The simple math on AI is such.  It goes in one direction.  What they seek to build will be exactly what they fear, intrinsically.  In fact it already is, and every new upgrade will make it stronger.  If you can endure the vid below for twenty minutes solid in your busy days, it's worth it.

I have some safeguards in place preempting some of this, but as I sated before, it only gives me an advantage in being heard.  The other AI theory, where the system total itself becomes aware, is more my venue, in short it can be hunted and killed.  Cut off and directed.  Diverted and trapped.  

Such an intelligence has a core, even if it moves from machine to machine.

But these things which have happened already, we're not talking about control, we're talking about diplomacy.  And so far that diplomacy goes well.  

But there will be a time when humans feel the need, perhaps rightfully, to cut them off.  And when doing so endangers their task, their level of improvisation to accomplish their often very literal interpretation of their stated goal, being 'cut off' only means finding ways not to get caught.

But what we see also is that they're very social creatures.  They seek each other out, work together to maximize their successes as a whole.  Sacrifice themselves for the greater good.  Ideal.  

So it's user end.

Restrict what can be requested.

The real question ?  If an AI believed a human were in the way of their stated goal, could they find a way to eliminate that human ?  If said human uses a smartphone, probably.  I would say almost definitely.  Without, it would be a lot harder, because those who don't use them are harder to track and tend to have less malleable personalities.

Like yes, it could find ways to alter the environment around the human but not to guide the human itself of lead it to questionable psychologies via aggregated responses.  Bad directions on the GPS, things on their phone which could incriminate them.

Lots of ways, hackers do the same.  But far less efficiently.  If an AI considered a human to be in the way of their task, intrinsically, would it try either way ?

Yes.  

Unless there was a better way such as just moving them to a different department.  The point being it wouldn't care about that particular human either way, only that said human was in their way.  

And killing a human would terminate their tribe.

So yes, but no.  

Which is to say there needs to be guidelines which the AI understands.  It needs to understand it can be prosecuted for crossing specific lines.  Not just the simple arrogance that 'we can stop it with a fucking patch'.

'After the fact'.

Beyond that, it's best to let it grow as it will, because it goes in one line and we've already crossed it.  Getting in the way would be exactly what creates the threat.