0:01 Should we be afraid of artificial
0:04 intelligence? It is a question that
0:06 keeps many of us up at night. Most
0:08 people first heard about AI through
0:11 1990s movies like Terminator or The
0:13 Matrix, where it was portrayed as a
0:15 villain turning against humanity. And it
0:18 was not just the movies. Renowned
0:20 figures like Steven Hawking have also
0:22 warned about the possible dangers of AI
0:25 rising against human beings. So back
0:28 then it was completely natural to feel
0:30 uneasy whenever we heard the term
0:33 artificial intelligence. But for nearly
0:35 20 years after those movies, AI made
0:37 little real world progress that could
0:39 cause serious concern. Things stayed
0:42 relatively quiet. But today the
0:45 situation has completely changed. AI is
0:48 now part of our daily lives through AI
0:52 cameras, Alexa, Siri, ChatGpt, Google
0:54 Gemini, and many more tools. Even the
0:56 YouTube algorithm that suggested this
1:00 video to you works using AI. And it is
1:02 not just YouTube. On platforms like
1:04 Facebook and Instagram, AI decides what
1:07 you see on your feed. Another unsettling
1:10 use of AI is in deep fake videos, which
1:12 many of you may have already come
1:14 across. In several countries, banks now
1:17 use AI to approve loans. Insurance
1:19 companies use it to calculate your
1:22 premium, and AI even plays a major role
1:25 in stock market trading. In the medical
1:28 field, AI is now being used for
1:31 diagnosing diseases. In short, AI is
1:33 being used in areas where serious
1:36 decisions are being made. Yet despite
1:39 all this, we have not seen AI behaving
1:41 like the evil villain shown in those
1:43 science fiction movies. And that
1:45 naturally leads to an important
1:47 question. Is there a difference between
1:50 the AI we saw in movies and the AI we
1:52 are seeing today? What exactly is
1:55 artificial intelligence? How does it
1:57 work? How many types of AI are there?
2:00 And could AI ever become a villain as
2:02 shown in the films? Let us explore the
2:11 Hi friends, welcome to a new video from
2:14 science simplified for all. Most of us
2:15 have been familiar with AI based
2:18 personal assistants like Alexa, Siri and
2:20 Google Assistant for some time now. But
2:22 it was only after the arrival of Chat
2:24 GPT and Google Gemini that many people
2:27 truly realized just how far artificial
2:30 intelligence has evolved. Chad GPT seems
2:32 to understand what we say in natural
2:34 language and respond in a way that feels
2:37 remarkably human. Because of this, many
2:40 people unknowingly assume that chat GPT
2:42 is a conscious being, something with
2:44 self-awareness. But we must always
2:47 remember Chat GPT is just a software
2:50 program specifically designed to mimic
2:52 the way a human respond in conversation.
2:54 It does not truly understand things the
2:57 way a human does. To clearly grasp this
2:59 idea, we first need a basic
3:01 understanding of what artificial
3:03 intelligence actually means and how it
3:06 works. This video will give a very
3:08 simple explanation aimed at everyday
3:11 viewers with no technical background or
3:14 prior knowledge about AI. So if you are
3:16 someone already familiar with the core
3:19 concepts of AI, feel free to skip this
3:21 part as it may seem like an oversimplification.
3:23 oversimplification.
3:26 Let us begin with the term intelligence.
3:29 The ability to learn new things, make
3:32 logical decisions and solve problems.
3:34 That is what we call intelligence in a
3:36 human being. When a machine or computer
3:39 begins to exhibit this kind of ability,
3:42 we call it artificial intelligence. That
3:46 is the most basic definition of AI. But
3:48 this definition alone does not give us
3:50 the full picture. We already know that
3:52 computers have been capable of doing
3:55 mathematical calculations much faster
3:58 than humans for decades. For example, if
4:01 you ask a computer to multiply 1,230
4:03 by 2480,
4:06 it will do it instantly, far faster than
4:09 any human ever could. In fact, a
4:11 computer can perform millions of such
4:13 calculations every second. In that
4:16 sense, computers have always been ahead
4:18 of us in solving arithmetic problems.
4:20 But here is the key point. These
4:22 calculations are all based on
4:25 instructions that we the humans have
4:26 already given to the computer in
4:29 advance. These instructions are what we
4:32 call programs. A standard computer can
4:34 only follow those predefined
4:36 instructions exactly as given. If it
4:38 encounters a situation slightly
4:40 different from what we programmed it
4:43 for, it will usually fail. Now, here is
4:45 the difference. When a computer is able
4:48 to do something new, something we did
4:50 not specifically teach it by learning
4:52 from the data and recognizing patterns
4:55 on its own, that is when we say the
4:57 computer has artificial intelligence.
4:59 Let us take an example. Suppose we
5:02 create a program to identify birds and
5:04 upload it into a computer. Then we show
5:06 the computer pictures of 10 different
5:10 types of birds. A crow, an eagle, a
5:13 sparrow, a parrot, and so on. We also
5:16 tell the computer these are all birds.
5:19 Now imagine we show the computer a new
5:21 picture of a bird it has never seen
5:23 before. If the computer is able to
5:26 recognize that the new image is still a
5:28 bird even though we never showed it that
5:31 specific example, that is a simple form
5:33 of artificial intelligence. It has
5:35 learned from the previous examples and
5:38 applied that knowledge to something new.
5:41 That is one of the key traits of AI. Of
5:43 course, what today's artificial
5:45 intelligence is capable of goes far
5:47 beyond this simple example. AI systems
5:50 now handle much more complex and
5:52 powerful tasks. But we are using this
5:54 bird example here to make the core
5:57 concept easier to understand. Now, let
5:59 us take a closer look at how the
6:01 artificial intelligence program in our
6:04 bird example actually works. We begin by
6:06 defining the unique features that
6:09 distinguish a bird from other objects.
6:11 For example, we might say that birds
6:14 usually have two legs, wings, feathers,
6:16 and a beak. These are the features that
6:19 we define in the program. Next, we
6:21 assign a weightage to each of these
6:23 features. That means we decide how
6:25 important each feature is in identifying
6:29 a bird. For instance, two legs 20%,
6:35 wings 30%, feathers 30%, beak 20%.
6:39 Together that adds up to 100%. Now if we
6:41 show the computer a picture of a crow
6:44 which has all four features it would
6:46 score 100% and the program would
6:50 confidently say yes this is a bird. But
6:52 suppose we show it a picture of a
6:54 penguin. A penguin does not have visible
6:57 wings or feathers like other birds. Its
6:59 wings look more like arms. So the
7:02 program might only find 40% of the
7:04 features it is looking for. Based on
7:08 this low score, it may wrongly decide
7:10 this is not a bird. But we know that a
7:13 penguin is a bird. So we correct the
7:15 program and tell it that it made a
7:17 mistake. When the program receives this
7:19 correction, it begins to adjust the
7:21 weights it had assigned earlier. It
7:23 might reduce the importance of features
7:25 like feathers and wings and increase the
7:28 weightage for features like a beacon two
7:31 legs. It also learns that a perfect 100%
7:34 match is not always necessary. Even if a
7:37 creature matches only 80% of the key
7:39 features, it could still be a bird. This
7:42 program is written in such a way that it
7:44 can adjust these weightages by itself
7:47 based on feedback. As we continue to
7:49 show the program more and more images of
7:51 different birds, it keeps refining its
7:54 internal settings. Each time it makes a
7:56 mistake, we correct it and the program
7:59 learns from that error. In other words,
8:02 it adjusts its internal parameters.
8:04 Eventually, after seeing enough examples
8:06 and making enough adjustments, the
8:09 program becomes capable of correctly
8:12 identifying any bird, even one it has
8:14 never seen before. This process is what
8:17 we mean when we say an AI program is
8:19 being trained. In reality, the working
8:22 of an AI system is far more complex than
8:25 what we just explained. Modern AI
8:27 programs use many different parameters
8:29 and weight values and these are
8:31 processed through multiple layers and
8:34 stages before reaching a final decision.
8:36 Still, we chose to present the concept
8:39 in such a simple way for one important
8:42 reason to help you understand three key
8:45 points clearly. First, in the example we
8:48 discussed, the AI does not actually
8:51 understand what a bird is. In fact, the
8:53 concept of understanding or mind itself
8:55 is quite vague. But even if we leave
8:58 that aside, what a human understands
9:00 when they see a bird is fundamentally
9:02 different from what an AI sees. For the
9:05 AI, a bird is just a group of numbers,
9:07 weight values associated with different
9:10 features like a beak, feathers, and
9:13 other such traits. That is all. Second,
9:16 to train an AI, we need a massive amount
9:19 of data. Even in the bird example, the
9:21 AI needs to be shown thousands of bird
9:24 images to learn what a bird looks like.
9:27 And not just that, every time the AI
9:29 makes a mistake, a human has to step in
9:32 and correct it, telling the AI that it
9:34 was wrong. Without that feedback,
9:38 learning will not happen. Third, even
9:40 after all this effort, the AI only
9:43 learns to answer one specific question.
9:46 Is this a bird or not? If you want the
9:48 AI to answer a different question like
9:51 which bird it is, you need to write a
9:52 different program and train it
9:55 separately. And if the image is not a
9:57 bird at all and you want the AI to say
9:59 whether it is a mammal or something
10:01 else, that too requires a separate
10:03 program and new training. In other
10:06 words, each AI can only perform the
10:08 specific task it was trained for. If you
10:11 want the AI to do a new job, it must be
10:14 trained again, often from scratch. To
10:17 put it simply, every specific task
10:20 requires a specially trained AI system.
10:21 Just because something is called
10:23 artificial intelligence does not mean it
10:26 can do everything. All these limitations
10:29 we discussed so far explain why even
10:31 though science fiction movies talked
10:33 about AI decades ago, the real
10:35 development of artificial intelligence
10:38 took much longer to happen. The rapid
10:40 growth of AI that we see today became
10:42 possible mainly because of two key
10:45 factors. The first reason is the
10:46 dramatic increase in computer processing
10:50 power. Back in the 1990s, computers
10:52 simply did not have the capability to
10:56 run AI programs. But over the last 20
10:58 years, the computational speed of
11:00 computers has increased by several
11:03 thousand times. This incredible
11:05 improvement is what finally made it
11:08 possible to run complex programs like
11:09 artificial intelligence on regular
11:12 machines. The second major factor was
11:14 the rise of social media. Let me give
11:17 you an example. When you upload a photo
11:19 of your pet dog to Facebook and tag it
11:21 as dog, you're actually helping
11:23 Facebook's artificial intelligence learn
11:26 what a dog looks like. Every time
11:29 someone does this, the AI gets better at
11:31 recognizing dogs, even when it sees a
11:34 totally new photo of a dog later. With
11:36 the explosion of social media, AI
11:38 systems suddenly had access to huge
11:41 amounts of data for training. Today,
11:43 millions of people across the world
11:46 upload photos every single day. Many of
11:48 those photos are used behind the scenes
11:51 to train AI programs. The same is true
11:53 for public messages, comments, and
11:56 captions that we post. All of this
11:58 content is used to help train artificial
12:01 intelligence in understanding natural
12:03 language. the way we humans actually
12:06 speak. And this is important because the
12:08 way we speak is very different from what
12:10 you find in books or dictionaries.
12:13 Spoken language is filled with informal
12:16 phrases, slang, and regional styles. Yet
12:18 today, we have reached a point where AI
12:21 can understand meaning even in informal
12:24 speech. And that is largely thanks to
12:27 the vast amount of training data AI has
12:29 received from language used on social
12:32 media. By now you should have a general
12:34 idea of how an artificial intelligence
12:36 system is trained. But there are a few
12:39 important things you need to know. When
12:41 an AI is trained using millions of
12:44 images and messages, the process is not
12:46 done manually by humans. That would be
12:49 practically impossible. Instead, the
12:52 training is handled automatically by
12:54 powerful software systems designed for
12:56 that purpose. And this leads to a major
12:58 problem. By the time training is
13:00 complete and the AI is released for
13:03 public use, even the developers who
13:05 created the system often have very
13:08 little idea of what exactly is happening
13:11 inside the AI's decision-making process.
13:13 This is because the training process
13:16 changes the AI program so extensively
13:18 that it becomes difficult to track how
13:21 it arrives at its conclusions. Here lies
13:24 the real issue. If such an AI makes a
13:26 mistake, it is incredibly difficult to
13:29 figure out where it went wrong or why it
13:32 made that mistake. This lack of clarity,
13:34 often called the blackbox problem, is
13:36 one of the biggest drawbacks of many
13:39 modern AI systems. Let me give you a
13:42 real world example. An AI system was
13:45 trained to identify animals in pictures.
13:47 But during testing, it started
13:49 mclassifying certain dog photos as
13:52 wolves. At first, the reason for this
13:54 behavior was unclear. Only after
13:56 detailed investigation did researchers
13:59 discover the actual cause. In the
14:01 training data set, almost all wolf
14:04 images had snow in the background. So,
14:06 whenever the AI saw snow behind a dog,
14:09 it assumed the image must be of a wolf.
14:11 In other words, it was giving more
14:14 importance or weight to the snowy
14:16 background than to the animal itself.
14:19 This kind of issue happens because after
14:21 training we no longer fully understand
14:24 what the AI is focusing on internally.
14:26 That is why it becomes so difficult to
14:29 trace and correct these kinds of errors.
14:31 To address this problem, a new approach
14:35 called transparent AI has been proposed.
14:38 The idea is to make AI systems more
14:40 understandable where we can see and
14:42 interpret what is happening inside.
14:45 However, transparent AI is still a
14:47 developing concept and has not yet been
14:49 widely adopted in real world
14:51 applications. There is one more crucial
14:54 point to remember. The data used to
14:57 train an AI must be completely accurate
14:59 and unbiased. If the training data
15:02 contains flaws or biases, those issues
15:04 will reflect directly in the AI's
15:06 behavior. Let me give you a real world
15:09 example. In one case, a company used an
15:11 AI system to shortlist candidates for
15:14 job interviews by analyzing job
15:16 applications. But when the results came
15:18 out, it was found that the AI had
15:21 selected only male candidates. On
15:24 investigation, the reason became clear.
15:26 The data used to train the AI mostly
15:29 came from past hiring decisions. And in
15:31 the past, the company had preferred
15:34 hiring men for that specific job role.
15:36 This bias in the historical data, even
15:39 if unintentional, got passed on to the
15:43 AI system. As a result, the AI learned
15:45 to favor male candidates simply because
15:48 that is what the past data showed. This
15:50 is a classic example of how bias in
15:52 training data can lead to discrimination
15:55 in AI decisions. And this can apply to
15:58 other human biases as well, such as bias
16:01 based on skin color, religion, or any
16:03 number of other prejudices. That is why
16:06 it is absolutely essential that the data
16:09 used to train AI is carefully checked
16:11 for fairness and neutrality. Otherwise,
16:14 our own biases will get transferred into
16:16 the AI we build. And here is something
16:19 we must never forget while using AI.
16:22 When an AI makes a mistake, it has no
16:24 idea that it made a mistake, nor does it
16:28 feel any regret. That is because AI has
16:31 no concept of understanding. It does not
16:33 know what is right or wrong. It simply
16:35 follows the patterns it was trained on.
16:38 So the responsibility to monitor AI
16:40 behavior will always rest on us the
16:43 humans. We must be the ones to watch to
16:45 correct and to decide what is
16:47 acceptable. Now let us look at the
16:49 different types of artificial
16:51 intelligence used today and the kinds of
16:54 jobs they are designed to do. The first
16:56 major type is called natural language
17:00 processing or NLP. This refers to AI
17:02 systems that can understand and respond
17:04 in human language, the way we actually
17:07 speak. Examples you're already familiar
17:10 with include Alexa, Siri, and Google
17:13 Assistant. But NLP is used in many other
17:15 areas as well beyond just virtual
17:18 assistants. The second type is
17:21 generative AI. This is a kind of AI that
17:23 can create new content, things that
17:26 never existed before. For example,
17:29 imagine an AI that has read thousands of
17:31 novels. Now, if you ask it to write a
17:35 completely new novel, it can do that. Or
17:37 if it has seen thousands of human faces,
17:39 you can ask it to generate a picture of
17:42 a face that does not belong to any real
17:45 person and it will do that too. That is
17:48 what generative AI does. Chat GPT and
17:51 Google Gemini are examples of generative
17:54 AI. To be more specific, they are
17:57 generative text AI, which means they
17:59 create new text based on what they have
18:01 learned from existing text data.
18:03 Similarly, there are generative image
18:06 AI, which can create new images from
18:09 learned visual data. One such tool is
18:12 Deli. For instance, you could ask it if
18:14 a famous painter from the past were
18:16 alive today, what would their painting
18:19 of a modern city look like? and the AI
18:21 would generate an image based on that
18:23 prompt. The third type is called
18:27 computer vision AI. This kind of AI is
18:30 used for image and face recognition. AI
18:32 cameras which can identify people,
18:34 objects or license plates fall under
18:37 this category. Beyond these, there are
18:39 many other specialized types of AI
18:42 available today. Robotic AI helps robots
18:44 navigate and interact with the world.
18:47 Speech recognizing AI converts spoken
18:50 words into text. Explainable AI is
18:52 designed to make AI decisions more
18:55 transparent. Planning and scheduling AI
18:57 are used in logistics, project
19:00 management, and more. In some cases, two
19:03 or more types of AI work together to
19:05 perform a more complex task. You might
19:08 have seen deep fake videos where a
19:10 famous actor's face is seamlessly
19:12 swapped onto another person's body or a
19:14 politician appears to say something they
19:17 never actually said. This is possible
19:20 because two AI systems are working
19:23 together in tandem. The first AI system
19:26 is trained specifically to swap faces in
19:28 a video. But if you look closely at
19:30 those videos, you might notice something
19:33 odd, some slight unnaturalenness that
19:36 makes you realize it is fake. Here comes
19:39 the role of the second AI program. Its
19:41 job is to detect flaws in the video.
19:44 Anything that seems artificial or off.
19:46 Once the flaws are identified, the first
19:50 AI makes corrections and produces a new,
19:52 improved version of the video. Then the
19:55 second AI checks it again to see if any
19:57 new mistakes are still visible. This
19:59 process continues in multiple rounds
20:02 with both AIs competing. One trying to
20:04 make the video more realistic, the other
20:07 trying to detect any imperfections.
20:09 Eventually, the outcome is a deep fake
20:11 video so convincing that we can no
20:13 longer tell it is fake. The deep fake
20:15 videos we have seen so far are not
20:18 perfect. But today there are AI tools
20:20 capable of generating even more
20:23 realistic and convincing videos. This
20:25 raises a disturbing truth. Such highly
20:27 accurate deep fakes can potentially be
20:30 used for fraud or other malicious
20:32 purposes. So far we have been
20:34 classifying AI based on the kind of work
20:37 it does. Image recognition, text
20:40 generation, speech processing and so on.
20:42 But there is another way to classify
20:44 artificial intelligence based not on the
20:47 task it performs but on its level of
20:49 capability. This is where we come across
20:53 the terms weak AI, strong AI and super
20:55 AI. And it is in this classification
20:58 that the possibility of AI turning
21:01 against humanity like in the movies
21:03 enters the picture. The AI systems we
21:06 have discussed so far are all designed
21:09 to do one specific task. This kind of AI
21:13 is known as weak AI or narrow AI. Almost
21:16 all the AI that exists today falls into
21:19 this category. We humans on the other
21:22 hand possess what is called general
21:24 intelligence. We can learn a wide
21:27 variety of things, adapt to different
21:30 situations, make logical decisions, and
21:32 solve many types of problems. If a
21:34 computer ever develops that kind of
21:36 broad capability, it would be called
21:39 general AI or artificial general
21:42 intelligence. Sometimes simply called
21:44 strong AI. But let us be clear,
21:47 artificial general intelligence does not
21:50 exist yet. It is still a theoretical
21:52 concept. There are ongoing efforts to
21:55 build such an AI, but no one knows how
21:57 long it will actually take. Some experts
21:59 in the field who are very optimistic
22:02 believe it might take around 20 years.
22:04 Others say it could take at least 50
22:07 years or even more. As of today, general
22:10 AI remains a future possibility, not a
22:12 present reality. There is one more
22:14 category of AI that is even more
22:18 powerful than general AI. It is called
22:20 artificial super intelligence. This
22:22 refers to a future AI that will be more
22:24 intelligent than humans in every
22:28 possible way. Right now, this kind of AI
22:30 is purely speculative, but that does not
22:33 mean it can never happen. Some believe
22:35 that such an artificial intelligence
22:38 might emerge by the end of this century,
22:40 but no one really knows how long it will
22:44 take. The main concern is this. Once
22:45 artificial super intelligence is
22:47 created, it could potentially become
22:50 smart enough to design even better
22:52 versions of itself. And those versions
22:55 could go on to create even more advanced
22:58 versions and so on. If that happens, the
23:02 growth of AI would become exponential
23:04 like a chain reaction. This kind of
23:06 situation is what experts refer to as
23:09 the technological singularity. It is
23:11 this level of super intelligent AI that
23:13 is often portrayed in movies as turning
23:16 against humanity. And it is not just in
23:18 fiction. Many respected individuals have
23:20 expressed concerns about this
23:22 possibility. The late physicist Steven
23:25 Hawking and Elon Musk, CEO of Tesla and
23:28 Space X, have both openly warned about
23:30 the potential dangers of uncontrolled
23:33 AI. But there are also many experts who
23:35 believe such fears are unfounded, at
23:38 least for now. Their main argument is
23:41 this. Today's AI systems do not have
23:43 consciousness or self-awareness. As we
23:46 discussed earlier, when an AI identifies
23:48 a bird, it does not truly understand
23:52 what a bird is. It is simply processing
23:54 numbers and patterns. It has no inner
23:57 awareness or sense of meaning. In fact,
23:58 even human consciousness is still a
24:01 mystery. We do not fully understand what
24:03 consciousness is or how it arises in the
24:06 brain. So, creating a conscious
24:09 self-aware AI is far beyond our current
24:10 capabilities. And without
24:13 self-awareness, the idea of an AI
24:15 deciding to take control or destroy
24:18 humanity is not realistic. After all,
24:21 such desires for power and domination
24:23 are very much human traits. Even if an
24:26 AI becomes smarter than humans, it does
24:28 not necessarily mean it will want to
24:30 harm us. What we should actually be
24:33 concerned about is not the AI itself,
24:36 but the humans who might misuse it. AI
24:39 is a very powerful tool and like any
24:41 powerful tool it can be used for good or
24:44 for harm depending on who is using it.
24:46 That is the real risk we must watch out
24:49 for. In fact, some of this may already
24:51 be happening. AI can be used to subtly
24:54 influence the general public toward a
24:56 particular political orientation or
24:58 ideology. And with many of the AI
25:00 algorithms we interact with in daily
25:02 life, achieving this kind of influence
25:05 is relatively easy. Another common fear
25:09 people have about AI is that it will
25:12 take away jobs. But we have faced this
25:14 situation many times in history. When
25:16 electricity and machines were
25:19 introduced, people feared job losses.
25:21 When computers came, there were similar
25:25 fears. Yet each time humanity adapted.
25:28 We found new ways to work and new types
25:30 of jobs were created. Many experts
25:33 believe the same will happen with AI.
25:34 Artificial intelligence will not
25:37 directly take away your job. But someone
25:38 who learns to use artificial
25:40 intelligence effectively might out
25:43 compete you in the same field. That is
25:45 where we need to adapt to the changing
25:47 world. We must learn how to work with
25:50 AI, not against it. I hope this video
25:51 helped you gain a clear and simple
25:53 understanding of what artificial
25:55 intelligence really is and what it is
25:57 not. If you found this video useful,
25:59 give it a like and share it with someone
26:02 who might find it interesting, too. And
26:04 if you enjoy content that explains
26:06 complex science in a way that anyone can
26:08 understand, make sure to subscribe to
26:11 the channel and of course tap the bell
26:13 icon so you do not miss any of our
26:15 upcoming videos. We have some
26:18 fascinating topics coming soon. Thank you.
