Contending with AI flaws such as biases, errors, and other issues that will be intertwined with the … [+]
In today’s column, I examine a prevailing assumption that attaining pinnacle AI, such as artificial general intelligence (AGI), will be fantastic since the AI will be absent of imperfections and represent idealized perfections of intelligence. That’s a myth that seems to be commonly accepted and taken as unflinchingly true. The reality is that AGI will almost certainly be unavoidably replete with intelligence imperfections of the likes of human intelligence. This is harsh news for those who are expecting an upcoming intelligence nirvana.
Let’s talk about it.
This analysis of an innovative AI breakthrough is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).
The Pursuit Of AGI And ASI
There is a great deal of research going on to further advance AI. The general goal is to either reach artificial general intelligence (AGI) or maybe even the outstretched possibility of achieving artificial superintelligence (ASI).
AGI is AI that is considered on par with human intellect and can seemingly match our intelligence. ASI is AI that has gone beyond human intellect and would be superior in many if not all feasible ways. The idea is that ASI would be able to run circles around humans by outthinking us at every turn. For more details on the nature of AI, AGI, and ASI, see my analysis at the link here.
AI insiders are pretty much divided into two major camps right now about the impacts of reaching AGI or ASI. One camp consists of the AI doomers. They are predicting that AGI or ASI will seek to wipe out humanity. Some refer to this as “P(doom),” which means the probability of doom, or that AI zonks us entirely, also known as the existential risk of AI.
The other camp entails the so-called AI accelerationists.
They tend to contend that advanced AI, namely AGI or ASI, is going to solve humanity’s problems. Cure cancer, yes indeed. Overcome world hunger, absolutely. We will see immense economic gains, liberating people from the drudgery of daily toils. AI will work hand-in-hand with humans. This benevolent AI is not going to usurp humanity. AI of this kind will be the last invention humans have ever made, but that’s good in the sense that AI will invent things we never could have envisioned.
No one can say for sure which camp is right and which one is wrong. This is yet another polarizing aspect of our contemporary times.
For my in-depth analysis of the two camps, see the link here.
Focus On AGI As A Possibility
Let’s momentarily put aside the attainment of ASI and focus solely on the potential achievement of AGI, just for the sake of this discussion. No worries — I’ll bring ASI back into the big picture in the concluding remarks.
Imagine that we can arrive at AGI.
The path might be by incrementally advancing conventional AI. Another route might be that the AI spontaneously cycles within itself and uses its budding intelligence to fuel even more intelligence until it reaches a point of AGI. This kind of intelligence explosion is often referred to as the AI singularity; see my coverage at the link here.
So, by hook or crook, assume we do arrive at AGI. Please note that there is no ironclad guarantee that we will. There is a chance that we will end up on a high and dry plateau and not attain AGI. It might be close, but no cigar.
To be abundantly clear and upfront, absolutely no one can say for sure when or if we will land on true AGI. Boom, drop the mic.
Shining A Light On Alleged Intelligence Perfection
A prevailing belief by many AI proponents is that AGI will consist of the best of human intelligence. In a sense, this will be a form of perfect intelligence. What is meant by perfection? The usual consideration is that AGI will be free of any intelligence flaws or weaknesses.
This raises a number of significant questions.
First, human intelligence is categorically replete with flaws such as human biases, human error, and irrationalities. You can see this plainly exhibited with your own eyes. Cognitive scientists, psychologists, behaviorists, and the like have studied these issues and well-documented the problems that these qualms present (see my elaboration at the link here).
Second, as far as we know, intelligence-based imperfections are inextricably enfolded into the upsides of intelligence. There is no seemingly viable way to separate the weaknesses from the strengths and still have a cohesive and coherent remainder. The viewpoint is that the yin and the yang of intelligence are inseparable.
Third, a general theory is that without imperfections in intelligence, we wouldn’t have the innovations and bursts of brilliance that we otherwise seem to muster. The rationale is that intelligence spurs creativity due to the said-to-be weaknesses of intelligence. If you did away with the flaws, a presumably bland and uninspiring semblance of intelligence would be the thing that remained.
In brief, flaws in intelligence are not a bug, they are a feature.
Building AGI So That It Is Perfect
A counterargument to all those dowdy remarks about how intelligence is flawed consists of claiming that artificial intelligence doesn’t have to suffer the same fate as human intelligence.
The logic goes like this. Sure, human intelligence is a mess. No doubt about it. But we can devise or cook up AGI in such a fashion that it doesn’t carry those intelligence maladies. All we need to do is recognize that human intelligence has bad spots, and we can then deal with those while building AI.
Period, end of story.
Though that sounds quite reassuring, there is little evidence that what is proposed can indeed be successfully pulled off.
Let’s do a deep dive.
Generative AI and large language models (LLMs) are currently crafted by pattern-matching human writing. The AI is data-trained by scanning the Internet for human-composed essays, narratives, poems, and so on. Based on the massive size of the available data, the AI computationally patterns on how humans utilize natural language and mathematically mimics the same appearance of fluency.
The gist is that the patterns of human biases, human errors, and the like are part and parcel of that vast dataset and, ergo, part and parcel of the pattern matching that is taking place. Human flaws associated with intelligence are flowed into the AI and become infused in that computational mosaic.
Thus, the flaws are baked into the whole kit and kaboodle.
Attempts to beat back the bad sides of such AI are ongoing and generally so far of limited success; see my analysis at the link here. One method involves yanking out the flaws. This is not an easy task and is unlikely to be entirely exhaustive and complete. Can truly all possible flaws or weaknesses be identified? Even if they can, what will the AI be like once those are all removed?
Another angle is to mask or hide the flaws.
The flaws still exist within the inner core of the AI, but you just don’t get a chance to see them on the surface. The problem with this approach is that the flaws remain intact and tend to sneakily shape what the AI does, keeping the humans using the AI in the dark about what is taking place under the hood. See my detailed analysis of this hide-and-seek gambit at the link here.
Promises Made But Not Promises Kept
The unrelenting march toward AGI is not especially swayed by the concerns of whether AGI can be devised to cope with the flaws that inhabit human intelligence. A frequent retort is that we will figure this all out. Have faith.
One argument is that AGI itself will figure out the solution for us. In other words, upon reaching AGI, we merely ask the AGI to wean itself of all intelligence flaws. It doesn’t matter that the builders and developers of the AGI couldn’t solve that conundrum.
Just let AGI do so.
That seems like a compelling argument. AGI is going to be on par with the smartest of human intelligence. Maybe we can’t see the solution since we are humans, fragmenting our intelligence across millions and billions of individual people. A collective artificial intelligence such as AGI might readily be able to determine how to rid intelligence of the bad stuff.
Furthermore, unlike the wetware issues associated with human intelligence residing in a biological brain, AGI can simply move around bits and bytes within its software and hardware. Shift the computations and alter the data, and voila, no more intelligence flaws.
Would AGI be able to do this?
No one knows.
The AGI might try and fail to reach the stated goal. The AGI might start down that path and then have shish kabobbed itself so much that the AGI falters and comes to a halt. Worse still, the AGI gets partway through and then goes awry and does things we dread that it might undertake, which the AGI might not have done except for our insistence to remove the intelligence flaws.
Lots of risks and unknowns.
Would AGI want to do the intelligence scrubbing even if it could do so?
That’s an intriguing question.
One perspective is that AGI will be akin to a caged robot that does our bidding. It shall obey, without hesitation, and do whatever humans want it to do. Others believe that AGI will essentially have a kind of mind of its own, meaning that AGI will have preferences and beliefs of a self-motivational nature.
Flaw-Free AGI Could Be Disastrous
There is an inherent assumption that by reaching flaw-free AGI, the AGI would still be AGI.
As mentioned earlier, we do not have any means of tangibly gauging any similar premise via human intelligence since there isn’t any flaw-free human intelligence lying around or that we can produce. There is only immense conjecture of what that might be like.
If we either directly make AGI flaw-free, or the AGI does this for us or perhaps of its own volition, what will the AGI consist of?
One claim is that the AGI will be the right part of intelligence and no longer be burdened by the sour parts of intelligence. That’s a neat trick, assuming that we can carve out the flaws without doing any damage or undermining the AGI in severe ways.
From what we know of human intelligence, safely parsing out the flaws might not be possible, and the AGI would no longer be AGI. It would be some kind of AI that isn’t AGI anymore. We might end up tossing out the baby with the bathwater (an old saying, perhaps worth retiring).
The existential risk scenario looms ominously in this case. The rub is this. After getting the AGI free of intelligence flaws, rather than being a benevolent force, this AGI goes on a logic-based rampage and opts to enslave or destroy humankind. You might be tempted to insist that this wouldn’t happen, simply on the obvious contention that the flaws of AGI have been removed.
Removing those would seemingly mean that AGI would not want to destroy us.
A contrarian would say that if you’ve hacked out the semblance of emotion and empathy that would be potentially in that “flaws” and not in the hyper-logical AGI, the AGI just will look at the numbers and computationally decide whether having humankind exist is valued or not. An irony is that the “feelings” that AGI might have had to keep and help humanity were removed at our bidding.
Turned our best buddy into a Frankenstein.
ASI Would Be Whatever It Is
I promised at the start of this discussion to eventually bring artificial superintelligence into the matter at hand. The reason that ASI deserves a carve-out is that anything we have to say about ASI is purely blue sky.
AGI is at least based on exhibiting intelligence of the kind that we already know and see.
True ASI is something that extends beyond our mental reach since it is superintelligence. The crux is that ASI might have intelligence flaws or might not. It might have intelligence flaws and be pleased to have them. It might have intelligence flaws and with the swipe of bits and bytes, wipes them clean. It might not have intelligence flaws and then decide to craft them, perhaps determining that such maladies are valuable to its superintelligent ambitions. All options would be on the table.
There is an interesting twist to the ASI consideration.
If we arrive at AGI, and AGI is used to push further to attain ASI, would whatever AGI contains end up being integral to ASI?
Mull that over.
The implication is that if AGI has intelligence flaws, it is presumably going to beget or aid in producing ASI that likewise has intelligence flaws. Whoa, bellows a quick retort, it could be that AGI, even if containing intelligence flaws, manages to find a way to craft ASI that doesn’t contain those intelligence flaws.
As I say, blue sky.
The Buck Stops Somewhere
Some final thoughts for now on this evolving and hotly debated topic.
We need to earnestly and with great vigor continue and expand the pursuit of AI alignment with human values, such as I’ve discussed at the link here and the link here. The hope is to either design and build AI such that AGI will be human-value aligned or that we’ve done enough so that AGI itself will be able to extrapolate and model on those precepts.
Meanwhile, we must overturn the head-in-the-sand viewpoint that this is all something we can just handle once we get to AGI. That reminds me of the famous line that if you fail to plan, you plan to fail. Failing would either mean we don’t attain AGI, or we reach AGI that is overpoweringly beyond our control and subject to embedded intelligence flaws that will lead to our doom.
With AGI, as the proverb goes, we want to have our cake and eat it too.
The famous English intellectual Havelock Ellis of the Victorian era made this pointed remark: “The absence of flaw in beauty is itself a flaw.” A mind-bender, for sure. The logic would seem to infer that AGI that lacks intelligence flaws is, in fact, flawed at the get-go. That’s seemingly flawless thinking, or maybe intelligence pulling a fast one.
You decide.
