What is generative AI? Your questions answered

Generative AI seems to have popped up everywhere in the mainstream lately—primarily via the popularity of ChatGPT, but also through the proliferation of text-to-image tools and AI avatars in our social media feeds. But beyond fun smartphone apps and ways for students to shirk essay-writing assignments, what can generative AI do? How does it work? How will it AI technology affect the way that businesses operate? And possibly most heavy on your mind: will generative AI models take your job?

As Ainsley Harris has written in Fast Company, in a profile of Mira Murati, the CTO of ChatGPT-developer OpenAI, “Nearly every aspect of economic activity could be affected by OpenAI’s tools. According to the McKinsey Global Institute, relatively narrow applications of AI, such as customer service automation, will be more valuable to the global economy this decade than the steam engine was in the late 1800s. [Artificial general intelligence] will be worth many, many trillions of dollars more, or so say its most ardent believers.”

What is generative AI?

Generative AI is artificial intelligence that can generate novel content, rather than simply analyzing or acting on existing data.

AI (short for artificial intelligence) broadly refers to the idea of computers that can learn and make decisions in a human-like way. Machine learning is a subfield of artificial intelligence focused on teaching computers to recognize patterns through data and algorithms. It differs from traditional programming in that the computer doesn’t need to be explicitly coded to address every potential scenario. Natural language processing (or NLP) refers to the specific ability of certain machine learning models to understand written or spoken language (as opposed to numerical data or specific, previously coded language).

Generative AI models take this pattern recognition a step further and use it to create text, images, computer code, and other novel content.

How does generative AI work?

If you’re using a generative AI model, you enter a prompt describing the output you’d like, and the program gives it to you, whether it’s in the form of text, code, images, or—increasingly—sound and video. But behind the scenes, things are a bit more complicated.

When programming a generative AI model, developers will feed it massive amounts of training data. For example, consider a large language model (or LLM). LLMs are the kind of generative AI used in ChatGPT and the basis of many generative AI tools in use today. These machine learning models utilize deep learning algorithms to process and understand language. LLMs are trained with immense amounts of previously written text to learn language patterns so they can perform tasks. Those tasks can range from translating content to responding in chatbot conversations—basically anything that requires language analysis of some sort. 

Generative AI models are comprised of multiple layers of “neural networks,” which are machine learning models that mimic the neurons in the human brain, using a network of nodes to process data through algorithms. This allows the generative AI to make connections between lots of different data points and learn which ones are the most important when responding to query or prompt.

Generative AI also makes use of deep learning, which is a type neural network whose data passes through several layers of processing—some of which are hidden from the programmer—before arriving at a response.

For LLMs, these neural networks work together to analyze text and predict outputs. They’re also trained with a left-to-right or bidirectional transformer, which works to maximize the probability of following and preceding words in context—just like a human could reasonably predict what might come next in a sentence. LLMs also have an attention mechanism that allows them to focus selectively on parts of text in order to identify the most relevant sections for summaries, for example. 

What are the applications of generative AI?

Even it’s early stages, generative AI can already do a lot and the tools available are incredibly diverse. Generative AI models can take in such content as images, longer text formats, emails, social media content, voice recordings, program code, and structured data. Then the AI model can output new content, translations, answers to questions, sentiment analysis, summaries, and even videos.

These universal content machines have many potential applications in business. Today the most common uses of generative AI are for marketing. In the future, there is potential for generative AI to impact healthcare and life sciences—to make diagnoses, for example, or find new cures for disease.

Not surprisingly, many of the early stages of generative AI began with large tech, or digital native, companies. Over the next several years, we see generative AI permeating traditional industries, like manufacturing, healthcare, and pharmaceuticals, for example. Once a generative AI model has been trained, it can be fine-tuned for specific content domains with much less data.

We are now starting to see specialized generative models for biomedical content, legal documents, and translated text, which will give rise to additional use cases in those industries and domains. They may help organizations to manage their knowledge and content more effectively so that it can be easily accessed by employees and customers.

What is ChatGPT?

The current wave of generative AI hype comes from the popularity of ChatGPT, a conversational chatbot created by OpenAI using an LLM that emphasizes back-and-forth dialog. GPT is short for “Generative Pre-Trained Transformer,” but the name itself requires some unpacking:

• “Generative” refers to its ability to generate text.
• “Pre-Training” means using the model from one machine learning task to train another, similar to how humans draw on existing knowledge when learning new things. In this case, GPT involves using a large corpus of text as training data.
• A “Transformer” is a kind of neural network that holistically learns about the relationship between all parts of a data sequence (in this case, the words in a sentence). It’s seen as a breakthrough for AI and generative AI technologies because it understands context and nuance better than previous approaches.

Basically, users can ask ChatGPT questions or they can ask it to generate text for them that fits certain parameters. ChatGPT then draws on all the data it was trained on in order to generate a response.

But ChatGPT—and generative AI tools in general—cannot come up with original ideas instead it uses its pattern recognition on from its training data and applies those patterns to the user prompts.

Image generation

Even though many current AI tools are based on large language models and respond to text input, generative AI can be used to create images. Diffusion models are one method for image generation from text prompts. These generative AI algorithms work by adding random noise to a set of training images, then learning how to remove noise to construct the desired image.

OpenAI’s text-to-image tool, DALL-E, uses GPT-3 (the same AI text model that underpins ChatGPT) to interpret users’ requests.

Generative AI assistants

We can all use a little help sometimes—that’s why the rich and famous have armies of personal assistants to help them out. We can’t all afford that, but artificial intelligence is coming to the rescue with a new wave of AI assistants to help get get things done, handling the tedious details of making things happen.

It’s a huge growth area, with billions of dollars invested in hundreds of companies creating generative AI models for specific tasks, such as AI marketing assistants, AI recruiting, HR, or even sorting out expenses. Others can handle tasks like taking notes in meetings, negotiating bills, or even doing math homework.

Chances are that whatever someone wants help with, there is, or will soon be, a generative AI assistant available to help.

Popular generative AI apps

In addition to ChatGPT, generative AI apps and tools have been popping up so fast it’s hard to keep track. From AI tools that can create images for use on social media platforms to video enhancers that are capable of deleting objects and creating virtual avatars, we’ve found some of the best generative AI tools for enhancing productivity, improving writing, and generating multimedia content, including sounds and videos.

Here are just a few of the most popular:

Generative AI text tools

• AISEO.ai: If you have content that’s already written, it makes sense to use it as widely as possible. AISEO’s content repurposer lets you take something like a blog post and rewrite it into a Twitter campaign, an email, or even a script for a YouTube video. AISEO also offers a neat tool called the Paraphraser, which uses generative AI to tweak your written content, including shortening, extending, and even changing the tone of it for different audiences. 
• Grammarly: This is an invaluable tool for anyone who writes. At the simplest level, it is a kick-ass grammar and spelling checker far superior to the ones built into Word or Google Docs. But that’s just the beginning: It can also advise on tone, readability, and engagement.
• Jasper.ai: As the name suggests, Jasper uses AI to provide inspiration. Feed the system a few prompts, and it can crank out anything from Instagram captions to essays on the meaning of life, providing multiple options.

Generative AI image tools

• Stable Diffusion: An open-source text-to-image application created by Stability AI. The official version has a laborious installation process and runs through a command line, but third-party developers have used the open source code to create more accessible versions for desktop computers and the web.
• Imagen: Another image generation tool that uses a diffusion model, this one created by Google. The company has chosen not to release its code or demonstrate it publicly for now, citing its potential to create inappropriate content.
• Dreambooth: A deep learning model, developed by Google, that can fine-tune images created through diffusion. Its most notable use case is the ability to generate new pictures of specific people based on existing photos—for better or worse.
• Lensa: An image editing app for iOS and Android from Prisma Labs that first launched in 2018.
• Nightcafe: An online AI image generator with a twist: It allows the user to use several different AI image generators, including Stable Diffusion, DALL-E, and others. Users can also apply a range of styles, so the experience is easy to get started with.
• Image.AI: AI image generators seldom produce good results the first time around. Instead, it is a process of evolution, refining the things that work and removing those that don’t. Image.ai does this, creating image after image until the user accepts the result.
• Photoleap: If you want to see how your house might look on the set of The Last of Us, Photoleap can help. The AI scenes feature of this iOS-only app uses generative AI to turn an image into a post-apocalyptic, fantasy cartoon or any one many other scene types.

Generative AI video tools

• Nova AI: Upload a video to Nova AI, and it will automatically caption, categorize and tag it. Nova AL struggles with music, though: when uploading a Leonard Cohen live video to it, and it subtitled “Let’s not talk of love or change” as “Let’s smart things we can’t.” Not quite what Leonard meant, but it has a certain poetry of its own.  
• Make-A-Video: Need a video of a confused bear in a calculus class? Meta’s new Make-A-Video tool can create a video from pretty much any text thrown at it.
• Synesthesia.io: This AI video generator for pitches creates pitch videos from text in a few minutes using a computer-generated avatar.

Generative AI sound tools

• Soundraw: This generative AI tool can create multiple pieces of incidental music in seconds. Pick a mood, a genre and a theme and the AI system creates 15 short music clips to choose from. 
• Murf: Speech generation was one of the first applications of AI, and the latest evolution is Murf, which creates AI voiceovers. This sophisticated service can create very natural-sounding speech from text in 120 voices and 20 languages.
• NaturalReader: Another voiceover service, but with a twist: The app snaps a smartphone photo and extracts the text and read it out loud in any one of the several hundred voices on offer. Their Personal Reader app can also read eBooks, PDFs, screenshots, and other texts.
• Cleanvoice.ai: Record yourself speaking and you will be horrified at the background noises, umms, ahhs and lip-smacking. Cleanvoice.ai cuts these out of your audio, leaving only the important stuff.

What are the drawbacks of generative AI?

Because the end user can’t see what ChatGPT and similar generative AI models are doing behind the scenes, it’s impossible to know the “source” of any given information they generate. ChatGPT has been known to generate false or offensive information, and even if you ask it to list its sources, the AI tool may invent a source whole-cloth.

Generative AI models also can’t discern bias in a source, leading to output that may not be objective.

There are some potential legal and ethical concerns related to generative AI and content creation as well. One is the ability to easily create “deepfakes”—images or video created by AI that appear realistic but are false or misleading.

Additionally, generative AI raises questions about what is original and proprietary content and may have a significant impact on content ownership. For example, if you use generative AI for creating content, is that content protected by copyright law? Or on the flip side, generative AI relies on existing content to use as training data, but what recourse does an author or artist have when their work is used as raw data to train an AI model?

Generative AI models can also be incredibly expensive to train. A 2020 study estimated that the cost of training a LLM with 1.5 billion parameters can be as high as $1.6 million.

Will generative AI replace humans in the workplace?

In a word: No. Or at least not yet.

Generative AI is a tool that can provide new ideas and help refine existing practices, but it can’t replace human intelligence in the form of the person who comes up with ideas and concepts. Instead, AI models can make this person more efficient and help them make more of the resources they have. 

AI assistants, for example, are excellent at specific tasks and definitely make work easier, but they are not adaptable: You can’t throw a new task at generative AI and leave it to figure it out like you can with a real person. That will probably take a few more generations. But for now, AI assistants can definitely make life easier by automating everyday tasks and leaving humans to tackle new, unique, or creative challenges. 

How can (and will) businesses use generative AI?

Babson College Professor Thomas Davenport and Nitin Mittal, head of U.S. artificial intelligence growth at Deloitte, are the authors of All In on AI: How Smart Companies Win Big With Artificial Intelligence. Their book examines how companies including Alphabet, Ping An, Airbus, Walmart, and Capital One leverage AI in business strategy, key processes, change management, and competition.

Here, Davenport and Mittal provide Fast Company with an excerpt from their book offering an overview on AI archetypes, capabilities, and general principles.

The general path to using AI (excerpt from All in on AI)

Reprinted by permission of Harvard Business Review Press. Excerpted from All in on AI: How Smart Companies Win Big With Artificial Intelligence by Thomas H. Davenport and Nitin Mittal. Copyright 2023 Deloitte Development LLC. All rights reserved.

The path to becoming all-in on AI is not particularly well trodden; we’ve estimated that fewer than 1% of large organizations would meet our definition of the term. However, there are capability maturity models for virtually every business capability, and we will describe a similar approach for AI. Advancing maturity in AI is based on a variety of factors, including:

• Breadth of AI use cases across the enterprise
• Breadth of different AI technologies employed
• Level of engagement by senior leaders
• The role of data in enterprise decision making
• Extent of AI resources available—data, people, technology
• Extent of production deployments, as opposed to AI pilots or experiments
• Links to transformation of business strategy or business models
• Policies and processes to ensure ethical use of AI

Capability maturity models tend to have five levels, and we see no reason to depart from that standard. They also tend to have low capabilities at Level 1 and high ones at Level 5, and we follow that pattern as well.

• AI Fueled (Level 5). All or most of the components we’ve described above, fully implemented and functioning—the business is built on AI capabilities and is becoming a learning machine;
• Transformers (Level 4). Not yet AI fueled but relatively far along in the journey with some of the attributes in place; multiple AI deployments that are creating substantial value for the organization;
• Pathseekers (Level 3). Already started on the journey and making progress, but at an early stage—some deployed systems, and a few measurable positive outcomes achieved;
• Starters (Level 2). Experimenting with AI—these companies have a plan but need to do a lot more to progress; they have very few or no production deployments;
• Underachievers (Level 1). Started experimenting with AI but have no production deployments and have achieved little to no economic value.

We might also add a “Level 0” to describe companies that have no AI activity whatsoever, but this is certainly a minority category among large firms in sophisticated economies.

The key difference with other maturity models is that we’re offering three alternative archetypes for the use of AI, but a company can be at various levels no matter what the primary focus of their efforts.

We would argue that in talking about AI-fueled enterprises, we are almost always describing Level 5 organizations. Like our examples, they are companies that have a wide variety of AI technologies and use cases in place, along with specialized technology platforms to support them. They do experiment, and companies striving to create may do more experimentation than those seeking operational improvements.

The goal of all these organizations, however—usually achieved—is to actually do business with AI by putting AI systems into production deployment. New business processes are employed. New products and services are introduced to the marketplace and used by customers. Senior executives are engaged and active in identifying use cases and monitoring performance. They have established data science groups, modernized their digital infrastructures, and identified large volumes of data for training and testing models.

Perhaps most importantly, there are alternative archetypes for employing AI, and somewhat different versions of capability models for different strategies. As we noted earlier, our view is that the three major archetypes can be summarized as:

1. Creating new businesses, products, or services
2. Transforming operations
3. Influencing customer behavior

While operational improvements are the most common objective for AI according to our survey research, it’s clear that at least some companies don’t just use AI to make their existing strategies, operations, and business models somewhat more efficient. Instead, they use it to enable new strategies, radically new business process designs, and new relationships with customers and partners. Those companies would assess their capabilities in terms of the degree to which they have successfully developed new strategies, business models, or products.

Operationally focused AI objectives would involve achievement of substantial operational  improvements, and customer behavior objectives would focus on how much actual customer behavior change has actually been achieved. Of course, that level of business transformation requires the active engagement and participation in strategic deliberations by senior management that Level 5 organizations typically display.

Thomas H. Davenport is the President’s Distinguished Professor of Information Technology & Management at Babson College, a visiting professor at Oxford’s Saïd Business School, a research fellow at the MIT Initiative on the Digital Economy, and a senior adviser to Deloitte’s Analytics practice. His bestselling books include Competing on Analytics and Big Data at Work.

Nitin Mittal is a principal with Deloitte Consulting LLP. He currently serves as the US Artificial Intelligence (AI) Strategic Growth Offering Consulting Leader and the Global Strategy, Analytics and Mergers and Acquisitions Practice leader.

Jared Newman, Laya Neelakadan, and Richard Baguely also contributed writing, reporting, and/or advice to this article.