The result of the trending news regarding Artificial Intelligence combined with my passion to learn about all things digital, has recently sparked my interest in learning about AI. This article is a research essay I wrote for my University which I turned into a blog post. In this article you will learn what AI has in store for our future, the benefits, and possibility of harm.
Artificial Intelligence: An Introduction
It is obvious that the world’s future involves extremely high levels of AI integration into society. With its fast-paced introduction and extremely rapid development, AI has already provided many benefits and tools. Already it has also been integrated into the world’s largest corporations. Microsoft, Google, Tesla, and Nvidia have all invested billions into their own AI Programs. However, this raises the question for AI’s future; will the benefits created outweigh the possible dangers? With this, there are two sides to the argument—people who are for AI, and people who are against it.
AI History and Industry Experts
The debate regarding artificial intelligence includes arguments of both optimism and skepticism. The argument dates back to the 20th century when the first intelligence model was invented. The introduction by Alan Turing with his Turing test during WWII was originally what sparked the debate on machine learning and its possibilities. As time went on, there have been bursts of huge innovation, followed by other dry periods with nothing additional added to the industry. In the modern era we are experiencing a ton of innovation, resulting in the topic quickly coming to dominate all forms of the mainstream media. The debate of AI danger and benefits continues to evolve, bringing up societal and ethical conversations about its potential to alter our world. On one side of the debate stand both David Baidoo-Anu of Queen’s University and Leticia Owusu Ansah of the University of Cape Coast. They the immense benefits that AI can provide. In their research, they focus on how it is currently reshaping the industries of teaching and education. On the other side however lies Toju Duke. In her book on this topic, she discusses the potential issues with AI development, while also including different methods for developing it safely.
Current AI Benefits: Education and Teaching
As it currently stands, what are some of the ways in which AI has already begun to innovate and provide benefits? In the research conducted by David and Leticia, they listed five current ways in which AI is already innovating and benefiting education. These include but are not limited to: personalized tutoring, automated essay grading, language translation, interactive learning, and adaptive learning. They later wrote, “Undoubtedly, ChatGPT and other generative AI are already pushing educational boundaries and initiating a significant paradigm shift in existing educational praxis.” This is a huge revolution in technology that can make the job of a teacher seamlessly much easier. One way of thinking about it is that anyone in the world with access to the internet at any time has access to a personalized tutor for any topic sitting in their pocket. Other beneficiary capabilities include natural language processing, text-to-video generation, image generation, and basic decision-making. It is obvious the hugely beneficial impact that AI is already providing, but what are the potential future benefits it can provide when the technology is developed further?
AGI: The Future of Artificial Intelligence
Currently where technology stands, AI can only complete and solve basic tasks. The future goal is to achieve a milestone known as artificial general intelligence (AGI). This next step in AGI would involve critical thinking and complex problem-solving that surpass human intelligence. This would further increase workplace efficiency and productivity. A report by the International Labour Organization said, “Generative Artificial Intelligence (AGI) is more likely to augment than destroy jobs by automating some tasks rather than taking over a role entirely.” Because of this, anyone or anything involved in the creation of AGI has a huge amount of responsibility.
So far we’ve learned what AI is, the benefits of AI, and its the future has in store. Let’s take a look at some of the possible dangers of it.
Possibilities of a Dangerous AI
People often associate the dangers of AI with a Robot takeover. This is because various fiction films created by Hollywood such as The Terminator and Age of Ultron depict AI dangers in this way. Although the chance of a Hollywood movie coming to life is only a hypothetical risk, AI experts seem to have many concerns. Geoffrey Hinton is widely known as the godfather of AI because of his countless awards, years of research in the industry, and decade of work conducted under Google’s AI program. In a 2024 interview he was asked how AI could kill humans. He answered by saying, “If it gets to be much smarter than us, it will be very good at manipulation because it will have learned that from us, and there are very few examples of a more intelligent thing being controlled by a less intelligent thing.” Because of the possible dangers AI presents, it is imperative to develop it safely.
Responsible AI Development
Toju Duke is one of the biggest spokeswomen in the conversation about responsible AI development. Like Geoffrey Hinton, she worked under Google’s umbrella, specifically as the Program Manager for Responsible AI. In her book, she wrote that ethics are crucial to implement into AI training data. Applying ethics can ensure fairness and remove bias in the outputs. Later in the book, she identified three ways in which AI can be developed to mitigate harm. These Include:
- Autonomous learning with Benign intent
- AI learns and develops further as it goes. Because of this, it is very unpredictable with its answers which can result in harmful outputs
- Human controlled with Benign Intent
- When it comes to training on new data, AI often outputs large amount of bias in its answers. Under these conditions, the models perform very poorly.
- Human Controlled with Malicious Intent
- This method of AI training is when it is utilized to do tasks to cause harm or chaos. In some cases, this has even caused AI to go rogue
Another training method which I find very interesting is called Zero-shot learning. This is when an LLM is asked to identify and recognize an object or group of data which it was not trained on during learning. Essentially, the AI has never seen the object or data that it is asked about. When this is implemented, AI is made to generalize and make assumptions by itself without access to any data. Because of this, it can make unfair assumptions that cause harm. If trained further without addressing the false outputs, AI can become dangerous.
12 Responsible Steps to Mitigate AI Dangers
The risks of dangerous AI training needs to be mitigated so that later on it does not include bias, unfairness, and other harmful results that can lead to danger. In chapter 5 of Toju’s book titled “Safety,” she identified 12 strategies to mitigate dangerous development. They are as follows:
- Data must be continuously improved so that LLMs do not develop unfair stereotyping and biases.
- In order to identify its limitations, track specific things that are included in the training data so that you can reference them and identify things that might be missing.
- Eliminate any misleading or harmful outputs to fine-tune the model.
- Perform distillation on LLMs using targeted datasets that have desired responses to sensitive prompts.
- The current performance of LLMs needs to be trained through different languages as they are mainly trained in only the most common (mainly English). Doing so would increase language performance.
- Adding more inclusive product designs also helps to remove unwanted statements.
- One way to increase efficiency and accuracy is to combine smaller language models into LLMs.
- To mitigate the environmental impact that large models have, companies should adopt compute-efficient architecture as well as sustainable energy.
- AI should be regulated by the Government through policies that forbid disinformation and malevolent intent.
- To reduce malevolent harms, AI companies should responsibly open-source the LLMs.
- When innovating their models, incorrect facts should be sifted through.
- Finally, the last way of fine-tuning models to mitigate harm is to increase privacy. Doing so would prevent personal data from being used in training.
According to Toju, implementing all of these methods should make outputs much less harmful.
Trusting Large Corporations with AI
As discussed in the beginning, many of the largest corporations are investing billions into their AI programs. Even more so, it recently surfaced in the news that Sam Altman, asked for a 7 trillion dollar investment into OpenAI. Because of this, many people question whether large corporations can be entrusted to develop AI safely. Take a look at Google for example. One of Google’s AI principles on its home page is that “We will seek to avoid unjust impacts on people, particularly those related to sensitive characteristics such as race, ethnicity, gender, nationality, income, sexual orientation, ability, and political or religious belief.” This is not always shown in the outputs of their AI. Circulating in the news recently was a lot of controversy regarding their new AI, Gemini, which introduced a text-to-image feature. According to news outlets such as the New York Post, the feature was shown to output “Woke” images. They also wrote that It created pictures that were “factually or historically inaccurate.” For example, when prompted to create an image of the pope, it created a picture of a “Southeast Asian woman and a black man wearing holy vestments.” Even though every Pope in history has been a white male, the AI went out of its way to promote diversity, showing its bias towards white men.
Conclusion
Understandably, there is a large debate about AI when it comes to possible benefits and dangers. In today’s modern era, the exponential advancement of AI technology has brought it to the center of mainstream media. Already in society there are so many benefits, and with its advancement there will be more to come. Despite this, there is still a large possibility that AI will turn out to be dangerous. If not developed with extreme caution, the future for AI could be dark. Because of this, there is a huge responsibility for developers to mitigate these risks in order to prevent a dangerous AI system in the future.