Can AI replace a human?

There have been many talks recently about AI replacing humans. Just put 1 – 2 trillion dollars more into it, and it will surpass any human. But is this really the case? For example, imagine that you don’t know anything about water. You asked AI the following question: “What can you do with water?” And AI may answer that you can put it on fire with matches.

And because you don’t know anything about water, you assume it is true and act on this information. Then, later, to find out that it is completely wrong. Moreover, very often, if you think before acting on information provided by AI, you can find a contradiction in AI’s answer.

As a result, there are multiple follow-up questions:

1. Why does AI answer like this?
2. Why do we trust it?
3. Why don't we use our brains and ask AI instead?

Part 1. How AI provides this kind of answer.

Maybe I need to provide an example here. Imagine you have a calculator. Somebody told you that it can calculate any math operation. You tried 2 + 2 and got the correct answer. You tried to calculate the square root of 9 and also got a correct answer. You tried several more operations, and they all seem correct. At this stage, you would assume that the calculator is good and you can use it in your everyday tasks.

But this calculator cannot calculate any operation at all. It was just trained on a huge dataset of math operations, and it simply knows the answers to most math questions. But because there are unlimited number of operations, it obviously cannot have an answer to all of them.

Remember, AI does not know how to calculate anything. All it does is try to find the next suitable word that most often follows your answer. Because AI was trained on 2 + 2 and does not have the answer for 2 + 2.1, it will place 4 as an answer for both cases.

If there is nothing close, then it may provide the word “red” as an answer. Just because this word was quite often around math questions on training dataset. Moreover, the calculator almost never provides the answer “I don’t know.” There is always some word that can be placed after your question.

And the worst thing is that most of the time results will be correct, and only sometimes incorrect. I think it is obvious that nobody will want this calculator because it will provide incorrect results on anything that was not in the original dataset. It will not say: “I cannot calculate” or “I don’t know”, instead it will provide a plausible or plain incorrect answer.

To be 100% correct, it was how AI worked in the early stages of AI. Obviously, AI companies understand this problem, and their solution is simple - feed more data to it. With more training data, the probability of a correct answer increases. But each increase increases probability by only a little bit while cost grow up substantially.

AI companies also add enforcement that tries to validate the answer to avoid getting complete nonsense as an answer likr the word “red” as an answer. It is getting better and better, but you can never rely on it because it does not calculate the answer mathematically. It just matches whatever is close enough or plausible enough in the training dataset to answer. But we don’t want close enough or plausible enough answers from the calculator. We want only correct answers.

And this is the biggest trap with AI. AI does not have intelligence. It is a super advanced pattern matching with enforcing. AI provides some output, enforcer validates it. If the output failed to meet prompt constraints, then the enforcer part will force AI to regenerate it.

But the enforcer also does not have any intelligence. In many cases, it cannot logically validate the text produced by AI. As a result, the text just needs to look plausible to pass enforcement.

If you don’t believe me, you can try to ask a question that AI cannot possibly know the answer to. For example, I asked how to convert the source code of one of the modules in our application to Python. The answer was completely incorrect, but it looks very plausible to anybody who does not know how our application is written.

Any human will answer “I don’t know” or “I don’t understand” in this situation and will not try to create a plausible answer. Moreover, often we will try to collect as much information as possible before answering. AI will answer with very little information and often incorrectly.

Part 2. Trust

And here is a logical question:  why do we trust AI? Because we treat the calculator as another person. Most people even say “Thank you” to it. When we need to understand if the person in front of us knows about something, we start to ask simple questions that we know the correct answer to and see their response. We not only monitor the answer for correctness, but we also monitor how that person answers.  If that person hesitates, thinks for a long time, often corrects themselves, etc., then we assume that this person is not a reliable source of information.

But if you ask somebody and they immediately provide an answer in a convincing voice, then we start to ask more complex questions. If that person still confidently answers them and never contradicts themselves, then we typically assume that this person is knowledgeable and we can trust them.

Companies that created AI are obviously trying to convince us that AI can do a lot of things. As you can imagine, we didn’t believe it, and we ran some tests. AI has access to a lot of encyclopedic knowledge, and as a result, it can correctly answer most of our test questions. And AI provides an answer in very confidently worded text with very plausible explanations for every answer. As a result, people automatically and instinctively will trust it because it is what they would do when another person will provide answer in a similar manner. They simply forget that the answer was generated by a machine.

But AI simply doesn’t have the capabilities to answer complex questions correctly because it requires a deep understanding of that subject, but AI is just super advanced pattern matching.

This is exactly the thing that makes people who use AI less productive in a lot of cases. People ask something of AI, and it produces a very plausible response. Then people act on it only later to find that it is complete nonsense, like the case with water I mentioned above.

For example, for my example with water, if you ask why water is flammable, it will provide a very convincing explanation. For example, it may explain to you that water consists of oxygen and hydrogen, and both are extremely flammable, so the combination of them is definitely flammable.

Part 3, Why don’t we think?

When we ask AI and got answer, we tend to just accept it without much critical thinking, especially if we don’t know much about that area. And this is also a psychological thing. If you failed to solve a problem by yourself and ask a guru about that problem, then we all tend to gladly accept their solution. It happened because we failed to find a solution, and then someone spent their own time to understand our problem, spent time thinking, and spent time explaining a solution to us.

So we just gladly accept it unless there is a direct and obvious contradiction. We forgot that AI is not a real person, but we unconsciously treat it as some kind of guru. And when this “guru” produces tens of lines of text with a detailed explanation of the solution for our problem, we are just overwhelmed with such an explanation and happily accept it.

But why don’t we think and ask AI in the first place? There are many reasons. Thinking is hard and painful for most of the people. It is much easier to ask AI. And a lot of people are just lazy. Also, while AI is thinking, people can watch memes, talk with colleagues, or do something else not work-related.

If somebody asks them why they are not working, they can always state that they are waiting for AI to complete research of certain subject to avoid future problems. Some people just try to cover their bottom part and put the responsibility on AI. “Our company is spending a lot of money on AI, so I’m using it “. Basically, AI provides an opportunity for many people to be lazy and not work.

Part 4. Other issues

There are more problems with AI. The biggest issue with any AI is finding quality training data. Some training data is verified. For example, it can be an encyclopedia or science books. They still may have typos or factual mistakes, but they are quite rare because they are verified by many people.

But AI is also trained on data from the internet. For example, before AI time, Stack Overflow was one of the most popular resources to find answers to complex programming issues. Multiple studies shown that about 90% of the most popular answers on Stack Overflow have mistakes.

It is hard to tell why that percentage is so high. Sometimes, handling all situations will require much more code and time from the person who answered. Sometimes, a person just doesn’t think about other cases and solves only the obvious ones. Sometimes, people answer just to collect reputation, and the faster you answer, the more upvotes you will have. There are many reasons.

Now, data from Stack Overflow is used to train AI models. Imagine that our calculator was trained on data that states that 2 + 2 = 3. Guess what answers it will produce? Yes, incorrect answers. But they will look correct or will work correctly in most of the cases. Until it won't.

Also, you can find all sorts of information on the internet. About the flat Earth. About lizard people and aliens. About secret societies that exploit other people and their governments. We all saw that kind of information on the internet. Just remember all these heated debates around Covid.

Now, imagine that all that information were used to train AI. It is just a super advanced pattern matching, and when it constructs a response, it will just put words that are most likely to be there from texts it saw on the internet. Basically, garbage in, garbage out.

To mitigate garbage from the internet, the AI team must provide verified facts to the AI model, and these facts will have a higher priority. In the case of our calculator, if the AI team found out that the calculator didn’t provide the correct answer to 2 + 2.1, then they will manually add it.

But it requires a lot of manual labor to provide verified facts, and typically it is outsourced to countries with cheap labor. They told them not to use AI, but many still do because they will work less and get more money. There is supposed to be someone who verifies these facts, but most of the time they don’t because it is very boring work and they do it only for very critical information.

There is also a lot of obsolete information. Especially in the IT world. For 10 years, we were supposed to do X. Then things changed, and now we're all supposed to do Y. But AI was trained on information that, in 95% of cases, tells us to do X and only in 5% to do Y. Obviously, AI will suggest X and not Y because it is based on probability. Until somebody adds verified facts.

Also, sometimes, AI simply cannot find the correct words to place because all of them contradict the prompt constraints. Then AI will place any word that does not contradict constraints. As a result, it will produce the correct text in English that has absolutely no connection to reality but sounds quite reasonable. And if you ask AI to explain the answer, it will provide a lot of confident text that explains why this garbage is correct.

Finally, right now, people are using AI to generate questionable data and then also place it on the internet (for example, back to StackOverflow). Then AI will be trained on that data too, further reducing the quality of its data. Then produce even more questionable data. Rinse and repeat.

In conclusion

AI tools are simply too expensive and too unreliable to use at scale. When human driving a car, they can make mistakes, but they are much more predictable, and our roads are built to take this into account. AI can theoretically do absolutely anything, like try to turn 180 degrees and apply brakes at the same time to flip the car over. It can stop and reverse on the highway or do something else.

People can make mistakes when manually calculating something, but their mistakes are predictable. Humans can make simple mistakes like typing 1 +10 instead of 1 + 1. But AI can make any type of mistakes. For example, 1 + 1 = 1000000. Why not. Moreover, today’s AI model can calculate 1 +1 correctly, but tomorrow’s model may provide something bogus.

And don’t let me start on security. Currently, AI tools accept the system prompt as plain text. And they typically work with plain text. It is possible to convince AI via text to drop the system prompt and do something else. It is especially bad with AI agents that accept information from the outside.

For example, imagine an AI agent that has access to your email and your file system. Then you told AI to read your emails to summarize them for you. But text in your emails can convince the AI to send files from your computer to somebody else. And this is not a theoretical issue; it is what actually happened.

AI developers can definitely improve it, but it requires even more computational power. But that will make an already very expensive tool even more expensive. From my research, it looks like this problem cannot be fixed completely. And this is very scary.

I’m not saying that AI is useless, because it is quite useful. It improves our lives considerably, but it most likely will not do what AI companies told us. Power screwdrivers improved our lives a lot, but they didn’t eliminate humans. They just make people more productive.

AI also made us more productive. But it also makes us lazy and makes us trust blindly in AI at the same time, which cancels any productivity gain and often makes people less productive.

Here are my 2 cents. My feeling is that AI pretty much reached a plateau and almost reached its fullest potential. AI cannot solve complex questions because it cannot learn. It cannot replace humans unless they are dumb.

And lastly, AI is very, very very very expensive. Right now, we pay from 1% to 5% of its true costs. All these AI company just burning hundreds of billions of dollars just to capture market share and evict competitors. After that, they must raise prices considerably.

Some of them are simply waiting on some technological or algorithmic breakthrough to make AI considerably cheaper. If you remember the history of computer hardware, you will remember how something that cost millions of dollars, but 10 years later, the same would cost thousands at most.

But these times were different. It was just the beginning of the information revolution. Now everything is at quite mature state. Chips are already hitting the limits of physics. People and tools squeezed everything from software. As a result, I don’t think there will be any breakthrough.

As a result, I think the AI bubble will pop, similar to the .com bubble in 2000. That bubble did not happen because .com and IT were useless. Quite contrary. It was just extremely overhyped. Exactly, like AI now. AI has it use, but we need to understand its constraints and limitations, and continue to use our brains.

I hope this was helpful.