Limitations Of AI

  

Introduction to AI

The science and engineering of creating intelligent devices, particularly intelligent computer programs, is what John McCarthy, the father of artificial intelligence, defines as artificial intelligence.

Artificial intelligence is a technique for teaching a computer, a robot that is controlled by a computer, or software to think critically, much as smart people do.

It is possible to create intelligent software and systems by first studying how the human brain works, as well as how people learn, make decisions and collaborate when attempting to solve a problem.

As artificial intelligence develops quickly, many businesses are eager to experiment with and evaluate what is currently on the market. Others, however, are not persuaded by artificial intelligence due to purported ethical concerns that could lead to accountability in a certain way. This thesis will address the limitations of artificial intelligence while describing how it is employed in several industries, including law, healthcare, the military, and others.

For AI to function, machine learning (ML) represents the most important technology. However, it can't work by itself. One technology that is incorporated into the algorithm to make AI function is the optical character reader (OCR). NLP, or natural language processing, is frequently required for the analysis and manipulation of language.

Robotic Process Automation (RPA) is also used to configure tasks. These technologies are still necessary even if they are far more advanced than they were previously, and this trend will continue as development calls for the adoption of even more advanced technologies. Additionally, the majority of AI applications currently have a limited scope, such as image recognition. This program does its job, but not in the sense that it will recognize any image that is available.

The current state of AI's emotional intelligence is a serious flaw (Bellapu, A.,2021). NLP is the basis for comprehending the meaning of human discussions, but since human feelings and emotions have unique profiles, communication is still distinctive communication. However, circumstances will change in the near future given that there is already a machine that can recognise facial expressions. AI needs to be controlled as it passes through such an algorithm, but one must manage the creation of these algorithms. The Microsoft chatbot Tay is the greatest example of mismanaged AI (Liu, Y., 2017). The goal of this chatbot was to interact with users and have a lighthearted conversation. Tay, an AI that operates by data feeds to it, was regrettably shut down soon after its release of immoral tweets.

Figure: Post from Tay, the Chatbot Twitter’s Account

These facts demonstrate that artificial intelligence can only perform what has been programmed into it. The goal of these studies is to create AI more like humans, however, the main distinction is that there won't be any empathy and feeling toward users. As a result, AI cannot make second decisions in many circumstances. Although artificial intelligence has made considerable strides, the human mind's adaptability is significantly more intricate, which explains why even after years of advancement, scientists are still unable to replicate how the mind functions.

Although there are already numerous robots or AIs performing all of the labor previously performed by humans, the gap between AI and the human brain continues to exist. However, this does not change the fact that AI still requires humans to function. Although it may be very helpful in a variety of fields, including manufacturing, healthcare, defense, and other services, AI still has its limitations, and they will continue to exist for many years to come. How far can humans go in order to overcome these constraints?

Limitations Of AI

Dependence on massive amounts of data

deep learning algorithms is affected. Sadly, bias is likely to exist also when data is readily available.

To address this issue, one-shot learning research has been conducted.

restricts the use of supervised learning algorithms to a small number of issues where labeled data c is either already accessible or where the value of the answer is high enough for businesses to invest in the preparation of semi-manually labeled data.

Unsupervised learning techniques are being enhanced in response.

Limited adaptability

A perfectly functioning system may become unresponsive to minor changes in the issue, necessitating a skilled adjustment to return it back to normal.

Research is being done to address this problem in the area of transfer learning. Recently, AlphaZero was capable of quickly learning the board games Go, Shogi, and Chess.

Inadequate transparency

deep learning algorithms are affected. Deep neural networks are intricate mathematical structures, thus it is difficult for people to simply understand their reasoning.

To remedy the lack of transparency, local interpretable model agnostic explanations (LIME) and focus strategies are being developed. This McKinsey essay provides some helpful illustrations of how these strategies operate.

Inability to integrate earlier knowledge

While some AI techniques rely only on stored prior knowledge, others, such as deep learning, have no way of incorporating expert summaries. This is a serious limitation when developing systems for fields where the majority of occurrences can be accurately explained by existing research.

Cost

Mining, storing, and analyzing data will be extremely expensive in terms of technology and energy utilization.

The GPT-3 model's projected training expense was $4.6 million. The training costs for a model like a brain would be significantly greater than GPT-3, coming into it at roughly $2.6 billion, according to a different video (see below).

Additionally, hiring them will hurt these organizations' finances because experienced engineers in these sectors are now expensive to come by. disadvantaged newer and smaller businesses here as well.

Defamatory Attacks

AI isn't particularly suited to adapt to changes in circumstances because it isn't human. An autonomous vehicle, for instance, might drift into the wrong lane and wreck if tape is simply applied to the incorrect side of the road. A person could not even notice the tape or respond to it. While the driverless vehicle may be considerably safer under typical circumstances, it is these extreme occurrences that should worry us.

This failure to adapt draws attention to a serious security weakness that has not yet been adequately fixed. While "tricking" these data models occasionally can be entertaining and safe (like mistaking a toaster for a banana), in severe circumstances (like defense objectives), it could endanger lives.

There is no agreement on privacy, ethics, or safety.

There is still work to be completed in determining the boundaries of AI use. Given current constraints, safety in AI is crucial, and immediate action is required. The majority of AI detractors also raise ethical concerns about its implementation, not just in terms of how it eliminates the notion of privacy, but also from a philosophical standpoint.

We believe that intelligence is innately human and distinctive. That exclusivity could seem contradictory to give away. One of the frequently asked questions is whether robots should be granted human rights if they are able to perform all tasks that people can, effectively making them equal to humans. If so, how do you define the rights of these robots? Here, there are no conclusive solutions. Given the recent development of AI, the study of AI philosophy is still in its infancy. I'm eager to watch how this area of AI progresses.

Algorithm bias

An algorithm is a series of instructions that a computer follows to execute a task, some of which may or may not have been developed by a human programmer. However, we never rely on algorithms if they are flawed or prejudiced since then you would only see unfavorable outcomes. Biases primarily result from the partial design of the algorithm by programmers, who favored some desirable or self-serving criterion. Large platforms with algorithms such as search engines and social media sites frequently have algorithmic bias.

For instance, in 2017, a Facebook algorithm has established an algorithm to delete hate speech. However, it was later discovered that the algorithm left hate speech against white men in place while allowing it against black youngsters. The algorithm permitted these hate statements because it was built to exclude just general categories like "whites," "blacks," "Muslims," "terrorists," and "Nazis," rather than particular subgroups of categories.

AI’s “Black Box” Nature

The ability of AI to learn from massive amounts of data, find underlying patterns and make data-driven judgments is well known. The AI system lacks an articulate or describes how everything arrived at this decision, which is a significant shortcoming despite the fact that it consistently and swiftly generates accurate findings. As a result, the question of how we can trust the system in extremely sensitive areas like national security, governance, or risky commercial endeavors is automatically raised.

Conclusion:

The development of artificial intelligence is accelerating quickly, and many businesses are eager to test its potential and assess its marketability. Others, on the other hand, are skeptical of artificial intelligence because they believe there are legal issues that could lead to transparency in a particular way. This research will examine the limitations of artificial intelligence as well as how it is employed in numerous fields, including law, health, and the military.

The possibility of job loss, particularly in industries like manufacturing and agriculture, appears to be AI's biggest fear right now. Employees who are AI-enabled have had some advantages over their human peers. They never get tired since they don't experience mental or emotional reactions. The likelihood of errors is greatly decreased. Since AI presents amazing opportunities across several businesses, it also introduces new challenges. Before then, problems with AI deployment were frequently blamed to staff members' disinterest or lack of participation in the procedure. on a continuous basis.

Companies typically embrace AI when they see the advantage of doing so but fail to take into account the processes and rules that make it up; this necessitates a complete understanding of all aspects of AI processes, from data gathering to transmission of exposed interactions. Always developing a long-term business strategy before starting expensive and resource-intensive AI efforts is the obvious answer.

Every AI design should take security and compliance into account. If not, businesses might try to reverse-engineer the answer or build from zero to come up with a solution that would be expensive artificial intelligence bias, and ineffectual.

Before new models are formed, artificial intelligence specialists who developed the technology should evaluate their data gathering, apply past programs that established 49 rules for reporting models that datasets, and ensure that biases are removed. AI is currently everywhere. Since the introduction of Google in 1997, artificial intelligence has met a variety of personal, amusement, and technical demands online. In other fields, deep learning and machine learning have been shown to offer significant advantages.