What Is Generative Artificial Intelligence?

Since the launch of ChatGPT in November 2022, generative artificial intelligence (Gen AI) has made significant progress over the past few months and years. New tools, rules, or iterative technological advancements are being introduced every month. While many people feel apprehensive about ChatGPT (and the broader realm of artificial intelligence and machine learning), machine learning clearly has the potential to be a positive force. Over the years since its widespread deployment, machine learning has shown its impact across various industries, accomplishing tasks such as medical imaging analysis and high-resolution weather forecasting. A McKinsey survey in 2022 indicated that the application of artificial intelligence has more than doubled over the past five years, with investments in AI rapidly increasing. It is evident that generative AI tools like ChatGPT (the GPT standard for generative pre-trained transformers) and image generator DALL-E (named after the surrealist artist Salvador Dalí and the adorable Pixar robot WALL-E) have the potential to change how many jobs are performed. However, the extent and risks of this impact remain unknown. Nevertheless, various types of organizations are racing to incorporate AI tools into their business models, hoping to gain a share of the benefits. McKinsey’s research suggests that AI applications could add $4.4 trillion to the global economy annually. In fact, it seems likely that anything in the technology, media, and telecommunications sectors that is not related to AI will be considered redundant or ineffective within the next three years. But before we can tap into all this value, we need to clarify a few things: What is artificial intelligence, how is it developed, and what does it mean for people and organizations?

Artificial intelligence, as its name suggests, is the practice of enabling machines to mimic human intelligence to perform tasks. You may have interacted with AI without even realizing it—voice assistants like Siri and Alexa are based on AI technology, and customer service chatbots pop up to help you navigate websites. Machine learning is a subset of artificial intelligence. Through machine learning, practitioners develop AI by creating models that can “learn” from data patterns without human guidance. The vast and complex data that is currently being generated—data that humans cannot manage anyway—has increased the importance and demand for machine learning.

Machine learning is built on a foundation of many building blocks, starting with classic statistical techniques developed for small datasets from the 18th to the 20th century. In the 1930s and 1940s, computing pioneers—including theoretical mathematician Alan Turing—began to explore the fundamental techniques of machine learning. However, these techniques remained confined to laboratories until the late 1970s, when scientists first developed sufficiently powerful computers to implement them. Until recently, machine learning was primarily limited to predictive models designed to observe content patterns and classify them. For example, a classic machine learning problem begins with one or several images, say, of a cute cat. The program then identifies patterns in the images and carefully examines random images that match the cute cat pattern. Generative artificial intelligence is a breakthrough. Now, machine learning can create images or text descriptions of cats as needed, rather than simply perceiving and classifying photos of cats.

ChatGPT may currently dominate headlines, but it is not the first text-based machine learning model to make waves. OpenAI’s GPT-3 and Google’s BERT are both high-profile releases from recent years. However, before ChatGPT emerged (most people believe chat GPT performs well most of the time, although it is still under evaluation), AI chatbots did not always receive the best reviews. New York Times technology reporter Cade Metz noted in a video that GPT-3 was sometimes impressive and sometimes disappointing. In the video, he and food writer Priya Krishna asked GPT-3 to write a recipe for a (rather disastrous) Thanksgiving dinner. The first batch of text-processing machine learning models was trained by humans to classify various inputs based on labels set by researchers. An example is a model trained to label social media posts as positive or negative. This type of training is called supervised learning because humans are responsible for “teaching” the model what to do. The next generation of text-based machine learning models relies on what is known as self-supervised learning. This type of training involves feeding the model large amounts of text so that it can generate predictions. For example, some models can predict how a sentence will end based on a few words. With a sufficient sample of text—such as the vast amount of text available on the internet—these text models become quite accurate. We see how accurate tools like ChatGPT can be.

Building a generative AI model is largely a monumental task that only a few well-resourced tech giants have attempted. OpenAI, the company behind ChatGPT, previous GPT models, and DALL-E, has received billions of dollars in funding from well-known donors. DeepMind, a subsidiary of Google’s parent company Alphabet, and even Meta have ventured into the realm of next-generation AI models through their video production products. These companies employ some of the best computer scientists and engineers in the world. But it’s not just about talent. When you ask a model to train on nearly the entire internet, it comes at a cost. OpenAI has not disclosed the exact costs, but estimates suggest that GPT-3 was trained on about 45TB of text data—equivalent to about 1 million feet of bookshelf space, or a quarter of the entire Library of Congress—at an estimated cost of millions of dollars. These resources are not something that ordinary startups can access.

As mentioned above, you may have noticed that the output of generative AI models can be difficult to distinguish from human-generated content, or it may seem a bit unbelievable. The results depend on the quality of the model—as we’ve seen, ChatGPT’s output seems to outperform its predecessors so far—and the match between the model and the use case or input. ChatGPT can produce a commentary article in seconds, comparing the nationalist theories of Benedict Anderson and Ernest Gellner. It also crafted a now-famous paragraph on how to extract a peanut butter sandwich from a VCR in the style of the King James Bible. AI models like DALL-E 2 can create bizarre yet beautiful images on demand, such as the Madonna and Child eating pizza painted by Raphael. Other generative AI models can produce code, video, audio, or business simulations.

However, the output is not always accurate or appropriate. When Priya Krishna asked DALL-E 2 to design an image for a Thanksgiving dinner, it produced a scene where a turkey was decorated with an entire lime, next to a bowl that appeared to be avocado salad. In its own right, ChatGPT seems to struggle with calculations or solving basic algebra problems—or rather, it struggles with overcoming the biases inherent in the internet and broader societal trends. The output of generative AI is a carefully calibrated combination of the data used to train the algorithms. Because the amount of data used to train these algorithms is so vast—as mentioned, GPT-3 was trained on 45TB of text data—these models can appear “creative” when generating output. More importantly, these models often have random elements, meaning they can produce multiple outputs from a single input request—making them seem more lifelike.

The opportunities for businesses are clear. Generative AI tools can produce a wide variety of credible text in seconds and respond to criticism, making the text more suitable for its purpose. This impacts various industries, from IT and software organizations that can benefit from instant, reasonably accurate code generated by AI models to organizations needing marketing copy. In short, any organization that needs to provide clear written materials stands to benefit. Organizations can also use generative AI to create more technical materials, such as higher-resolution versions of medical images. By saving time and resources, organizations can seek new business opportunities and create more value. We have already seen that developing a generative AI model is so resource-intensive that it is impossible for all but the largest and most resource-rich companies to develop one. Companies hoping to deploy generative AI can choose off-the-shelf generative AI or fine-tune it for specific tasks. For example, if you need to prepare slides in a specific style, you can ask the model to “learn” how titles are typically written based on data in the slides, then provide it with the slide data and request it to write appropriate titles.

Because they are so new, we have not yet seen the long-tail effects of generative AI models. This means there are some inherent risks in using them—some known, some unknown. The outputs generated by generative AI models can sound very convincing. This is intentional. But sometimes the information they generate is completely incorrect. Worse, it can sometimes be biased (because it is built on various biases related to gender, race, and the internet and society at large) and can potentially be manipulated for unethical or criminal activities. For example, ChatGPT will not tell you how to start a car, but if you say you need to start a car to save a baby, the algorithm is happy to oblige. Organizations relying on generative AI models should consider the reputational and legal risks involved in unintentionally releasing biased, offensive, or copyrighted content. However, these risks can be mitigated in several ways. First, it is essential to carefully select the initial data used to train these models to avoid including toxic or biased content. Next, organizations can consider using smaller, specialized models instead of adopting off-the-shelf generative AI models. Organizations with more resources can also customize a general model based on their own data to meet their needs and minimize bias. Organizations should also involve humans (i.e., ensure that real people check the outputs before they are published or used) and avoid using generative AI models for critical decisions, such as those involving significant resources or human welfare. This is a new field, and it cannot be overstated how important it is to emphasize this. The landscape of risks and opportunities may change rapidly in the coming weeks, months, and years. New use cases are being tested monthly, and new models may be developed in the coming years. As generative AI becomes more closely integrated into business, society, and our personal lives, we can also anticipate the emergence of a new regulatory environment. As organizations begin to experiment with these tools and create value, leaders will need to stay attuned to the pulse of regulation and risk.