Danny Chadwick
In the last few months, apps using Generative AI exploded onto the market. AI photo app Lensa and OpenAI’s chatbot, ChatGPT, made a huge splash because they make high-quality text and images on demand. Now Microsoft and Google are playing catchup. But what is Generative AI, and how does it work?
What is Generative AI?
Types of Generative AI and How They Work
Applications of Generative AI
Potential Risks and Ethical Considerations
Final Thoughts: Welcome to the Future
What is Generative AI?
To put it as simply as possible: Generative AI is an AI (so-called “artificial intelligence”) that creates unique content based on a prompt from a user. For example, the prompt that you give Lensa to make those cool AI profile pics a selection of selfies. In ChatGPT’s case, a prompt could be “write a sonnet about bagels in the style of HL Mencken.” The resulting text and images are wholly unique and generated by the AI. And it’s not just text and pictures that generative AI can create. Other AI products can create uncanny voice recreations, and there are even services waiting in the wings that can make video content based on text prompts.
Generative AI combines two powerful AI technologies: machine learning and the ability to create new content. AI programmers use machine learning to build models that can recognize patterns and trends in existing data, while content generation allows for the creation of unique items like a composition or an image. When an AI has a large enough sample size to draw from (its training set) it can recreate pretty much anything it can recognize. And because the data set to train AI models like ChatGPT are so large, it can mix and match elements from multiple sources to deliver something that is both unique and recognizable as the whatever the prompt asked for.
Types of Generative AI and How They Work
Generative AI algorithms come in many forms but fall into three general categories: generative adversarial networks (GANs), variational autoencoders (VAEs), and transformer models like GPT-4. Each type of generative AI algorithm has its advantages and disadvantages.
GANs are a type of generative AI that uses two deep-learning neural networks to generate new data. The first network, called the generator, is trained to create new data that resembles existing content, while the second network, called the discriminator, is trained to distinguish between real and generated data. As programmers train their AIs, the generator learns how to produce increasingly realistic images that fool the discriminator into believing they are real. This process is known as a “minimax game” since each network tries to outsmart the other while minimizing its own mistakes.
One potential disadvantage of GANs is that they can sometimes produce unrealistic or blurry images. For example, a GAN trained to generate images of human faces might sometimes create pictures with an extra pair of eyes or a distorted facial structure. Human hands can look like a downright nightmare. However, it’s still early days for this technology, and issues like this will be ironed out in due course.
VAEs are another type of generative AI used to generate new, unique data. Unlike GANs, VAEs use a compressed representation of its input data to generate something new that resembles the original. VAEs are most often used to make images and videos, but they can also generate text. A potential limitation of VAEs is that their data may not be as varied as those generated by GANs because VAEs learn a more constrained representation of the input data. Additionally, VAEs sometimes suffer from the distorted image problems that GANs run into.
Transformer models like GPT-4 are a relatively newer iteration of generative AI that have attracted a lot of eyeballs due to their impressive performance on many natural language processing tasks. ChatGPT is the current gold star example of a transformer-based AI product. These models are based on a type of neural network architecture called a “transformer.” They’re designed to process massive data sequences, are trained on a enormous text dataset, and can make coherent and contextually relevant responses to a prompt.
The advantage of transformer models is that they can generate diverse and high-quality text. However, they can suffer from biases and inaccuracies in the training data, leading to inappropriate or erroneous outputs. Additionally, the massive amount of computational resources and data required to train and run these models can make them difficult and expensive for some applications.
Applications of Generative AI
Generative AI is already being used in a host of popular services. There is the aforementioned chatbot ChatGPT, made by OpenAI and its sister image generator DALL-E. There’s also a slew of AI image editors, including Lensa (iOS, Android), Wonder (iOS, Android), and more. Those have all been around for a while. But when ChatGPT took off, Silicon Valley decided it was time to unleash the new tech and announced one new AI product after another.
Just since the start of this year, Microsoft and Google both announced AI enchantments to their search engines. Followed soon thereafter by smaller search providers DuckDuck Go and Brave. Microsoft has added AI image generation to Bing and Edge, as well as AI components to its office suite. Even Opera is adding ChatGPT to its desktop browser. Plus, Shutterstock and Adobe have come out with AI art generators trained on the work of compensated users.
But generative AI can go far beyond image creators, chatbots, and search assistants. Professionals from all walks of life can use these tools in their work. Generative AI has potential applications in product design, enabling companies to create custom products tailored to each customer’s needs. Additionally, it can be used by healthcare professionals by aiding in diagnostic and treatment development.
Moreover, generative AI can create personalized content, such as news articles or music playlists. By analyzing a user’s preferences and behavior, generative AI algorithms can generate content tailored to their interests, increasing user engagement and satisfaction. Generative AI can help create new content for the entertainment industry, such as movie scripts or video game levels. The ability to generate unique but similar products allows companies to create more content faster and consistently higher quality.
That’s just the tip of the iceberg when it comes to the potential applications of generative AI. The technology could also find useful places in many other industries and professions. It wouldn’t be hyperbolic to compare this technology, when implemented in scale, to the invention of the printing press or the development of the assembly line in terms of how it could transform how we create and consume content and perform our jobs.
Potential Risks and Ethical Considerations
Of course, with any new technology comes the risk that it will be misused or impact some groups negatively. One of the chief concerns surrounding generative AI is that it could displace writers, artists, and other creative types who make their living making articles, art, scripts, books, and more. Another potential downside to generative AI is that it could be used to make deep fakes of celebrities and politicians that would be indistinguishable from videos and images of the real people and use them to trick the public. And, of course, there always hangs the science-fiction question of what happens if we allow AI to start making decisions without proper human oversight. Will it turn on its creators or make decisions that harm humans, thinking it will help?
The good news is that most of the ethical questions surrounding AI are perennial concerns of technological progress. Job destruction almost always accompanies advances in automation. But it also comes with more advanced tools for those who stick with the craft. Writers, artists, and other creatives now have a potent assistant to use to help them with their vocations, not necessarily destroy them. Plus, the fakery of images of celebrities and politicians has been around since the first photo editing software programs. And we’ve been prepping for the AI takeover in movies since before the first Terminator movie. And while they are valid questions and concerns, they will likely be resolved in a way that benefits everyone or at least dealt with in a way that doesn’t involve massive job loss and the coming of an AI-overlord government.
However, one huge snarl many AI products will have to overcome is copyright. Because generative AI is trained on a massive dataset of text, images, sounds, and more, copyrighted material makes up a distinctive portion of what generative AI draws from to make new creations. Granted, the nature of generative AI precludes an exact word-for-word recreation of a copyrighted work, but everything generative AI creates is made up of bits of copyrighted material. Or at least the AI learned how to write and draw based on the works of humans. This could lead to potential lawsuits from writers and artists who feel that their work was stolen to train the AI and that they deserve compensation or demand that the AI “forget” what it learned from their work.
But, it could be argued that the AI isn’t rote-copying protected work and that machine-learning is equivalent to human learning, just like if a writer read a book and got inspired to write their own along similar lines. It will probably come down to a court battle where a judge will have to decide, “what’s the difference between an AI learning from mimicking, and a human doing it?” And that’s just the tip of the iceberg when it comes to untangling all the legal implications that generative AI will surely raise. Now would be a good time for attorneys to brush up on their computer science.
Final Thoughts: Welcome to the Future
Generative AI may be as scary as it is impressive and fascinating. But it’s here now, and it’s not going away. Given the rate of adoption in the opening months of 2023, it’s not a hard prediction to make that by the end of the year, Generative AI will be incorporated into much of your daily life. And by the end of 2024, it may be difficult to remember life without this technology.
