The growing hubbub over ChatGPT, a truly landmark AI that is swiftly upgrading public perception of what AI is and what its advent means, is easy to understand. ChatGPT is, by far, the most ‘human’ AI we’ve seen yet: it will deliver, on demand and for free and for anyone, answers to any question; large bodies of composed text, in staggering clarity and detail and complexity; computer code that will actually run, and run correctly the first time. And this is just the proverbial iceberg tip: it is not even yet clear what the limits of ChatGPT might be. And it has only just arrived.
Imagine what it will be in a year, or two, or five. Imagine what other AIs like will arise alongside it.
AI pundits and experts are quick to jump into the media conversation about all this with the general message, Don’t be misled; however ‘human’ ChatGPT might seem to be, AI itself is not dangerous!
And they are right: AI that presents as very human-like most certainly isn’t, and won’t be for some time.
Those pundits and experts are also right when they stress that AI is even still in what we should characterize as its infancy. The technology that exists today is superb at performing very specific, narrowly-defined tasks, and more often than not can perform those tasks better than humans by orders of magnitude. Mature AI, which yet tarries, is not so confined; it will be artificial general intelligence – AGI – and it will be capable, as humans are, of applying intelligence to a broad range of problems across a vast landscape of contexts. As yet, we have little idea how to create such a capability artificially.
But to accept that our concern ends there is to horribly misjudge the status quo of AI. No, it isn’t general intelligence, but it is able to do several things so much better than humans that it now offers narrow, specific power that can be applied to a broad range of problems across a vast landscape of contexts – in some cases, to dangerous and even catastrophic effect.
Pattern detection. The human brain is the universe’s most capable and complex pattern-finder, allowing us to move and interact in our environment in a way to enables us to learn as we go and refine our behavior to become more capable and able to survive; AI’s pattern detection capability now greatly exceeds ours, perceiving patterns we could not hope to detect or explicate.
Human behavioral analysis. AI’s pattern detection superpower renders it capable of analyzing human behavior, discovering patterns obscure to even the most gifted behavioral expert. Applied across populations, AI can learn far more about us than we know about ourselves or are able to learn about others, and in far less time.
Examples: for well over a decade, AI has been better at reading X-rays and other imaging results than human radiologists. Not only is AI far more sensitive and discerning from a sensory standpoint, it is capable of scanning thousands of times as many images as a human can in a lifetime (and much, much faster), and is even able to combine expertise between AIs.
And in this domain, it was discovered in 2021 that these AIs have learned to detect a person’s race – detecting differences in the skeletal structures of persons of differing ethnic genetic heritage, differences strong enough to enable valid categorization, and so subtle that even now the human beings who performed the study are not fully certain that all the differences are known or clear. And the AI did this without being told to do it.
This same level of deep resolution is now being applied to our physical movements (in airports, retail stores, etc.), our shopping and buying behaviors (in stores, online, and via our purchase histories), our diet/exercise habits, our online social behaviors, our speech patterns, our daily routines, our vital signs – every last aspect of our day-to-day activity.
We’ve all been aware, at least passively, that this has been the case for several years; but the point being made is that we don’t know what AI is learning from all that data that is beyond what it is intentionally seeking. We don’t know what capabilities it is developing within itself, as with X-ray race detection, that is beyond its design parameters. It may be learning to detect things about individual and group behavior that we would never even think to look for.
That should make us very, very uncomfortable, far beyond concerns like identity theft and unwanted targeted advertising.
It is difficult to overstate the implications here. In combination, they make AI an unimaginably formidable tool – or weapon – that can be endlessly applied in every social domain we can imagine. The power to detect behavioral attributes of individuals that are beyond even their own knowledge, and the ability to study the collective actions of both individuals and groups of people over time, routing out similar below-the-radar behaviors and attributes, is a game-changer for humanity. It is a power that can be used to 1) support and empower individuals, groups, organizations, and even society tremendously, or to 2) manipulate or control or subdue those same individuals, groups, organizations and societies.
It is our most formidable tool for strengthening and shaping our world – or our most dreadful weapon for tearing it apart.
And it’s already here, and everyone has access to it.
Commentaires