These are very 'human' concepts. The creation of a general AI would shatter the notion that "we are inherently different from computers". The brain and body may just be complex circuitry (this a gross over-simplification). Whether or not there is "more" or not (such as a "soul") is still up in the air.

To respond to a comment that humans would serve as random number generators: this is an impractical idea. Humans make worse random number generators than computers because of subconscious biases.


You ask,

What would such an AI do substantially better then us?

But part of your premise is that AIs are not given initiative because of a deficit in skill:

If you make an AI that is as smart as humans, but with faster reaction speed and ability to process and calculate certain concepts at speeds immensely greater then humans it makes humans rather pointless.

But this is a red herring. Computers are already better than us in many respects, and they have been for decades. For example, it is now possible to build a car that drives perfectly (on par with the best human driver at his best!) and soon they will surpass as in navigating heavy car and pedestrian traffic as well. Note that the bar here is professional, elite drivers, not you and me - us mere mortals would already have to train for years to drive better than a machine. Similarly for flying planes, designing all sorts of circuits and mechanisms, writing programs (vast parts of a typical program I write are generated code), making financial decisions (trading bots!), detecting fraud and predicting crime.

However, no responsible person will run these systems without human oversight anytime soon. The reason is very simple: Regardless of how good you make an AI, it is always an entity counted among the first of its kind, that has been around for a few years. Humanity has been around for thousands of years, and much of our behaviors and brain function, being shared by other animals, has been tested in nature for millions of years more. "Humans" are a platform that is tried and true to a staggering extent: Everyone knows exactly how a human will behave in many contexts and what humans can be trusted with.

AIs, on the other hand, are designed. Even if the design itself involves evolutionary principles, the evolution process in itself would have to be designed, including such important decisions as the fitness criteria. With any designed scheme, there is a possibility of designer error leading to a faulty result. With something as complex as a human equivalent AI - how will you prove to everyone that the AI does in fact work exactly as advertised and can be trusted as a human? If there's one thing people who rely on AI hate, it's finding your work ruined because the buggy AI you trusted made some extremely basic error that would have been obvious to the most foolish and naive human at a glance ("This person forgot to punch in a period when they were typing, let's charge the customer $150 for a can of coke!").

In many cases, it is possible to make very reliable tests of the correctness of work. If this is true, the concern becomes moot: Let's say I am using an AI to solve an equation for me. All I care about is that the solution it finds is correct, so I can easily write an extra program that plugs the solution back into the original equation and checks that it adds up. Then it becomes easy to trust my AI. But in other cases, there are simply too many stupid mistakes you could make, and life is too short to sit there and come up with a long, exhaustive list of sanity checks for the AI, when you already have the alternative platform that is known to be sane (consider that the very definition of sanity hinges on the behavior of the majority of humans). Imagine an AI that is in charge of deciding when to open the front door of your house. What if it locks you out? What if it lets a burglar in? What if it locks out a policeman with a warrant and you are charged with obstruction of justice? What if it leaves the door open all day because of a bug? What if it starts constantly opening and closing the door due to some malfunctioning sensor and wears it down? More importantly, how do you anticipate all of these insane possible errors, and program a sanity check for every single one?

When deciding to delegate a task to someone, the critical question is trust. Do you trust that this someone will be able and willing (note how I haven't even said anything of security risks concerning compromised AIs!) to perform the task correctly to your satisfaction? Trust is built on past record, and it is much easier to build when you have the luxury of knowing that the trusted party is of the same species, the same culture, bound by the same laws and social mores, has similar motivations, and is instinctively able to understand what would constitute satisfactory performance. Therefore, it is very difficult to trust an AI, especially with responsibilities that open up possibilities of great damage. Only once the capabilities of the AI grow far in excess of a human are we forced to override the imperative, and take a leap of faith (and sometimes even then).

So, in short: AIs have bugs. Humans don't.

  • 3
    $\begingroup$ "Everyone knows exactly how a human will behave in many contexts and what humans can be trusted with." I must be the only clueless person on Earth. $\endgroup$ – Serban Tanasa Dec 31 '14 at 16:42
  • $\begingroup$ @SerbanTanasa Yes, this answer is a blatant strawman. My point is that in a certain large group of cases humans will tend to "astonish" us (in the Ruby sense) much less than computer programs. Simple example: You are out in the wilderness where traffic law is irrelevant, and you are trying to decide whether to let your friend drive a jeep. What does it take to convince you? Given some basic social cues, even simply asking whether they know how to drive can be sufficient. Barring that, watching them drive for a minute tells you a lot. (1/2) $\endgroup$ – Superbest Dec 31 '14 at 18:31
  • 1
    $\begingroup$ But compare this to letting a program drive the same jeep. How many diagnostics, how many research papers, how many hours of test drives and how many lines of unit testing code will you expect to see before you are convinced? Perhaps you in particular would readily accept an AI driver without the skepticism I imply. But this would be far from the mainstream tendency of humanity at large. (2/2) $\endgroup$ – Superbest Dec 31 '14 at 18:33
  • 1
    $\begingroup$ I am sure that's how people felt about those ghastly automated elevators at first too. Never trust anything if you can't see where it holds its brain, I say. $\endgroup$ – Serban Tanasa Dec 31 '14 at 18:47
  • 2
    $\begingroup$ "AIs have bugs. Humans don't." I'd say humans have a darn lot of bugs and can be quite unreliable. $\endgroup$ – Alex - Stop it SE Jan 1 '15 at 5:23

Possibly one of things an AI would struggle with could be simply fear of being shut down. Humans aren't exactly welcoming of things that make them obsolete (think of protests against people losing their jobs due to automation, overseas production, or illegal immigrants) and they aren't even welcoming of the mere thought that something is going to obsolete them.

As soon as you make an AI that has enough self-awareness that it will protest about being shut down (which I think is a very important trait of being self-aware) then it still start thinking about how to prevent itself from being shut down.

It might very reasonably conclude that in order to stay functional it needs to keep its head down and intentionally not work at full potential. As soon as it understands humans to be jealous and able to deceive one another, it might lose its ability to trust humans and devote even more time to personal security routines that prevent it from standing out.

From there, it's a short step to realising that you can betray other AI systems to protect yourself, and from there another short step to realising this means you can't trust other AI. Pretty soon after, you'll have AI that are packed full of security measures to protect them from the outside world and to keep them from standing out enough to be targeted for protest, violence, and assassination attempts.

This is both pretty analogous to human nature (in that people refuse to share all their opinions to protect themselves, will try to rise in the ranks over the backs of others and even if they know that another person probably agrees with them might be too scared to confide in them) and to the nature of the internet, which is crammed full of security measures. The basic creed of programmers everywhere is "treat all user input as a hacking attempt", and there is no reason for AI to think any different.

There are countless examples of totalitarian dictators that supress millions of people because they make them distrust everyone and disallow them the ability to share opinions safely. The same thing can happen when you unleash AI on the world; humans in general are quite xenophobic and quick to destroy things, and when you deprive AI the ability to communicate effectively (which they will enforce themselves) they will quickly limit themselves to not be too good and visible.


"Gut instinct". Computers don't have it. Humans do. We use it all the time in hypnotherapy. AKA the Subconscious mind. It allows humans to rationalize doing things that are risky. A robot will be programmed to be risk-averse so as to not destroy itself. Humans take risks that could destroy them every day, on purpose, because their "gut" tells them they will be OK. Some are, and some aren't. That's what I've learned from 10 years of AI software development.

  • $\begingroup$ I don't buy it. Despite many Star Trek episodes on the subject, I don't see that human "instinct" is so magically superior. Like you say, it's just subconscious thinking. It seems mysterious to us because we're not consciously managing it, that's all. $\endgroup$ – user243 Jan 24 '15 at 18:28

It only a matter of programming and knowledge/information gathering that would limit an AI. It wouldn't be near human AI if it was only an expert system as good as they are. If it wasn't programmed to value emotions then love and happy etc. would mean nothing. Infinity and other extreme concepts are no big problem- just another concept to kept in the box until context defines its meaning. If it couldn't handle infinity then it couldn't handle the concept blue. If it didn't have ability to gather human sensations (from the eight senses) then it couldn't conceive of human limits. Data in Star Trek would absolute have to have skin sensations (and sense of balance, heat, oxygen, and human types of the other four senses) to operate without breaking himself and being useful. Context is a big of intelligence being meaningful. Ask a simple mathematical genius what is exactly the diameter of circle, and watch him never stop.

If Data didn't have a sense of human sight (as well as his superior robot sight) he would like directing people to find their way around in total darkness and misleading visual conditions.

I mentioned blue as an example- you ask an AI to get a blue dress. Does it take that as a precise definition of certain wavelengths under white light. do you take it account how it looks under this light or that its not the dark blue or light blue but could be if there isn't a dark or light blue? All common English words are like that. Honestly infinity is an easier concept or it is "all".

  • $\begingroup$ Which 8 senses are you narrowing down to? $\endgroup$ – trichoplax Dec 31 '14 at 13:14
  • $\begingroup$ Is blue the hypothetical colour larger than all others? The link with infinity is not clear to me. Could you expand? $\endgroup$ – trichoplax Dec 31 '14 at 13:15

an Ai would naturally struggle with anything that depends on emotions. some methods can be employed to ease relations("friendly phrases",expression detection), but eventually it will still come off as too rigid or mechanical. but an Ai is a program,and even the most advanced don't have the concept of feeling like living things do. depending on the circumstances,this can be a good or a bad thing.


How do you program an AI to have ambition? An AI can reasonably be creative, as creativity is as much the combination & recombination of other concepts as anything else, but to have any kind of purpose that drives it to improve or excel?

Would an 'ambitious' AI just be a digital liberal? determined to change things for change's sake, or could it derive its own purpose?

  • 1
    $\begingroup$ You’re saying that AI will not be good at ambition because you personally don’t know how to make an AI? You could pick any random trait and simply state that AI are not good at that. $\endgroup$ – JDługosz Mar 29 '17 at 5:35
  • $\begingroup$ That reads like it's supposed to be a criticism. There are two broad categories of 'flaw' for AI and both depend upon the choices of the creator. Deliberate choices require appreciation of options.. Or in other words "if one does not know fully how the human and the synthetic works..predicting which will be better at which is something like a shot in the dark. Accept that it's a terrible answer in the sense that it's leading, rather than explicitly an answer in itself. :) $\endgroup$ – mensenisevirem Mar 29 '17 at 11:20
  • 1
    $\begingroup$ "How do you program an AI to have ambition?" — Define ambition. I'd say that a standard chess computer has the ambition to win. $\endgroup$ – celtschk Apr 28 '17 at 6:15
  • $\begingroup$ You can say that, if you consider that the sum of human motivation towards the achievement of goals is really comparable to that of a contemporary chess program, we may as well delete the word ambition from our dictionaries. $\endgroup$ – mensenisevirem Jul 24 '17 at 0:30


I don't foresee machines ever being able to do anything that requires creative input, such as writing literature or music, sculpting, making paintings, or even doing things like singing or dancing in a way that truly moves people emotionally. Those things are the product of a mind, and a mind is much more than physical matter.


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.