Taming the Artificial Intelligence Tiger: Embracing Change & Navigating Challenge

We are so used to using Siri/ Alexa, ChatGPT, navigation tools, language translation tools, autocomplete on emails in our daily lives, that it is difficult to imagine a day without them. All these have Artificial Intelligence at their core, and there is no doubt that AI adds a lot of value in our day to day lives. At the same time, stalwarts of AI like Geoffry Hinton & Sam Altman, have raised their apprehensions about the future use of AI. This was the trigger that pushed me to share my thoughts on this. I start by giving some background to give a headstart to someone new to the field of AI, and then delve into the applications and challenges of AI.

Introduction to AI

Since the advent of computers, humankind has witnessed an unprecedented transformation across various spheres of life. It is impossible to imagine the world of banking, communications, travel, education and so many other areas without computers. So what is different in Artificial Intelligence?

Let me set some context by explaining the basic difference between classical programming and artificial intelligence & machine learning.

In classical programming, the programmer tells the machine what to do in terms of logic and sequence of steps to take, and the machine does exactly as it is told to. So essentially, it is a rule-based system, and hence deterministic, meaning the same input would always provide the same output. The main input in classical programming is the ‘how’ (rule).

While a small subset of AI systems are rule-based, most AI systems learn from data using Machine Learning (ML), meaning we don’t tell the system how to give the result. Rather, we provide multiple inputs of data, and associated results, and tell the system to figure out how! So the main input element is now the ‘what’ (data the system needs to learn from), and not the rule (‘how’).

To summarize in one sentence, classical programming focusses on knowing the rule, while ML focusses on finding the rule.

Merriam-Webster dictionary defines intelligence as the ability to learn or understand or to deal with new or trying situations. Hence, the term intelligence in Artificial Intelligence (AI). Artificial, because we put bounds on what directions the machine can think in by giving data to learn from. More on that later in the article.

How does ML Learn?

Before getting into the various kinds of AI, let us broadly understand how an ML system learns the rule. To put it very simplistically, an ML system basically looks for patterns in the input data we provide, and tries to form an equation that generates the output from the associated input. The system then finds the coefficients of that equation. Once that is done, within a reasonable error margin, we say the model is trained. For any new data input, the system would use that equation and use the coefficients to provide the output based on the input data.

Applications of AI

AI can be immensely useful in processing different types of data, be it in the form of tables, graphs, language/ text, images, audio or video samples etc. The applications of AI using these different data types can be categorized under two broad categories.

• Inferencing: When we use AI to get some insights into that data, but don’t generate any new data.
• Forecasting/ Generating: When we use AI to generate data that doesn’t exist, eg forecasting the weather for the next 5 days or generating an image of what a city would look like in the future, language translation.

Let me give some concrete examples of the use of AI across various domains.

In healthcare, AI algorithms can analyze vast amounts of medical data and assist in diagnosing diseases and predicting outcomes. AI-powered medical imaging systems can provide first level assessment by analysing the ECG, X-ray or CT scan of a patient, even before the doctor sees the patient.

AI has also transformed the business landscape. Companies across industries are leveraging AI to enhance customer experiences & optimize operations. Intelligent chatbots and virtual assistants have become common, offering immediate support and personalized recommendations. AI-powered tools ?help businesses make data-driven decisions, identify trends, and forecast market demands.

AI has played a pivotal role in the development of autonomous vehicles. Self-driving cars are becoming a reality, with AI systems enabling real-time decision-making, navigation, and collision avoidance. Video and Image processing technologies are enabling us to enhance road safety, reduce traffic congestion, and improve the overall efficiency of transportation systems.

In the financial world, AI algorithms can analyze vast financial datasets, detect anomalies, and identify patterns that humans might overlook. This helps in fraud detection, risk assessment, and algorithmic trading.

AI is revolutionizing the field of education as well. Intelligent tutoring systems can adapt to students' individual learning styles, providing personalized feedback and guidance. AI-powered language learning platforms help users practice speaking and writing skills, providing instant feedback. Virtual reality and augmented reality technologies, coupled with AI, are transforming the way students learn by creating immersive and interactive educational experiences.

AI is enabling companies to do real-time sentiment analysis based on social media feeds, translate languages in real-time to enable effective communication by breaking language barriers. Content companies are using AI to generate and summarize content, and even create personalized content.

AI has revolutionized the entertainment industry, by creating content, providing personalized recommendations, creating algorithms to play games with humans, and enabling Augmented reality & Virtual Reality (AR/ VR) systems. Deep Blue and AlphaGo are significant advances in this area.

On the personal front, virtual assistants, autocomplete (and now even auto-respond) on mails, navigational apps, applications like Alexa & Siri are all examples that have AI at their core.

Consider the world’s first robot citizen, Sophia, as an advanced AI outcome that combines most of these applications including NLP and text to speech to talk, computer vision to see, and advanced robotics to walk.?

But AI has its own share of problems as well like fake content creation, and data-protection and data privacy issues. Let us explore some of these challenges next.

Challenges of AI

To err is human, to really mess up, you need a computer.

I came across this funny line on the Internet. This coupled with the buzz around AI and the concerns raised by people like Geoffrey Hinton & Sam Altman, set me thinking about really how big a mess can a computer create.

Let us look back at classical programming. If a programmer tells the machine to do something which was incorrect, or there is a flaw in the logic that he gave to the computer, you get bad results, but almost all of them are explainable and attributable to a bad code/ logic, and hence correctable. The key is the term explainable because you know the rule. So if the logic was wrong and a rule deducted 20 cents instead of 10 for each transaction, it would do the same always. And this can be corrected if you can find the flaw in the rule.

Contrast this with AI, where we don’t even know the rule, in most AI systems. The whole process is a black-box approach. Given the complexity and scale of AI models, how an algorithm reached a particular result is beyond human comprehension (at least for now). Just to give an idea of scale, an image recognition model has parameters (analogous to the coefficients in the equation) in the range of about 500 million, which are learnt by the machine. NLP systems on the other hand have a much larger number of parameters, eg in the case of ChatGPT-4, it is close to a whopping 170 trillion parameters! Imagine this complex rule (equation with so many coefficients), and the ability of the human mind to process and understand that equation.

Making AI explainable is an active area of research in the field and is called explainable AI (XAI). Till that time, let us just remember Murphy’s rule (what can go wrong will go wrong), and put safety checks in place.

The artificial in Artificial Intelligence

Another problem with AI is the way AI learns. Recall that the system learns from the data you give to the system, thereby limiting its ability to learn in all directions. Instead, the system thinks that everything is contained within the data that is given to the system as input, and is the absolute truth. This can lead to the presence of prejudice or favouritism in the AI systems, which is called AI Bias. As an example, if a system is trained by giving data of credit card defaulters where 80% of the students defaulted, the system is very likely to predict a student as a defaulter, because, for the system, 80% of students are defaulters. So a system can acquire intelligence, but that is limited by what it has seen.

Bias in AI can perpetuate discrimination and reinforce existing prejudices. Any intentional or unintentional bias in systems can be used to further existing beliefs or biases, potentially putting society at risk.

I’m not a robot

I find it funny when a computer asks me to select a few boxes containing traffic lights to identify that I am not a machine. But there is no way for me to ask when I am talking to someone that I am talking to a real person or looking at a fake image or a deep fake video.

The generative examples of AI that we discussed earlier, can create fake images from scratch which can appear very real. Deep fake videos can superimpose audio on a separate video, and edit the video to sync the lip movement to create a deepfake, where you can essentially make any person say anything you want in his voice.

This can have far-reaching implications on law and order and society in general.

Computer vs Computer

Imagine a scenario where AI generates a fake image. Then we feed this image into another AI engine to detect if it is a fake image, and provide feedback to refine the generated fake image. This process can be iterated multiple times till even the AI detection algorithm cannot detect if an image is fake. This is the basis of Generative Adversarial Networks (GANs) in AI, which could pose another threat in the use of technology, where humans become mere spectators watching a machine fight a machine.

Hackers and criminals have started to use these systems to prevent credit card fraud detection engines, creating fake images, deep fake videos and even launching cyber security attacks.

Who is responsible?

Another aspect to be looked at is the accountability part of AI. As an example, if a driverless car causes an accident, laws have to be drafted to cater to scenarios, which are very different from the conventional world.

Data Privacy Issues

These issues are not specific to AI, but wherever there is data, we need to have a mechanism in place to protect data and ensure the privacy and right of the individual to decide if he wants to share his data with others. The problem is just more severe, because you can decipher a lot more by finding patterns and putting together information from multiple sources, even inadvertently eg user behaviour

A machine doesn’t have a heart or ethics

One crucial aspect that distinguishes humans from AI systems is the presence of emotions. Emotions play a fundamental role in human experience, influencing our behaviour, interactions, and decision-making processes.

The absence of emotion in AI systems has both advantages and limitations in various applications. As an example, AI's lack of emotions allows it to make decisions based solely on objective data and logical reasoning, free from human biases or emotional biases that humans may exhibit.

At the same time, Emotionally-driven tasks, such as understanding and responding to human emotions, can be challenging for AI systems. Although AI can analyze facial expressions, voice intonation, and textual sentiment, it lacks the depth of understanding and empathetic response that humans possess. A machine can talk to you using very polite language, and a very apologetic tone, but it doesn’t have the apathy or ‘understanding’ of the issue at hand. AI systems may struggle to accurately interpret emotional cues or context in human interactions, leading to misunderstandings or inappropriate responses.

Intelligence is also defined as the ability to comprehend and understand, and we are talking Artificial intelligence!

The Ethics Debate

Continuing the thought in the previous section, a machine would do what it is trained for. So, if you train a machine to do illegal tasks, it will do them without worrying about the ethics or legality of the issue. As AI continues to advance, the responsibility lies with human developers, researchers, and policymakers to ensure ethical considerations are in place.

We don’t know everything

Recall the discussion on explainable AI. In a system, where we don’t know fully how it works, we definitely don’t know what can go wrong. There can be unintended consequences of using AI, and many companies of repute have already started to draft policies around responsible use of AI. The only catch is these policies are specific to each company, and what looks like responsible behaviour for one company might not be true for another.

We have just seen Weak AI

There are three broad types of AI

• Narrow or Weak AI, which is designed to perform a specific task or a limited range of tasks, and lacks general intelligence. All examples we know till date (Siri, Alexa, ChatGPT, AlphaGo, Watson or even Sophia) of AI fall in this category.
• General AI (also known as Strong AI), refers to AI systems that possess human-level intelligence and can understand, learn, and apply knowledge across various domains. It is still a hypothetical concept, and there is ongoing research and development in this area.
• Superintelligent AI, goes beyond human-level intelligence and refers to AI systems that surpass human capabilities in virtually every aspect. Superintelligent AI is also a topic of theoretical speculation and research.

Given the active research in the field of AI, and considering ChatGPT, AlphaGo, Watson and Sophia as rapid steps towards General AI, we need to tame the tiger now, before the tiger becomes more powerful.

Conclusion

AI is a powerful tool that has the potential of solving multiple complex problems, and freeing a lot of human mind hours to do something more and better. However, we need to have a responsible reaction, clear guidelines, and regulatory and legal frameworks in place to ensure we don’t end up creating a monster, that ends up harming its own creator. Governments, legislators, regulators, innovators, technology providers, and even users of technology need to come together to keep the tiger tamed.

Hope you found this worth your time. Happy to hear your thoughts and feedback.