Is AI getting out of hand?

I think there are some classifications that can be useful regarding AI (but my TL;DR is that they were a problem 6 years ago, and it has gotten worse).

• There is an area called Narrow AI. These have their problems, but aren't what's on the media nor the most dangerous. However, they had one benefit to companies that current, more general purpose models also have: law falls behind of making companies accountable for their output. This includes tools from spam filters, recommender systems (e.g. for adds and feeds), image filters; and even alpha fold (all of them) or computer vision for cars can be considered narrow because their training data is restricted to a task.
• Another area, is General-purpose AI (GAI). In some cases, the architectures are similar to narrow AI, but their training style and dataset is more general (for example, all the text, including computer code, and images of the internet).

Within GAI there are a diverse range of opinions, ranging from "they are stochastic parrots" to "they are conscious and super intelligent". 

There are articles showing that sometimes they don't quite learn general concepts but heuristics (also papers opposing these). As said, it's still debated and does not mean they don't have capabilities (they do, and are a problem).

In other words, how general their intelligence is, and whether they could be conscious, isn't in my view a fully settled research problem.

Some consider their overt behaviour or actions and capabilities as the whole picture. Then, if the machine says it feels something, does it really? If instead of the computer (with its CPU and electrical circuitry and components), the computation was performed by a system of mechanical pistons, would you consider this artefact (which is equivalent to the computer) to feel something?

That's a derivation of Searle's Chinese Room Argument, but there are more compelling or updated ones as well.

Those questions are hard to answer, and I find the "It's this way!" (especially by CEOs of billion-dollar companies) isn't trustworthy, and they also are not the top experts.

However, to the more mundane question of: can they cause a lot of disruption (e.g. job market) and trouble? The answer is probably yes, as we have already seen. The problems that already existed with Narrow AI (e.g. accountability, decision transparency and explainability) have been expanded by the increase in capability of the systems. For example, with a simple prompt you can hack certain systems, get a recipe to create bio-weapons or do all sort of other things.

I think there are some classifications that can be useful regarding AI (but my TL;DR is that they were a problem 6 years ago, and it has gotten worse).

• There is an area called Narrow AI. These have their problems, but aren't what's on the media nor the most dangerous. However, they had one benefit to companies that current, more general purpose models also have: law falls behind of making companies accountable for their output. This includes tools from spam filters, recommender systems (e.g. for adds and feeds), image filters; and even alpha fold (all of them) or computer vision for cars can be considered narrow because their training data is restricted to a task.
• Another area, is General-purpose AI (GAI). In some cases, the architectures are similar to narrow AI, but their training style and dataset is more general (for example, all the text, including computer code, and images of the internet).

Within GAI there are a diverse range of opinions, ranging from "they are stochastic parrots" to "they are conscious and super intelligent". 

There are articles showing that sometimes they don't quite learn general concepts but heuristics (also papers opposing these). As said, it's still debated and does not mean they don't have capabilities (they do, and are a problem).

In other words, how general their intelligence is, and whether they could be conscious, isn't in my view a fully settled research problem.

Some consider their overt behaviour or actions and capabilities as the whole picture. Then, if the machine says it feels something, does it really? If instead of the computer (with its CPU and electrical circuitry and components), the computation was performed by a system of mechanical pistons, would you consider this artefact (which is equivalent to the computer) to feel something?

That's a derivation of Searle's Chinese Room Argument, but there are more compelling or updated ones as well.

Those questions are hard to answer, and I find the "It's this way!" (especially by CEOs of billion-dollar companies) isn't trustworthy, and they also are not the top experts.

However, to the more mundane question of: can they cause a lot of disruption (e.g. job market) and trouble? The answer is probably yes, as we have already seen. The problems that already existed with Narrow AI (e.g. accountability, decision transparency and explainability) have been expanded by the increase in capability of the systems. For example, with a simple prompt you can hack certain systems, get a recipe to create bio-weapons or do all sort of other things.