Artificial Intelligence Limitations

Artificial Intelligence which is sometimes called as Machine Learning can be defined as ‘the valuable part of machine learning is predictive modeling where we use historical data to train a model and use the model to make predictions’.

It is not restricted to predictive modeling alone. It can also encompass pattern recognition. It is employed in so many varied fields : autonomous systems, human interaction, goal driven systems and so on.

Its ability to look for a pattern or anomaly from a large set of data provides it superior ability when compared with humans. It can interact quite well with human language where it is known as Natural Language Processing. Simply put Natural Language Processing is the process of using specialized computer algorithms to identify key elements in everyday language and extract usable meaning from unstructured input.

A working example for the NLP is the ‘TargetWoman Health App’.

Slowly and steadily AI has started permeating into our daily life.

The AI technology to create autonomous vehicles exists now. So does reasonably mature facial recognition AI . The whole technology is almost ripe that even mobile phones use facial recognition to unlock.

The Riddle of AI

Project Maven – a project initiated by the Pentagon that uses AI to differentiate people from objects uses the enormous data it had compiled from its drone footage. It may have started a controversy about the eventual use of autonomous entities for military purpose.

What does it take to connect the dots and weaponize the AI into creating an autonomous unit to eliminate targets ?

Not a lot.

Consider the following scenario:


Scenario: One of the key intelligence agencies of the government alerted the Autonomous Sniper Unit (ASU) that a high value target has been seen near a public park. The ASU called home its human handler who cross checked with his superiors in the intelligence agency for confirmation. There was palpable tension in the air as he waited for the agency to give the ‘execute’ order. It was not a decision he can take on his own and this will be the first time, the intelligence agency has assigned the task to an ASU.

The final call for action came after 30 minutes which appeared like a long time. He sat before the laptop in a darkened room. The ASU was triggered to Phase 1. The complete operation is divided into 3 phases.

Phase 1: ASU is assigned the target by showing a series of photographs. It uses advanced facial recognition software to identify the target from a cluster of high resolution cameras mounted in a SUV. It can identify the target even if he wears disguise as it uses AI which has been trained with thousands of images. It has been field tested with hundreds of samples and the final version is so advanced that it identified the ‘target’ every time in all the tests. There is no way it can fail.


Phase 2: Once the target is acquired in its vision, the ASU will track the target in the SUV. It has additional autonomous miniature drone which will follow the target so that the main ASU can comfortably stay out of sight and still track the target. It has been given the ‘execute’ order. It is upto the unit to identify the exact moment when it will use its specially modified sniper rifle to kill the target – without causing any collateral damage to civilians nearby.

As with human snipers, the process is very complicated. The sniper is helped by the spotter, who assists in observation of the target, surrounding area and atmospheric conditions. In ASU, it uses an array of cameras – both visual and IR as well as other sensors which supply data on atmospheric conditions. When the ASU was trained, it has to score bullseye 100 % . It was tested in almost all conditions – desert, snow, high mountains where the atmospheric air pressure is low and at sea level. The AI was thoroughly trained along with the modified high power rifle until each part worked with the other with the utmost efficiency. Again there was nothing left to chance.

It also has thermal imaging system so that it can track the target even if the target is hidden. The AI uses complex algorithm to precisely identify the target in all conditions based on visual cues.


Phase 3: In this critical phase, the ASU waits for the final ‘execute’ order to complete the mission. The AI at this stage works out the trajectory of the bullet and ‘calculates’ the exit point of the bullet behind the target. It makes small calculated maneuvers in the ‘position of the gun’ and makes required maneuvers in the SUV as needed. The gun will emerge from the sun roof for the shot briefly to take the kill shot. Once the human handler tells the ASU, the shot is taken. It records every minute detail for later analysis. The ASU returns to base after its successful mission.


The ASU, at this stage is a figment of our imagination now. But the technology for all its parts exists today. They are readily available in various forms. All one needs is to connect the dot and collate the various hardware and AI to complete such an autonomous Sniper units.

Artificial Intelligence Limitation:

Artificial Intelligence primarily depends on a set of data to train (machine learning). To quote a real life example, when we started the NLP work for the medical application, we hunted for a large set of words/phrases related to medical/health topics. We converted volumes of medical books to the format required by the software. We spent a long time to format the data as needed. The key here is the data. Unless you start with the right data and preprocess the data and massage that for learning, you will end up with poor results. It would end up in GIGO –Garbage in Garbage Out.


The computing power required to learn from millions of words in real time is humongous. But the technology is improving in leaps and bounds. These days you get the complete AI backed into a slice of silicon chip for the specific task. You get to see them even in mobile phones.


Now that we have seen AI is claimed to be used in all most all areas of human interest, does it really reach a stage where it can be a threat to humanity like they show in movies ?
Limitations of AI
All intelligent creatures need a raison d’être (reason for living). After satiating the basic physiological needs, the creature will look for higher levels of gratification for its existence. Maslow’s hierarchy of needs as a motivational theory of psychology lists ‘ a sense of belonging and a sense of accomplishment’ as important needs after physiological needs.

Most mammals require stroking as part of their survival process. Observe how the cat starts purring when you stroke the cat. The psychological gratification is essential for the complete health of the mammal. If not nourished, it will wither away.


We created Artificial Intelligence to help us. A lot of people assume that one day it may reach a stage where it connects the dots and decides to perpetuate its existence without any interference from the human master.

There are others who point to the complete lack of raison d’être for its continued existence. It has no psychological aspirations or it has no quantifiable desire to gratify. Besides the key reason why it will not go against its baked in instructions is the St.Augustine’s theorem. It simply states that even as humans we have a definite built in limitation to our intellectual quest. So any ‘creation’ that we come up with will equally suffer the parent limitation. Thus AI will not break away from the limitations we have imposed deliberately or inherently.

Artificial Intelligence (AI) and Health Information

Artificial Intelligence (AI) seems to be everywhere these days – from driving your autonomous car to detecting diseases in clinical conditions. Artificial Intelligence has permeated almost all fields and has evolved into a basic component of business growth. The term AI encompasses a wide spectrum of converging technologies – Deep Learning, Machine Learning, neural networks and Natural Language Processing (NLP).

Definition of Machine Learning:
Jason Brownlee sums up like this: ‘The valuable part of machine learning is predictive modeling where we use historical data to train a model and use the model to make predictions’

Tom Mitchell in his classic book on machine learning says:

‘The field of machine learning is concerned with the question of how to construct computer programs that automatically improve with experience’

Bereft of confusing buzzwords, AI is basically a technique to look for a pattern from training on a set of data. The key here is the data. Unless you start with the right data and pre process the data and massage that for learning, you will end up with poor results.

Moravec’s paradox: Hans Moravac discovered that high level reasoning requires very little computation, but low-level sensorimotor skills require enormous computational resources. In another instance some researchers found it odd that they were successful at writing programs that used logic, solved algebra and geometry problems and played games like checkers and chess. Logic and algebra are difficult for people and are considered a sign of intelligence. They assumed that, having solved the “hard” problems, the “easy” problems of vision and commonsense reasoning would soon fall into place. They were wrong and one reason is that these problems are not easy at all, but incredibly difficult. The fact that they had solved problems like logic and algebra was irrelevant, because these problems are extremely easy for machines to solve.

To quote from a personal experience, I was having difficulty finding the right corpus to train our AI for our Athena -Simplified Health Information project. I was cribbing about this Artificial intelligence problem when my partner blurted out – ” that is because it is artificial; it is not a natural intelligence”.

I usually throw a piece of meat taken from the refrigerator to move my 1 year old German Shepherd dog to the backyard every evening. On most days, he would sit there watch me fetch the piece. I decided to fool him once and touched the piece of meat but instead fetched a sliver of cabbage. He couldn’t have noticed the switch as I did it in a slight of hand and threw the piece out. He didn’t move from his position. He must have anticipated my switch somehow. I was clueless – about how a dog could predict his owner’s behavior without any extra sensory input.

May be there is an inherent limit to our creative prowess – something like St.Augustine’s conundrum. There was a time St.Augustine sat near the seashore contemplating the mystery of the Holy Trinity when he saw a little child running back and forth from the sea to a spot on the seashore. The boy was using a shell to carry water from the sea and fill a small pit that he had made in the sand or it so appeared to St.Augustine.

Augustine came up to him and asked him what he was doing.

‘I’m going to pour the entire sea into this hole’ the boy replied.

‘What’? said Augustine. That is impossible, my dear child, the sea is so vast and the shell and the hole are so little.

The boy replied It would be no more impossible than what you have been contemplating. It would be easier and quicker to draw all the water out of the sea and fit it into this hole than for you to fit the mystery of the Trinity and His Divinity into your little intellect; for the Mystery of the Trinity is greater and larger in comparison with your intelligence than is this vast sea in comparison with this little hole.

Natural Language Navigation
We, at TargetWoman have been working at the Natural Language Processing from the year 2004 onwards – around the time we started. Simply put Natural Language Processing is the process of using specialized computer algorithms to identify key elements in everyday language and extract usable meaning from unstructured input.

Microsoft says this about languages: “Understanding’ language means, amongst other things, knowing what concepts a word or phrase stands for and knowing how to link those concepts together in a meaningful way. ‘

For example, the word “cell” can mean different things to different people. It can mean a prison cell, compartment in a honeycomb, smallest organizational unit of a movement, a component of a battery, mobile phone or the smallest structural unit of an organism to name a few of the definitions. The surrounding words help to exact the true meaning of the word in relation to the context. In other words, you will have to decide the meaning of the phrase only after analyzing the entire sentence.

We started working on the concept of Natural Language Progression way back in the year 2004 – long before the term Artificial Intelligence became a buzz word. We worked on the premise that to convey an idea or a thought – words are only a transport layer. It is not the word per se that conveys the idea/thought but the collection of words, their proximity to each other and the context. With that said, we turn our attention to using computers to understand the logical progression of our language which some call as Natural Language Processing. We have created a working model which can understand the subtlety of the language we speak – at least limited in context to health and medical topics.
Sample this: Our Natural Language Processing Engine when triggered by the ubiquitous word ‘sex’ comes up with the following words – pregnant with meaning (Natural Language Association):

Natural Language Navigation:
In TargetWoman, we have thousands of pages dedicated to different topics of interest to women. It is always not so easy when it comes to navigation inside such large repository of content. A conventional way of navigation is to split the content into various clearly defined topics like: health, home improvement, careers and travel to name a few. But every topic may have dozens of pages or some even in hundreds of pages. Most large websites addressed this issue with a site wide search facility. We implemented this search engine at the outset itself. But it was not enough as it turned out from the server log. A better method needs to be found which would take into consideration how people navigated inside a site. This is how we turned our attention to Natural Language Navigation. We fed the key topics of every single page to our Logical Progression Engine – another term for Natural Language Navigation. This in turn sorted the content based on ‘keywords’ and their frequency of occurrence. The latter part provides a strong correlation to the keyword with the nature of content. This is how a typical search engine evaluates/indexes content. The data was updated everytime a new page was published.
Artificial Intelligence for Website navigation
A real world example here:

If you click on the ‘Browse by topic’ for an article: Example:

“Browse by Topics: Aerobic+Water+Exercise” would return the following content:

1. Home Exercise Equipment
2. Swiss Exercise Ball
3. Morning Exercise and Metabolism
4. Physical Fitness Exercise
5. Abdominal Exercise
6. Circuit Training
7. Xiser Workout

From the above abbreviated list, you can see that the NLN system picked out only content related to the physical exercise alone and did not pick content from Parenting with the keyword exercise (home exercise for students).

We would discuss how we collected the vast data and how we created the algo for this project in subsequent blogs.
Until that time, try our NLP here:

You can discuss in the comments section how our Natural Language Navigation helped you to find information on health topics.

Our Parent site TargetWoman - the leading women portal presents painstakingly researched extensive information in the form of thousands of condensed pages. It offers the widest and the most detailed information on subjects women care.