17
$\begingroup$

With the increasing complexity of reCAPTCHA, I wondered about the existence of some problem, that only a human will ever be able to solve (or that AI won't be able to solve as long as it doesn't reproduce exactly the human brain).

For instance, the distorted text used to be only possible to solve by humans. Although...

The computer now got the [distorted text] test right 99.8%, even in the most challenging situations.

It seems also obvious that the distorted text can't be used for real human detection anymore.

I'd also like to know whether an algorithm can be employed for the creation of such a problem (as for the distorted text), or if the originality of a human brain is necessarily needed.

$\endgroup$
10
$\begingroup$

Informally, AI-complete problems are the most difficult problems for an AI. The concept is not mathematically defined yet, as e.g. NP-complete problems. However, intuitively, these are the problems that require a human-level or general intelligence to be solved.

Real natural language understanding is believed to be an AI-complete problem (this is also discussed in the paper Making AI Meaningful Again by Jobst Landgrebe and Barry Smith, 2019). There are a lot more AI-complete problems. For example, problems that involve emotions.

$\endgroup$
  • 3
    $\begingroup$ NP-complete problems are called 'complete' because they are the hardest in a precise sense: that any other problem in NP can be reduced to one of them by a deterministic, polynomial-time algorithm. Therefore an oracle that could solve any one of the NP-complete problems in polynomial time could be used to solve every problem in NP in poly time. Is there any sign that your notion of AI-completeness makes objective sense in a similar way? $\endgroup$ – Mike Spivey May 5 at 16:54
  • 1
    $\begingroup$ @MikeSpivey This is a good question. I've not read an extensive amount of related literature, but, in the paper Towards a Theory of AI Completeness that I link you to in my answer, the authors state "Those measures can also tell us how much of the problem we already managed to delegate to a computer; using reductions, progress in one problem can translate to progress in another". So, in that paper, at least, there is the desire to develop a formal complexity theory for AI. $\endgroup$ – nbro May 5 at 17:10
1
$\begingroup$

This is more of a comment and philosophical opinion, but I don’t believe that there are any problems an AI couldn’t solve, that a human can. Being new to this forum, I cannot make it a comment on the question (and it would probably be too long) — I preemptively ask for your forgiveness.

AI Eventually Will Mimic Humans (and surpass them)

Humans by nature are logical. Logic is learned or hardwired, and influenced by observation and chemical impulses.

As long as an AI can be trained to act like a human, it will be able to act like one. Currently, those behaviors are limited to technology (space, connections, etc), which the human brain has been optimized to rule out or disregard certain “fluff” automatically enabling it certain super capabilities. For instance, not everything seen is registered through the brain; often, the brain performs differential comparisons and updates for changes to reduce processing time and energy. It will only be a matter of time before AI can also be programmed to behave this way, or technological advancements will allow it to not need some of this function, which will allow it to leapfrog humans.

In the current state, we recognize humans are sometimes irrational or inconsistent. In those cases, AI could mimic human limitations with configured randomization patterns, but again, there really won’t be a need since it can be programmed and learn those patterns automatically (if necessary).

It all comes down to consumption of data, retention of information, and learned corrections. So, there is no problem that a human can perform (to my knowledge) that AI couldn’t theoretically ever perform. Even in the case of chemistry. As we are manufacturing foods and organs, an AI could also, theoretically, one day reproduce and survive via biological functions.

Instead of the question being binary about human capability vs that of artificial intelligence, I’d be more interested to see what people think are the more challenging things humans can do, which will take AI time to accomplish.

$\endgroup$
  • $\begingroup$ Hi vol7ron. I downvoted this answer because it is full of speculations, thus I don't think it is a very useful answer and I don't agree with a lot of your statements. Do not take it personally. Anyway, AFAIK, you can comment, because you have more than 50 points of reputation. $\endgroup$ – nbro May 6 at 14:20
  • $\begingroup$ @nbro comments have length limitations and I could only comment after posting because I was not a member, I receive the +100 after. $\endgroup$ – vol7ron May 6 at 14:25
  • $\begingroup$ You receive 100 points of reputation once you create your account, AFAIK. However, I would like to point out a few of your statements that I don't agree with and I think are inconsistent or ambiguous. First, you say "Humans by nature are logical". What do you mean by logical here? Then you say "In the current state, we recognize humans are sometimes irrational or inconsistent". It seems to me that your arguments are not very consistent and thus contradictory. Furthermore, your main point is: AI will be able to do whatever humans can do because I have no proof against it: it is a poor argument. $\endgroup$ – nbro May 6 at 14:29
  • $\begingroup$ @nbro thank you for your feedback! I would say that the heart of my answer addresses the question — “no” is the answer. It is not purely on speculation, but I would argue that the question was inviting speculation. The title says “will ever be able to”. How can anyone predict the future? It is speculative. My speculation is founded on track record and capabilities within the bounds of AI operations. It considers current limitations that prevent acting like humans in their current state and the speculation is that won’t be an issue in the future. $\endgroup$ – vol7ron May 6 at 14:30
  • $\begingroup$ Given that I don't think that speculations are, in general, useful, I down-voted. I would not have down-voted if you had been more consistent and had argued your statements. For example, I don't even understand what you mean by "learned corrections" in the statement "It all comes down to consumption of data, retention of information, and learned corrections". I would also like to note, in case you are not aware of it, that not everything can be inferred from data (that is, ML is often insufficient: e.g. it can't detect causal relations, but only correlations). $\endgroup$ – nbro May 6 at 14:36

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