Slashdot Mirror
What Does Artificial Intelligence Actually Mean? (qz.com)
An anonymous reader writes: A new bill (pdf) drafted by senator Maria Cantwell asks the Department of Commerce to establish a committee on artificial intelligence to advise the federal government on how AI should be implemented and regulated. Passing of the bill would trigger a process in which the secretary of commerce would be required to release guidelines for legislation of AI within a year and a half. As with any legislation, the proposed bill defines key terms. In this, we have a look at how the federal government might one day classify artificial intelligence. Here are the five definitions given:
A) Any artificial systems that perform tasks under varying and unpredictable circumstances, without significant human oversight, or that can learn from their experience and improve their performance. Such systems may be developed in computer software, physical hardware, or other contexts not yet contemplated. They may solve tasks requiring human-like perception, cognition, planning, learning, communication, or physical action. In general, the more human-like the system within the context of its tasks, the more it can be said to use artificial intelligence.
B) Systems that think like humans, such as cognitive architectures and neural networks.
C) Systems that act like humans, such as systems that can pass the Turing test or other comparable test via natural language processing, knowledge representation, automated reasoning, and learning.
D) A set of techniques, including machine learning, that seek to approximate some cognitive task.
E) Systems that act rationally, such as intelligent software agents and embodied robots that achieve goals via perception, planning, reasoning, learning, communicating, decision-making, and acting.
A) Any artificial systems that perform tasks under varying and unpredictable circumstances, without significant human oversight, or that can learn from their experience and improve their performance. Such systems may be developed in computer software, physical hardware, or other contexts not yet contemplated. They may solve tasks requiring human-like perception, cognition, planning, learning, communication, or physical action. In general, the more human-like the system within the context of its tasks, the more it can be said to use artificial intelligence.
B) Systems that think like humans, such as cognitive architectures and neural networks.
C) Systems that act like humans, such as systems that can pass the Turing test or other comparable test via natural language processing, knowledge representation, automated reasoning, and learning.
D) A set of techniques, including machine learning, that seek to approximate some cognitive task.
E) Systems that act rationally, such as intelligent software agents and embodied robots that achieve goals via perception, planning, reasoning, learning, communicating, decision-making, and acting.
To begin with, referring to "human intelligence" is pointless as we do not agree on what this is. Including "rational thinking" as part of the definition won't help either since the process of asserting rationality is non trivial. "To think as humans" and say that all artificial neural networks does this is to insult neurologists. It might work to say that neural networks are loosely inspired by how we think human brains work.
However, I do like if the definition include a metric for how the system can adapt and learn. Just don't confuse the two. Neither adaptation nor learning says anything if the system is about to converge to a stated policy. Sometimes you don't want adaptation but forceful behavior. Sometimes you realize your policy is crap but you don't want the system to cheat anyway. So are these attributes even useful to consider when you want to regulate AI (or human intelligence for that matter).
Why would the regulation of AI be any different from the regulation of human intelligence? Let the AI prove itself, then hold it or its "parents" accountable.
Back in high school in my computer programming class, we were taught arrays, to do this we made the game of memory where we had 16 cards with 8 matching values, which were randomized.
Then we were to pick 2 cards if we got a match we would had got a point. Then the computer picked two cards.
Normally most of the students just had the computer pick randomly. I felt ambitious as programming was my thing that made me the Alpha geek back then. So I made it keep track of the cards when it found them and learned from its mistakes, thus being a difficult game to play.
This isn't AI, but it seems to fit definition A. as would most video games of any challenge. Also most business intelligence apps that find patterns would classify.
If something is so important that you feel the need to post it on the internet... It probably isn't that important.
So let's discuss the why before we just start regulating stuff that 99.999% of the time will not need any regulation for any public safety, or even ethical purpose.
What is the purpose of regulating computer software? AI in most cases these days means computer software that has been trained with examples to process a data set rather than programmed to process one. It is just more efficient than figuring it out and programming an algorithm directly for more variable input. And once the training is over and optimized the algorithm is usually frozen so that it can be applied in a tested and predictable way. So AI is rarely about algorithms that are trained during production use.
And this part of the proposed definition makes it a blanket definition for all computer software not just AI: "Any artificial systems that perform tasks under varying and unpredictable circumstances, without significant human oversight".
So really hard to see how you regulate "AI" without a blanket regulation on all software development.
If we are talking about simulating complete multi-functional animal brains, especially human, then I think ethics do come into play. Perhaps our discussion should focus on that as something that should be regulated.
I think we have an societal interest in working to prevent the abuse of animals and people. And it could be that at some point, maybe very soon, we can effectively simulate a human or large animal brain and even good people might fail to realize the real perception of suffering, real suffering, they are causing in a thinking being stuffed into a computer.
That said, do we really want regulations preventing AI from becoming more like us? Is this inherently wrong? As every parent is acutely aware suffering is part of life and learning and we feel for our children because we have been there and understand how hard it can be. It is hard to imagine the human brain learning without negative feedback, without at least some bare minimum of physical and emotional pain.
Is the greater good in preventing any suffering or just limiting it to what is absolutely necessary for us to learn? It seems preventing all suffering is no different than preventing life. And allowing suffering more than what is necessary for life is also wrong.
Is there a golden mean between these extremes? And can that be regulated through the force of government?
IMHO, self-aware just means being able to examine some internal states and store or report that information.
"Thermostat, what's the current temperature?" :: "72 degrees" - that's self-aware. It just isn't much of a self.
Even adding "able to modify internal states based on examining them" is something more than self-awareness.
Self-aware is not the same as "conscious." Consciousness implies assigning internally conceived meaning to, and abstract manipulation of, such states.
I've fallen off your lawn, and I can't get up.