Has AI Already Surpassed Us?
Has AI Already Surpassed Us?
Yes, but more importantly, no.
A few days ago, Stanford University’s Institute for Human-Centered Artificial Intelligence (HAI) released its seventh comprehensive AI Index report. This report, being such a thorough and comprehensive assessment of the breakthrough technology, has been the litmus test for AI development for years. If you want to know how good AI is, you should turn to HAI’s AI Index report. So you can imagine my shock when journalists started to report that the latest instalment declared that AI is better than humans in almost every measurable way. If that sounds a bit off, it’s because it is. AI still can’t hold a candle to use. So why these results?
This isn’t the first time this index has found AI can outperform humans. Back in 2015, it found that AI could achieve a higher success rate at image classification. In other words, AI could more accurately identify items within a single image than we humans can. Over the years, it has been found that AI could achieve better-than-human success rates in more and more tasks, such as basic reading comprehension (in 2017), visual reasoning (in 2020), and natural language inference (in 2021). In this latest report, they found that humans could only surpass AI in two areas: visual common-sense reasoning and competition-level mathematics. However, it was also found that AI could outdo us in image classification, natural language inference, medium-level reading comprehension, multitask language understanding, and visual reasoning.
When put in such simple terms, it sounds like AI is only a few steps from totally eclipsing human intelligence.
But, when you look at real-world applications, this notion of AI surpassing humans falls apart.
Take Amazon’s “just walk out” store, which I covered a few days ago. Its AI technology was meant to tot up a shopper’s purchases as they took them from the shelves, enabling Amazon to charge their account as they left with the items, rather than the customer having to use a check-out. This is a fairly basic image recognition AI aided by shelf sensors, yet it ultimately failed. A recent report found that over a thousand remote workers had to be hired to monitor the video feeds from a single store and verify 70% of the customer’s purchases, as the AI was consistently getting it wrong. This method of using humans to keep an eye on the AI not only proved that AI can’t outperform humans at image recognition but also how economically backwards it is, as hiring so many remote workers was so expensive that Amazon has ditched the entire “just walk out” concept.
This notion is further backed up by the Harvard Business Review, which has stated multiple times that AI is too unreliable to make decisions independently and, if used to replace workers, will damage a business. Instead, they postulate that AI should be used to better inform human workers, who are better at making these critical decisions.
Then there are self-driving cars. Again, these are image recognition and visual understanding AIs, which are meant to perform recognition and response tasks that most humans can pass with relative ease. Yet, no self-driving car has yet been as capable, reliable or safe (in all situations) as a human driver. In fact, some sources suggest that Tesla’s top-of-the-line self-driving product, FSD, is around 10 times more dangerous than a human driver!
These are just a few examples of AI falling seriously short of human performance; there are many others. But this means there is a vast disconnect between what this highly respected Stanford report is saying and the real-world results we are getting from AI.
So why the disconnect? Well, it’s all about the tests Stanford uses. You see, these tests have to be conducted in such a hyper-controlled and constricted application to make such analysis possible. In this setting, AI can thrive. There are very few external variables, and errors don’t compound to send them wildly off course. But, in real life, things aren’t so clear-cut, and you must catch errors before they become a significant problem to perform even the most basic tasks.
This is where human intelligence and AI intelligence differ massively. We humans might not be able to recognise objects as accurately in a single image as AI, but we don’t make single observations. We evaluate and test our understanding and decisions constantly and mostly subconsciously. This means we can catch and correct our errors, enabling us to perform complex, multilayered and highly interconnected tasks way better than AI. AI simply doesn’t have the depth of understanding or general intelligence required to make such self-corrective decisions. This is why self-driving cars repeatedly misread road signs or misidentify other road users and conduct manoeuvres which seem utterly inept. It’s also why AI is not more intelligent or capable than a human in the real world, but for the most part, it is in a very controlled and utterly unrealistic environment, such as Stanford’s tests.
So, has AI surpassed humans? Stanford’s index is utterly thorough and isn’t false in any way, shape, or form. But it isn’t representative of how AI is used, as such, its results don’t answer the question. It’s like saying a monkey is more capable than a fish because it can’t climb a tree. That analogy doesn’t even work because we have trees, whereas the use cases where AI can surpass humans are so restrictive that they require a human to enable them and utilise the results. So, no, AI hasn’t surpassed humans. What’s more, for AI to surpass humans, it requires a level of cognition which it hasn’t yet got (depth of understanding or general intelligence), and right now, it seems possible that AI could never feasibly achieve such a level as we have no viable way to give it such cognitive abilities.
Thanks for reading! Content like this doesn’t happen without your support. So, if you want to see more like this, don’t forget to Subscribe and follow me on BlueSky or X and help get the word out by hitting the share button below.
Sources: New Atlas, Planet Earth & Beyond, HBR, Tech Target, CNBC, Planet Earth & Beyond, The Verge, The American Prospect