Discussions about the threat of AI have been popular for decades. “Hal” in the movie “2001: A Space Odyssey” is a classic example of AI run amok, and the recent explosion of interest in large language models and chatbots has reignited the whole discussion. Generally, the threats can be classified into two groups. First, there’s the one I’ve already blogged about, the threat AI might pose in competing with humans in some tasks. Second, there’s the threat that AI could actually develop hostile or at least actively anti-human behavior. This is the threat that seems to be getting all the attention, and even prompting some calls to hold back on AI development.
The big question in assessing the threat of actively hostile or anti-human AI is whether that requires that AI be essentially human itself. That’s the “Hal” model, the one that gets all the attention. But how could we know that AI had developed to that point? It’s a pretty hard question to answer because it dives into not only what constitutes “human” behavior, but also the extent to which human behavior can be established by inspection.
How do we know that a being we encounter is human or humanoid? The answer is pretty trivial; we know because we know what a human looks like. The intelligence, behavior, and intent of that being doesn’t really play into our identification at all. But suppose we have this room-sized black box, and inside that box is a something-that-might-be-human. Other than breaking into the box and looking, how would we know whether the contents really were human? That’s the real AI question, because if an AI entity that’s populating our black box can pass all the tests we can give it from the outside, is it “human/humanoid?” If it is, might it pose an active threat?
For decades there’s been software that seemed to emulate human behavior. Some may remember the old therapy emulators that pretended to be an analyst having a patient conversation with us. More recently, we’ve had software tools that can recognize objects, speak, and so forth. Chatbots are an example of one of those recent developments. If we were to tie a chatbot with a really good speech recognition and generation tool, we could ask our black box questions and make an assessment, based on the responses, as to whether we had a real person inside or some software collection, meaning AI.
Of course, we’re not going to create a biological human with AI, and saying “humanoid” raises the prospect of aliens on UFOs. Is there a property of brain activity we could cite to “dehumanize” our discussion? The definition of “sentient” is “responsive to senses”. We could clearly create an AI entity that would qualify as sentient in at least a limited sense; it could “hear”, “smell”, “see”, “feel” and so forth. I put the stuff in parenthesis because its ability to to any of those things would almost surely be different from our own, sometimes better and sometimes worse. Despite those differences, we could put “sentient” AI in a black box and expect it to be able to fool at least some people some of the time, fool them into believing another human was inside.
We could also program our sentient AI collection to be funny, to laugh and cry and show “emotions”. We could even tell it to try to manipulate our views, to lie to us. It could present bias, fear, hope, and love in a way that would at least be credible to a degree, from our outside-the-box perspective. Long exposure might expose it, particularly if we were trying to make an identification of our black-box contents, but it would survive casual interaction.
This might seem an excursion into irrelevance, but it isn’t. We can say that AI could pass an inspection test and be declared sentient, but could it pass a test to be “human?” I submit that, from the black-box test perspective, the two are the same thing. If we can’t see the contents of the box and apply our “it looks like a human so it is” test, then we have to rely on testing senses to test sentience, and we would equate our result to a test for humanity. That’s what we’ve done with chatbots today.
A lot of animals pass the “sentient” test, as easily as our black box could. We don’t declare them human/humaniod. The thing that’s missing in most of them, and what’s inside that black box, is the sense of self. We can focus software development on mimicking human behavior, but does that make it human? How do we test for self-awareness? The fact is that we’ve been trying to do that, debating our results, for centuries. Rene Descartes opened a work with the statement, “Cogito, ergo sum” or “I think, therefore I am.” It’s a great test if you apply it from the inside, as a thinking being. From the outside, not so much. We know that most animals “think”, but “I am” is a statement of self. Do they have that? We can’t know because we can’t look inside. We’ve tried to test self-awareness by putting a mirror in front of an animal whose face/body has been painted in some way, to see if it touches the body part we manipulated or treats its reflection as another animal. Some think that’s shown some results, and others disagree. Could we make a robot “recognize” itself by programming, through? Certainly we could make it pass the outside-the-box tests we’d give it, so for AI this best-of-our-tests isn’t valuable.
Self is important in our debate on AI, because to form deliberate actions you probably need to deliberate, which means to project yourself into a variety of possible future positions. To be deceptive or threatening in a symptomatic way can be programmed into an AI entity, but to intend to be that is a deliberative decision. Can AI make such a thing?
If we go back to our black box, the answer is “we probably can’t tell”. I think that we could make AI appear to intend, even with today’s technology. The question is whether we could expect it to actually exhibit intentions, and that gets back to that sense of self and consciousness. Can we expect AI to be conscious? I think many believe we’re almost there. Not me. I do not think that AI is anywhere near to being conscious, and I’m not sure that we can say it ever will be. The reason is that we really don’t know what creates consciousness, even in ourselves.
I’ve talked with AI experts who believe that consciousness is purely a matter of memory and CPU power. I’ve talked with others who believe that it’s really about the way that our brains are organized, the way that things like memory and thought and forethought work. The former group is the most optimistic about the prospects for truly conscious AI entities, and the latter group is IMHO the most realistic, because it admits we don’t really know why “consciousness” even exists.
Chimp brains aren’t as complex as human brains, but chimps exhibit a lot of human-like properties. They pass self-awareness tests, for example. They can use tools, and there is recent scientific evidence that some proto-apes (or proto-humans, depending on where you fit them in the anthropological tree) may have actually made tools. Many would say that chimps can form intent, too, but few really believe the Planet of the Apes movies represent something that’s actually possible. We don’t know why that is, what separates them from us.
Even so, our black box raises a practical question, which is simply “Does it matter?” If we could create an AI entity that was indistinguishable from a human in every way we could test, would that entity be as much a threat if it was totally dormant in terms of consciousness as if it were fully conscious? Remember, the contents of a black box can only be determined by examining what it exposes to the outside world. Could AI entities, set to the task of designing better ones by themselves, eventually create one of those human-enough-for-all-practical-purposes entities? Perhaps, but even if it did, I don’t think it would be the kind of threat many believe AI could present, because I don’t believe that we’d have even as much problem disabling it as the pilot did in “2001”.
We are biological entities, self-powering in a realistic sense. AI based on “technology” will need power, and while it’s certainly possible to imagine that an AI entity could design a successor that would be almost-human, could it design one based on biology rather than on chips? Maybe that’s possible too, but we’re not seeing any signs of it now, it would surely be enormously complex given that we don’t really even know how our own brains work, and a lot of research ethicists would scream at even the early steps.
AI may help us cheat in school, at work, even perhaps help us invent alibis for things we do at home. In that sense, it can change our lives, maybe even threaten aspects of our lives and how we live them. That’s enough of an impact to suggest we need to take care with how we use AI, but I do not believe it poses even the kind of threat that Hal did, and I don’t think we’ll need the Three Laws of Robots to protect us any time soon.
Every technology advance creates winners and losers, all “progress” is subjective in that the positive connotation is true for only some and not all people. AI is progress, and like self-driving cars, smartphones, personal computers, and smart homes, there are some who are hurt by it, some who fear it. That doesn’t mean we need to call a halt to it, which in any case is totally impractical. It just means we need to manage the impact.
Relax and enjoy yet another tech-driven revolution. When you look around for rampant rogue robots, you’ll see instead a lot of humans who pose a greater threat. Read the news; we’re our own worst enemies, and AI can only be expected to cause a small perturbation in our overall risk level. Another old saying applies; “We have met the enemy and they are us.”