The Real Problem With Artificial Intelligence

It’s not that it is “too intelligent.”

In March of this year, Elon Musk and 1,000 other AI industry leaders signed a letter calling for AI development to be halted until protocols can be implemented to ensure AI doesn’t pose a risk to society. This has led many people to draw parallels to Nick Bostrom’s 2014 book Superintelligence, in which he stipulates that AIs that are more intelligent than humans constitute an existential catastrophe for us. But the actual danger of AI is far from this and far more pressing. You see, AI is not as intelligent as we think it is, and these dumb machines in smart clothes already pose a significant risk to us. But how?

Firstly, let’s set the ground rules for this article. I am discounting any statements made by people with a significant stake in the AI industry. While many of the points they make about AI are very valid, their opinion on the potential of AI is massively swayed. I mean, take Elon Musk; he has literally billions of dollars wrapped up in companies whose worth is entirely based on AI — namely Tesla and X.ai. Of course, he is going to hype up just how powerful they are, and the same goes for every AI CEO or AI startup scientist. Instead, the sources for this article are from AI researchers and well-regarded public scientific figures. While this isn’t a foolproof way of getting a non-biased view, it means we can at least get as close as possible. That being said, this is still an opinion piece, and you are more than welcome to disagree with me.

So, let’s start with what an AI is. AI uses virtual neural networks that work similarly to a biological brain to recognise patterns in data and decide a response to that data. A great example is mammogram screening for breast cancer. The early signs of breast cancer are incredibly hard for doctors to spot, as they look almost identical to health cells. But an AI was trained on apparently healthy mammogram images of patients who went on to develop breast cancer, and it can now identify breast cancer better than a doctor.

Recently, more complex AI systems have emerged, such as self-driving AIs and GPT4. GPT4 is a remarkable writing AI that can pass the infamous Turing test, hold engaging conversations, write people’s homework for them and even write full-blown articles. I have even looked at trying to use GPT4 to help write more articles and cover a more comprehensive range of topics for my readers, though I haven’t yet (more on that later). Meanwhile, self-driving AIs, such as that of Tesla or Waymo, can safely navigate through complex traffic junctions at peak traffic.

This is why many people are saying that AI technology is now so advanced that it is getting better than humans, and we are reaching a tipping point where AI could pose a nuclear-level threat. Even Geoffrey Hinton, who was one of the computer scientists who invented the basis of modern AI technology, quit his job at Google to be able to openly speak about the threats of AI as he “suddenly realised that these things are getting smarter than us.”

But does this stance hold up to the evidence? No, it doesn’t. But his worries are understandable, and we should take heed. Let me explain.

AI has something called the black box problem, which makes them far from intelligent. Take a self-driving AI. If it were truly intelligent, it would understand the rules of the road and have learnt the highway code. However, we can’t take apart the AI apart and find this learnt rulebook within it. In fact, the AI doesn’t actually know the rules of the road at all. All it does is pattern and recognition and response.

This means we can’t predict how AI will respond, as the rulebook it is playing to is unknowable.

This is one of the possible reasons why last year, Tesla’s FSD was filmed taking a manoeuvre that would have run over a cyclist, despite the fact the systems had identified the cyclist. Luckily, the driver was paying attention and took back control, saving the cyclist. Why this happened, we don’t know, but if FSD were truly intelligent and understood the rules of the road, it wouldn’t have driven like that.

You can see the same thing in GPT4. While it is very, very good at mimicking natural human writing, it doesn’t have the logic to construct a cohesive argument. It makes up false facts, and its line of reasoning can be incredibly flawed. What’s more, it doesn’t or struggles to cite its sources or even makes them up. This is because GPT4 can only mimic human writing and does not have the intelligence to actively engage in conversation or cohesive writing. As it stands, it’s currently only suitable for non-factual writing that appears to be entertaining to read, hence why I don’t use it. Yet.

One final example of this is the famous AlphaGo, which was the first AI to beat a human player and the ancient and complex board game of Go. You’d think, as it has learnt something comparatively constrained compared to driving or conversation, that it would have figured out all the rules of Go and how to play the game. But that isn’t so. Researchers found a way to make an absolute novice beat AlphaGo almost every single time by using an incredibly simplistic strategy that was entirely within the rules. If AlphaGo actually knew the rules of Go, it would have easily seen what this novice was doing and beat them. Instead, it played incredibly erratically and lost.

For more on AlphaGo and GPT4’s problems, I highly recommend the incredible Kyle Hill’s video on the topic.

Again, AlphaGo isn’t intelligent. It doesn’t understand what’s going on. All it is doing is pattern recognition and response. And like all the other AIs, we can never fully understand why it responds the way it does.

This is why AI is dangerous. It isn’t that it is some sort of super-intelligence that can outperform humans. Not yet, at least, and many researchers think we are decades away from such AIs, and even then, they will be limited. Instead, it’s that they are potentially unfit tools for the job and can inadvertently cause disaster.

Take GPT4. Imagine letting it control a political campaign. It would churn out voter-appealing misinformation at rates we have never seen before, without understanding the consequences. But we can take this problem even further. What if we let our economy, military, media, healthcare, transport or schooling be controlled by AI? How do we know it won’t act erratically? We can’t, as it can never understand what it is actually doing.

AIs are dumb machines in smart clothes. They don’t understand the world around them. They can mimic patterns, but this is not intelligence. This isn’t just my view; it is shared by many in the AI space. The risk comes when we treat AI as intelligent and let them have autonomy. As such, this unchecked way of using AI poses a monumental risk.

The problem here is that in order for AI companies to be worth the amount of money that has been put behind them, their systems need to be given autonomy! What’s more, many companies and individuals are happy to use AI in this way, as it can either lower overheads and increase profits or make their lives easier.

This is why that letter with 1,000 AI industry leaders’ signatures on it is so crucial. You see, if there is a way to level the playing field, and ensure every AI has to have the correct checks and protocols in place to ensure it isn’t being misused or with too much autonomy, then this threat should be significantly reduced. That way, no one AI company can get the lead in the market by letting their AI gain massive levels of autonomy before it, or the people using it, are ready. But, as of right now, AI development hasn’t been paused, and these protocols and checks haven’t been created. So as it stands, AI does pose a significant risk to us, but not out of “superintelligence” but out of incompetence from the machine itself and those using them incorrectly.

Enjoyed this article? Get articles & ad-free video articles 24 hours before anyone else on Substack, or check out my YouTube channel for daily video articles.

**Article originally published on Substack**