Generative artificial intelligence

Generative artificial intelligence is a type of artificial intelligence (AI) technology that can make content such as audio, images, text, and videos. It involves algorithms such as ChatGPT, a chatbot that can produce essays, poetry, and other content requested by a user, and DALL-E, which generates art. AI emerged gradually over more than half a century. Generative AI is a type of machine learning, which involves using data and algorithms to imitate how humans learn and become more accurate. While machine learning can perceive and sort data, generative AI can take the next step and create something based on the information it has. However, it remains expensive, and only a few well-financed companies, including OpenAI, DeepMind, and Meta, have built generative AI models.

Background

British computer scientist, philosopher, and polymath Alan Turing proposed a question in his 1950 paper “Computing Machinery and Intelligence” that appeared in the journal Mind. He wondered, “Can machines think?” Turing suggested a test that he called the Imitation Game but in modern times is called the Turing Test. A subject posts questions to a human and a machine (computer), not knowing which is answering. If the person cannot tell which answers are from the person or computer, the machine passes the test and can be said to think. In the paper, Turing predicted that most of the time humans would be unable to tell machines and humans apart in this respect by the twenty-first century. However, when Turing wrote his paper, computers had several limitations. Notably, they could execute commands but could not store them. Computers were also prohibitively expensive and only government facilities, large technology companies, and major universities had access to them.

The first AI program, Logic Theorist, was created in 1956. It was written by Allen Newell, Cliff Shaw, and Herbert A. Simon to perform automated reasoning, specifically to prove theorems from the three volumes of Principia Mathematica (1910–1913) by Bertrand Russell and Alfred North Whitehead. Research and Development (RAND) Corporation provided the funding for the development of Logic Theorist, which was presented in 1956 at the Dartmouth Summer Research Project on Artificial Intelligence (DSRPAI). One of the hosts and organizers, John McCarthy, coined the term artificial intelligence at this conference. This inspired many in the field of computer science to focus on improving computers and programming. Over the next two decades, computers were developed to have better memory and speed and became more affordable, while programmers improved functionality. For example, Newell and Simon developed the General Problem Solver while others improved computers’ ability to interpret spoken language (natural language processing). The Defense Advanced Research Projects Agency (DARPA), a US Department of Defense agency, was sufficiently impressed to fund AI research. However, computers were still not powerful enough to store and process information quickly enough to approach AI, and funding dried up in the 1970s.

Development of AI and other computer uses was largely following Moore’s Law, which rather than a law was an observation engineer Gordon Moore expressed in 1965. Moore noted that the number of transistors in an integrated circuit or microchip had doubled roughly every two years and predicted that this would continue for at least another decade. Although the doubling went on for decades, through the latter half of the twentieth century, computer scientists were repeatedly limited by the storage and processing speed of the computers then available.

In the 1980s several developments increased computing potential. John Hopfield and David Rumelhart promoted deep learning techniques that allowed computers to learn from experience. Expert systems introduced by Edward Feigenbaum followed a human-like decision-making process. Japan’s government heavily invested in computer processing in the 1980s, pouring $400 million into its Fifth Generation Computer Project (FGCP). While these efforts moved the process forward, they had little practical effect. However, they inspired young engineers and scientists.

The 1990s and 2000s saw progress in AI development as the memory and speed of computers finally developed to more than meet most needs. IBM’s Deep Blue computer program defeated grand master Gary Kasparov in 1997, raising the profile of AI and its potential. Windows implemented Dragon Systems’ speech recognition software that year as well.

In the twenty-first century, AI became commonplace. Algorithms helped companies such as Netflix suggest movies and shows that customers might enjoy or items that Amazon shoppers might wish to purchase. Self-driving cars arrived on the market. Companies, organizations, and individuals used data mining in myriad ways, such as marketing and making political decisions.

AI has several subfields, including deep learning, machine learning, and neural networks. Deep learning is a subfield of neural networks, which is a subfield of machine learning. Machine learning is a neural network with at least three layers. Neural networks try to function in the way that the human brain does so they can learn from available data. Additional layers help to increase accuracy. Deep learning is commonly used to improve automation, such as by assessing processes to increase efficiency. This technology is used by digital assistants, self-driving vehicles, and voice-enabled devices such as television remotes. Deep learning can use unstructured data such as images and texts. For example, if tasked with organizing images of vehicles, a deep learning algorithm would decide what features are most important to consider, such as size. Machine learning uses structured data or data that it can organize into a format that it can use to make predictions. In other words, a person would decide what features are most important and create a hierarchy for the machine learning algorithm to use. A deep learning algorithm evaluates its accuracy and, if presented with a new photo, can more precisely evaluate the subject.

Machine learning and deep learning models can learn in different ways. The types of learning are supervised, unsupervised, and reinforcement. Supervised learning uses datasets that a person has labeled, and the model’s education is guided by a human. Unsupervised learning, or self-supervised learning, finds patterns in the data and groups the data accordingly. Reinforcement learning uses feedback to refine its actions and become more accurate.

The layers of deep neural networks are interconnected. They use forward propagation, which means that each layer uses the computations of the previous layer to increase accuracy. The input and output layers are called visible layers. Data is fed into the input layer while results are made in the output layer. Backpropagation is the process of going back through layers to train the model based on algorithms that calculate prediction errors. The combination of forward propagation and backpropagation can refine predictions and gradually increase accuracy.

Overview

While generative AI has been in use for some time, it only emerged in the public consciousness in the early 2020s because of transformers and the language models they make possible. Transformers, a type of machine learning, allowed researchers to train enormous models using unstructured data such as billions of texts. Transformers also permit attention, or the ability of models to create multiple links, such as when book links are made between the words in a sentence, throughout the chapter and the book, and with other books. Thanks to large language models (LLMs), which may have trillions of parameters, generative AI models can create realistic images and videos, unique text, and other content.

ChatGPT is possibly the best-known generative AI product, in part because it was made available to the public to use. Its developer, OpenAI, released the free chatbot in November 2022. It uses public data to generate responses to prompts. People have used it to write poetry and other amusements and for work, such as writing real estate listings and creating lesson plans. Subscribers to the paid version, ChatGPT Plus, can use a range of plug-ins such as Expedia to book trips and OpenTable to make restaurant reservations. The Code Interpreter plug-in uses Python to generate responses to prompts. For example, users have provided data sets to the chatbot and requested analysis to find trends. ChatGPT can generate graphs based on datasets such as stock price fluctuations and clean up data. For example, it can delete empty rows in a spreadsheet. Code Interpreter can generate videos and GIFs from images and images from videos and GIFs. These and other uses have prompted people in some professions, such as journalism and data analytics, to fear they will be replaced by AI.

While ChatGPT has been instrumental in popularizing generative AI, it has also demonstrated some drawbacks of the technology. For example, the chatbot has produced responses that are false, such as fabricated court cases in a legal filing. It has encouraged students to cheat. And individuals have used generative AI to create deepfakes, fake audio and video clips or images, to fool people. For example, deepfakes appear to show celebrities and politicians saying or doing things that they have not. After Russia invaded Ukraine in 2022, deepfakes appearing to show Ukraine’s president announcing his country’s surrender circulated online. Experts are concerned that deepfakes will be used to swing elections or cause stock market swings. In April 2023 the Republican National Committee (RNC) released an AI-generated ad, and Republican presidential candidate Ron DeSantis’s campaign released an ad containing AI-generated images of Donald Trump hugging Dr. Anthony Fauci, an infectious disease expert many on the right despise. In response, the Federal Election Commission (FEC) was working on possible regulations regarding the use of deepfakes in political ads.

Bibliography

Anyoha, Rockwell. “The History of Artificial Intelligence.” Harvard University Graduate School of Arts and Sciences, 28 Aug. 2017, sitn.hms.harvard.edu/flash/2017/history-artificial-intelligence/. Accessed 21 Aug. 2023.

“The Imitation Game: A Rare Alan Turing Article at CMU Libraries.” Carnegie Mellon University, July 2020, www.library.cmu.edu/about/news/2020-07/imitation-game-rare-alan-turing-article-cmu-libraries. Accessed 21 Aug. 2023.

Jesuthasan, Ravin. “Cutting Through the Hype Cycle of Generative AI.” Forbes, 19 Aug. 2023, www.forbes.com/sites/ravinjesuthasan/2023/08/19/cutting-through-the-hype-cycle-of-generative-ai/?sh=7fccaa762755. Accessed 21 Aug. 2023.

Lawton, George. “What Is Generative AI? Everything You Need to Know.” TechTarget, July 2023, www.techtarget.com/searchenterpriseai/definition/generative-AI. Accessed 21 Aug. 2023.

Martineau, Kim. “What Is Generative AI?” IBM, 20 Apr. 2023, research.ibm.com/blog/what-is-generative-AI?. Accessed 21 Aug. 2023.

Sundar, Sindhu, and Aaron Mok. “What Is ChatGPT? Here’s Everything You Need to Know about ChatGPT, the Chatbot Everyone’s Still Talking about.” Business Insider, 21 Aug. 2023, www.businessinsider.com/everything-you-need-to-know-about-chat-gpt-2023-1. Accessed 21 Aug. 2023.

Swenson, Ali. “FEC Moves Toward Potentially Regulating AI Deepfakes in Campaign Ads.” PBS, 10 Aug. 2023, www.pbs.org/newshour/politics/fec-moves-toward-potentially-regulating-ai-deepfakes-in-campaign-ads. Accessed 21 Aug. 2023.

Turing, A.M. “Computing Machinery and Intelligence.” Mind, vol. 49, 1950, pp. 433–460.

“What Is Generative AI?” McKinsey & Company, 19 Jan. 2023, www.mckinsey.com/featured-insights/mckinsey-explainers/what-is-generative-ai. Accessed 21 Aug. 2023.