A new era of artificial intelligence is upon us. It comes in the form of generative AI, a subset of AI that can potentially change the way we live, work and make critical business decisions.
Since the days of the first chatbot more than six decades ago, generative AI has been flying under the radar -- that is, until ChatGPT made its debut in November 2022. ChatGPT is the user-friendly interface to the generative pre-trained transformer (GPT), a type of machine learning model. Users worldwide discovered that ChatGPT could produce an array of content and answer seemingly any question. The answer is sometimes accurate and sometimes not but always in believable, sophisticated language.
Welcome to the world of generative AI, technology that can generate poems, product descriptions and all manner of text in a matter of seconds. It can also produce images, audio and video. These rapidly evolving capabilities have opened opportunities for businesses in text-to-image generation, personalized content creation and code generation. Among the beneficiaries are data scientists, application developers, marketers, sales teams, digital artists, designers and the media.
On the flip side, generative AI has heightened risks of potential copyright infringements, data privacy violations, discrimination, deepfakes, phony messaging, deceptive practices and cyberattacks.
Generative AI's history in brief
We're still in the early days of using generative AI, but this innovative technology didn't develop overnight. One of the earliest examples of generative AI was Eliza, the chatbot created by Joseph Weizenbaum in the 1960s. These early chatbots were limited by vocabulary, lack of context and an overreliance on patterns, but they continued to evolve.
In 2010, thanks to advances in artificial neural networks and deep learning techniques, the technology could automatically learn to parse text, classify image elements and transcribe audio. In 2014, Ian Goodfellow introduced the generative adversarial network (GAN), which could generate realistic images of people, voices, music and text. This inspired interest in how generative AI could be used to create realistic deepfakes that impersonate people's voices, as well as people in videos.
In 2017, Google reported on a new type of neural network architecture that brought significant improvements in efficiency and accuracy to tasks like natural language processing. Researchers showed how a transformer neural network translated between English and French with more accuracy and in just a quarter of the training time compared to other neural networks.
The breakthrough transformer architecture has evolved rapidly since it was introduced, giving rise to better pre-training techniques, like Google's Bidirectional Encoder Representations from Transformers, or BERT, in 2018, and large language models, like GPT-3, in the following years.
How generative AI works
It helps to understand the difference between generative AI and traditional AI. Traditional AI can identify patterns, make decisions, analyze and classify data, and detect fraud. Generative AI can create entirely new content. It can produce chat responses, images, diagrams, synthetic data and deepfakes.
Generative AI starts with a prompt that could be in the form of text, image, video, musical notes or any other input that the AI system can process. AI algorithms then return new content in response to the prompt. That content can include essays, solutions to problems, or realistic fakes created from pictures or audio of a person.
Early versions of generative AI required submitting data through an API or a more complicated process, but pioneers in generative AI are now developing better user experiences that let you make a request in plain language. After the first AI response to the request, you can customize the results by providing feedback about the style, tone and other elements that you want the generated content to reflect.
Popular generative AI interfaces
You may have heard about three of the more popular interfaces to large language models: ChatGPT, Dall-E and Bard.
ChatGPT is an AI-powered chatbot application built on OpenAI's GPT-3.5 language model. OpenAI has provided a way to interact with and fine-tune text responses using a chat interface. ChatGPT incorporates the history of its interactions with users into an immediate response, simulating a real conversation. Earlier versions of GPT were only accessible with an API. That changed with the advent of GPT-3, followed by a more souped-up GPT-4.
Dall-E is a multimodal form of GPT-3 that can generate images from text prompts. It identifies connections across different media, such as visuals, text and audio. It was built on OpenAI's GPT language model in 2021. A more capable version called Dall-E 2 was released in 2022 that lets users generate imagery in multiple styles.
Google, the inventor of transformers, open-sourced some of these language models for researchers but never released a public UI for them. However, when Microsoft announced a significant new investment in OpenAI and integrated a version of GPT into its Bing search engine, Google rushed out a public-facing chatbot called Google Bard.
Business applications and benefits of generative AI
Generative AI is becoming more accessible to enterprises, thanks to emerging innovations, like GPT, that can be tuned for different business applications. Specific industries can directly benefit from generative AI capabilities:
- Businesses can implement generative AI to improve customer service and technical support, as well as produce email responses to customer inquiries.
- Manufacturers can combine data from cameras, X-rays and other metrics to identify defective parts and root causes of problems more accurately and economically.
- Financial services can use AI to improve operations and build better fraud detection systems.
- Film and media companies can produce content more economically and translate it into other languages using an actor's own voice or compose music in a specific style or tone.
- Law firms can design and interpret contracts, analyze evidence and suggest legal arguments.
- Pharmaceutical companies can identify promising candidates for new drugs.
- Architectural firms can design and adapt prototypes of buildings more quickly.
- Gaming companies can design levels of content faster and more efficiently.
Generative AI's limitations
The convincing realism of generative AI content and its lack of transparency can make detection difficult. AI-generated outcomes can contain inaccuracies, plagiarize content, amplify biases, lack proper sourcing, violate privacy laws, infringe on intellectual property rights, disrupt existing business models, generate fake news, and invite cyberattacks and other malicious activities.
Simply put: If you don't know how the AI arrives at a conclusion, you can't completely trust the outcome. And relying on that outcome can be dangerous. Therefore, when introducing generative AI into workflows, it's important to follow some best practices to ensure accuracy, transparency, ease of use and success:
- Clearly label all generative AI content affecting employees and customers.
- Vet the credibility of generated content by identifying the primary sources of information.
- Look for ways to mitigate bias that can be introduced with AI applications.
- Double-check the quality of AI-generated code and content by using dedicated software.
- Learn the strengths and limitations of every tool you're planning to use.
That said, the depth, ease and speed of generative AI suggest that the advantages could outweigh the pitfalls of this emerging technology. The trials and tribulations of early rollouts have already inspired research into better tools for detecting AI-generated text, images and video.
Advances in AI platforms will help improve generative AI capabilities in business applications. But the most significant impact will come from embedding these capabilities directly into the AI tools already in use. Yet, the bottom line for the successful use of generative AI is whether the results can be trusted in real-world applications. More trustworthy outcomes require tools and procedures that are better at tracking the source and credibility of data that's fed into AI systems.