By IT Brew Staff
less than 3 min read
Definition:
While generative AI, or GenAI, is often used interchangeably with AI, the two are different. For starters, “AI” encompasses a number of artificial-intelligence disciplines, including predictive AI. GenAI depends on massive datasets of text, images, and more. Neural networks, which are machine-learning models designed to recognize patterns, are “trained” on that data to produce new, contextually relevant content.
Some common uses of GenAI include text, video, audio, software code, and images. For specialized uses such as drug discovery or cybersecurity, a properly trained GenAI app can also generate synthetic data, which is new data that mimics the properties and structures of existing data (such as a molecular structure or a list of sensitive personal information).
Types of generative AI include:
- Transformers: When most people say “generative AI,” they’re often referring to technology based on transformers, such as the large language models (LLMs) underlying ChatGPT and other popular chatbots. Transformers rely on attention mechanisms to train a model on what to focus on; they’re effective in natural language processing as well as outputting text into a variety of formats, including code or spreadsheets.
- Generative adversarial networks: A generator network generates fake data, which a “discriminator” then evaluates against verified data, with the aim of creating data indistinguishable from the “real thing.” GAN is used in applications such as deepfaked videos.
- Variational autoencoders: This involves two neural networks. The first compresses raw training data into parameters, while the other reconstructs that data from its compressed form. It’s often used in data compression and enhancing images.
While GenAI can benefit teams that need to produce text, images, code, and video quickly, the technology has some well-documented drawbacks, including the risk of hallucinations (i.e., nonsensical and factually incorrect outputs), biased outputs due to homogeneous training data, and an inability to see how it arrived at particular conclusions.
Numerous lawsuits alleging intellectual property violations have been launched against GenAI producers; for example, Anthropic, which created the popular Claude chatbot, had to pay $1.5 billion in a class action settlement filed by authors who said the company misused their books and intellectual property to train its AI.
In addition, GenAI may produce code seeded with vulnerabilities, and some of its products—such as deepfakes—are often used in cyberattacks. Protecting tech stacks against generative AI-powered attacks has become a key focus of cybersecurity pros.
