When I sat down to write this post, I had intended to tackle the topic of generative artificial intelligence (genAI) controversies, and more specifically, what the benefits and drawbacks were based on its current uses, risks, and potential future trends. But to talk about the impacts of genAI, both positive and negative, we first have to be on the same page about what exactly AI is, and how genAI is different.
And as it turns out, the phrase “artificial intelligence” can mean a lot of different things, depending on who and how you ask. So today I want to dive into what exactly constitutes AI.
To define what we now call artificial intelligence, it helps to first look at its origin. And while one could argue that AI models are based on logic and statistics fields that go back centuries, most historians, computer scientists, and mathematicians agree that the modern concept of AI dates to the 1950s. According to Stanford’s One Hundred Year Study on Artificial Intelligence, leading logicians, mathematicians, scientists, and philosophers of the time met at a 1956 Dartmouth College workshop seeking to answer the question of whether computers might one day be capable of human-like thought or intelligence, officially coining the term “Artificial Intelligence.”
But exactly what qualified depends on how you define intelligence. Known as the father of modern computing, Alan Turing published his essay “Computing Machinery and Intelligence” in 1950, which theorized a process for computers to learn and communicate their findings. In this essay, he designs an “imitation game,” later called the Turing test, which would have a human researcher ask questions of an out-of-sight human and a computer, and if the researcher was unable to determine which was the computer, it was deemed to be capable of human intelligence (read more about the Turing test here). In this way, AI can be thought of as programs capable of producing human-like responses.
Around the same time, in 1959, computer scientist Arthur Samuel designed a computer program that learned how to play checkers through trial and error, which became one of the founding examples of a branch of AI known as machine learning. Instead of a computer programmer creating code that used if-then logic to code in all possible moves and give directions based on prior moves, the code was instead created so that the program would play checkers matches, store data about which moves led to good and bad outcomes, and then use that dataset to make moves in the next game that had a higher probability of success. The more games played, the larger the dataset, and the program could predict with higher accuracy which moves would lead to a winning game (read more in Samuel’s own writings here). This was a very mathematical approach to the concept of learning through trial and error, which humans do constantly. In this way, artificial intelligence was measured through the ability to learn and improve artificially, or without human intervention.
By the mid-1960s, the first program was ready to attempt the Turing test. ELIZA, a program that we would now call one of the first chatbots, used a newly developed branch of AI called natural language processing (NLP) to simulate the conversation patterns of a psychotherapist (this article explains the nuance and impacts of the ELIZA program). NLP is the system of translating human speech or text into a computer script that can be processed and responded to, and while ELIZA was a very rudimentary example, it was groundbreaking for the time. And in the same decade, research by MIT was being done on a parallel branch of AI called computer vision, which translates images and video into a computer script, just as ELIZA translated text.
Different branches of AI like that of machine learning, natural language processing, and computer vision, among others, have had momentous gains, as well as large setbacks, over the last half century, as described by IBM. Despite an “AI Winter” in the 1980s where funding for research dried up, progress was still made to create more efficient, faster, accurate, and effective coding and scripts for computers to follow, allowing for the development of programs that used multiple branches of AI simultaneously to make our lives easier. And in many ways, AI has been in our lives without us really thinking about it. Spam filtering in our emails uses machine learning to recognize patterns based on historical email data, while Netflix uses the same tools to give us viewing recommendations based on our watch history. Siri has been using natural language processing to understand our verbal requests since its creation in 2011, while we’ve been using computer vision to unlock our phones with Face ID since 2017.
Which brings us to the AI boom of the last decade, and specifically, the rise of generative artificial intelligence. Whereas previously AI was consuming and interpreting data and either providing direct answers or predictions based on historical patterns, MIT News explains that new developments in the mid-2010s allowed for programs to be trained to generate entirely new data.
Combined with an increase in computing power and resources, genAI has exploded in popularity and capacity, and is now utilized for everything from writing emails to creating deepfake celebrity videos. As you can imagine, programs written by flawed humans designed to create new datasets have immense potential, but also immense risk. But that will be a topic for next time.