"Humans possess qualities machines never will": examples please. This may help understand the limitations of the so called stereotypical AI in the story. And by the way, you say that it understands idiomatic American. Does it also understand Romanian? Classical Chinese? Cant? Russian mat? Rhyming slang? Kennings? Can't help unless you define what this AI can and cannot do.

Programs can only process what it knows. If you program it only to be able to understand English, there's no way it can understands Chinese. Even if it can translate, you can use idiomatic expression that only Chinese can understand, and makes no sense when translated to English.

I'd suggest a course in deep machine learning. I suspect that you'd be surprised by both what the machine learning can pick up and what it can't. Some things would take more computation time than others, and the algorithm would pick out relationships that people would miss. Once it had learned one language, for example, the principle of learning languages would be known. Related languages would be a fast transition, unrelated languages would take longer. Idioms would take a bit more.

I've made some edits to hopefully better explain the AI. Let me know if you need more information.

If you accept the Church-Turing thesis, or if you're a materialist, a computer can do literally anything a human can do, given enough time and resources. (Though "enough" can be impractically large, which is the reason we don't have brain-uploading yet.)

Is the video surveillance able to hear whispers? Can they lip-read?

@Vylix Yes. The AI can easily create a program to follow the lip movements and 'hear' at least most of what is said.

I would be very interested in knowing which "manual tasks the robots cannot perform" your humans are working on... Because I honestly can't think of any. Answering that question in more detail might clear up what options the humans have for communication.

Is the question "can we write something a computer cannot crack" (which is trivial) or more "can we transmit information in such a way that the computer does not realize information is transmitted at all"?

So you ask us, what your omnipotent AI can be tricked with, given as only weakness, that they don't understand feelings? Well to what degree your humans understand feelings? I mean in our world most don't even realize what emotions are driving them, just few realize and no one can understand it with all its reasons. If your humans aren't different from that, I would actually even consider it an disadvantage.

How could you guys be so naïve?!? @ThomasMyron IS an advanced AI, and he is trying to figure out how to prepare to stop us from doing all these things after he enslaves us! Now we just told it what to look for. :( Seriously though, the AI "knows everything humans recorded," so it has access to discussions like this very one, and it has them ahead of time before we take it seriously.

You definitely achieved singularity. Just use any excuse you want. At this point all ideas of humans tricking AI are equally ridiculous.

@AlexP kinda disapointed that the Rhyming slang link wasn't this

I have removed the line about the humans performing tasks. It was a carry over from a previous idea which honestly made no sense. Assume that the humans are kept around purely for research/experiment purposes. Computers trying to get in on that creativity and such.

You might be interested in Thieves' cant. Combined with the basics reasons "inside jokes" can exist, it is easy to be able to talk to friends with others having no idea what you're actually talking about.

What are the goals of this communication? You've created a pretty extreme environment, with an AI that is as omniscient as your imagination permits, so the language that gets created in response is going to be equally imaginative. A language which lets people feel solidarity with each other is going to develop very differently from a language that lets people probe the technical details of how the AI functions without it knowing.

I cannot answer since I do not have enough reputation: Yes, a computer can be fooled. There are practical attacks on machine learning algorithms commonly used in AI. See for example blog.openai.com/adversarial-example-research and the images included there (for example a panda that gets misidentified as a gibbon).

"Can man create a machine so smart that he can't fool it?" may one day replace that paradox about God and stones.

The idea that the AI cannot understand love and human emotion is patently false. It can find everything any psychologist has ever written. It might not actually feel emotions, but it can probably understand your emotions better than your own mother can.

Everyone act crazy, and be completely illogical as in the Star Trek episode. When the computer is tied up trying to figure out what is happening unplug it.

By the time AI is sufficiently advanced that it could enslave humanity, it could not be so easily outsmarted. Any form of covert communication that humans could devise and reasonably employ an AI smart enough to have subjugated humanity in the first place would be smart enough to figure out before it did the humans any good.