HOW GENERATIVE AI WORKS: WHAT CREATIVE PROFESSIONALS NEED TO UNDERSTAND
Doug Woodham | September 2026
Generative AI does not think or create as humans do. Yet statistical prediction allows it to produce writing, images, music and ideas that can compete with human work. Understanding how that happens is essential to judging which parts of creative work AI can augment, and which it may replace.
Generative AI has produced a curious divide among creative professionals. Some condemn it and take pride in refusing to use it. Others have quietly incorporated it into their work but hesitate to say so publicly. Strong opinions are plentiful. Understanding of the technology is not.
At its simplest, generative AI consists of systems trained on vast bodies of existing material that can produce new text, images, music, video, or code in response to a prompt. These systems do not think, imagine, or experience the world as people do. They have no childhood memories, ambitions, or desire to make something new. Yet they can write elegant prose, generate arresting images, compose music, suggest ideas, and imitate human emotion with increasing skill.
Machines do not need to create as humans do to produce work that audiences experience as creative. By modeling statistical patterns across enormous datasets and predicting what plausibly comes next, they generate outputs that resemble human originality closely enough to compete with it in practice. For creative professionals, that distinction matters because it helps explain which parts of creative work these systems can reproduce easily, which still depend heavily on human judgment and experience, and why arguments about whether a machine is “really creative” can obscure the more consequential question: what happens when its output is good enough for clients, employers, or audiences?
The sections that follow explain how generative AI became possible, how large language models use statistical prediction to produce remarkably human-like output, and why that output can appear intelligent or creative even though the systems do not think as people do. The article then turns to the economic question that matters most for creative professionals: what happens when machine-generated work becomes good enough to compete with human work?
How Computing Power Made Generative AI Possible
For most of history, creative production was bound by time. A novelist drafted only as quickly as sentences formed in her mind. A designer sketched at the pace of hand and eye. Even in collaborative studios, output expanded by adding more people, each constrained by the finite hours of attention and effort. The clock governed creative labor because idea generation unfolded at the pace of biological cognition.
For much of computing’s early history, progress faced a similar constraint. Early processors executed instructions one after another, like passengers boarding a bus, making complex tasks slow and costly. Simply adding more processors offered limited relief, as software struggled to coordinate them effectively. The conceptual leap came with parallel processing. By dividing problems into smaller components and executing them simultaneously across multiple processors, engineers compressed the computational clock. What once required hours could be completed in minutes. The constraint did not disappear, but it loosened. Time became something that could be engineered rather than merely endured.
Today, parallel architectures allow firms to marshal thousands of processors across distributed data centers, scaling computing power to meet demand in near real time. In data-intensive fields such as climate modeling and genomics, this shift proved transformative. Computation was no longer bound to a single sequential stream of operations; it could unfold across vast, coordinated networks. By the mid-2010s, artificial intelligence was advancing rapidly, thanks in part to these gains in computational scale. Much of the early excitement centered on machines that could see. Breakthroughs in image recognition enabled systems to identify faces, categorize photos, and assist autonomous vehicles in interpreting their surroundings. Yet when it came to understanding and generating human language, machines continued to struggle.
The difficulty was context. Language is not merely a sequence of words but a system of shifting relationships in which meaning depends on what surrounds it. Consider the word spring: short, familiar, yet entirely dependent on usage. A hiker might discover a mountain spring; a student could look forward to spring break; an athlete might spring into action. Each usage is correct, but together they reveal a central truth about language: meaning is situational rather than fixed. Stripped of context, even the simplest word can mislead, because nuance lies not in the word itself, but in the company it keeps.
Early language models processed small clusters of words sequentially and often lost track of broader relationships. They could recognize local patterns, but they struggled to sustain coherence across longer passages. The breakthrough came in 2017 with the introduction of transformer models, whose attention mechanism allowed models to evaluate relationships across entire passages simultaneously.
That improvement changed the economics of language generation. Better performance made larger models worth building, while parallel processing made them possible to scale. Once generation became both coherent and fast enough to use at scale, what had been experimental became usable. And what became usable could enter markets.
How Large Language Models Generate Text
The title of the 2017 paper was playful: Attention Is All You Need, a subtle nod to the Beatles’ anthem All You Need Is Love. Its eight authors, working within Google’s research ecosystem, proposed the transformer architecture that would become the foundation of modern large language models. Its attention mechanism allowed a model to weigh relationships among words across a passage rather than process language only step by step. When scaled up, the architecture produced dramatic gains across language tasks and was eventually adapted to generate images and video, compose music, and write software.
Today’s large language models produce prose that bears little resemblance to the stilted, mechanical output of early chatbots. When a user types a query into a system such as ChatGPT, the response can feel almost magical. Beneath the surface, however, lies a cascade of mathematical operations performed at staggering speed. What reads as fluent text is, at its core, the output of a statistical engine optimized to predict what comes next. This process, known as next-token prediction, is the mechanism at the heart of how large language models generate language.
The same predictive mechanism that makes large language models fluent also helps explain why they sometimes invent facts with confidence, a problem explored in my companion article, Why Generative AI Hallucinates: What Creative Professionals Need to Understand.
The process begins with disassembly. The model breaks incoming text into tokens, discrete units that may be whole words, subword fragments, or punctuation marks. Because LLMs operate through mathematical computation, each token is mapped to a numerical representation known as an embedding, which captures statistical relationships based on patterns of usage, context, and proximity learned from training data. Those patterns are derived from billions of lines of text drawn from books, websites, and online conversations.
The transformer architecture then processes the relationships among these tokens. A mechanism known as attention allows each token to weigh the relevance of others in the sequence. This sensitivity to context helps the model resolve ambiguity, discerning, for example, whether “spring” refers to a season, a water source, or sudden motion by interpreting the word in relation to those surrounding it. These representations then pass through multiple layers of computation, each refining the model’s representation of the context. By the final layer, the system produces a ranked list of candidate next tokens, each assigned a probability.
This is where language generation begins. The model does not simply choose the most likely next token every time. Instead, it samples from a range of candidates, usually favoring the most probable but sometimes selecting less likely alternatives to introduce variation. This controlled randomness, governed in part by a parameter known as temperature, helps make the output less repetitive and more humanlike. It also helps explain why a model may give slightly different answers to the same prompt. The selected token is added to the sequence, and the process repeats, one token at a time, until the response is complete.
To the user, the illusion is striking: polished language appears with fluency and coherence, yet there is no corresponding inner voice, intention, or experience behind it. A model can produce a witty sentence, an apt metaphor, or a paragraph that feels emotionally perceptive because it has learned statistical patterns in how humans use language. It does not feel sadness and then search for words to express it; it arranges the language of sadness into forms readers experience as meaningful.
Training large language models requires an immense volume of text. One researcher estimated that it would take an adult more than 2,600 years of nonstop reading to consume the material used to train GPT-3; newer models rely on even larger datasets. During training, a model repeatedly predicts subsequent tokens and adjusts billions of internal parameters when its predictions are wrong. Repeated across that enormous corpus, this process allows the model not merely to reproduce familiar phrases but to generalize to combinations it has never encountered in precisely that form.
Billions of trained parameters interact to produce the model’s behavior, making it difficult to trace why one response emerges rather than another, even for its creators. Yet statistical prediction at sufficient scale can produce output that appears remarkably intelligent and creative. That raises a harder question. How well do these systems perform when measured against the same kinds of tests traditionally used to assess human intelligence and creativity? The answer points toward a frontier that is at once exhilarating and deeply unsettling.
How Generative AI Performs on Tests of Intelligence and Creativity
Once systems built around prediction began performing tasks associated with reasoning and imagination, an uncomfortable question followed. If the underlying mechanism is statistical, should the output count as intelligent or creative?
How AI Performs on Tests of Reasoning
Standardized tests have long served as proxies for intellectual aptitude in academic and professional settings. Exams such as the Uniform Bar Exam, the SAT, and Advanced Placement tests are designed to assess analytical reasoning, language comprehension, and, in some cases, domain-specific knowledge. Their appeal lies in consistency, offering admissions offices, licensing bodies, and employers a common yardstick for evaluating skills that have historically correlated with academic success and professional performance.
When OpenAI released GPT-4 in 2023, it reported that the model had scored in the top 10 percent on the Uniform Bar Exam, a two-day assessment used by states like New York and Texas to determine whether a candidate is qualified to practice law. The result was striking not simply because the exam is difficult, but because the model was not trained specifically to take it. GPT-4 also performed strongly across a wide array of academic benchmarks, including the SAT, GRE, and Advanced Placement exams in subjects spanning the sciences, social sciences, and humanities, often placing in the top quartile of human test-takers.
What does this say about machine intelligence? These exams measure more than factual recall. They demand abstract reasoning, contextual judgment, and the ability to apply concepts to unfamiliar situations. That a machine can perform strongly on such tasks suggests that statistical learning can generate capabilities that closely resemble what we conventionally describe as intelligence.
How AI Performs on Tests of Creativity
Creativity presents an even more provocative case. It has long been regarded as a uniquely human trait, rooted in imagination, intention, and lived experience. Yet large language models complicate that assumption. Given open-ended prompts, such as proposing novel uses for everyday objects or improving a familiar product, they can brainstorm and extrapolate with ease, producing responses that are imaginative yet coherent, novel without being nonsensical.
Their creativity is not rooted in introspection or desire. It emerges instead from statistical patterns distilled from vast troves of human language. With well-crafted prompts, they can display a striking degree of what might be called synthetic originality. When evaluated using established psychological benchmarks, a growing body of research shows that these models can generate ideas that rival, and sometimes surpass, those associated with gifted humans. In one such study, GPT-4 was assessed using the Torrance Tests of Creative Thinking, a long-standing measure applied to both children and adults, and scored in the top one percent.
One of the most striking features of large language models is their ability to write convincingly in a wide range of voices, from a New Yorker essayist to an angry teenager. When prompted, they can evoke tenderness, anger, embarrassment, joy, or grief with uncanny precision. Through exposure to enormous quantities of text, they have internalized the patterns, cues, and cadences through which humans express emotion. This narrative dexterity mirrors a central achievement of human fiction: the ability to articulate emotional states beyond one’s own experience. That LLMs can accomplish this without consciousness makes the result all the more remarkable.
In literature, we admire writers for creating emotionally resonant characters they themselves are not. Tolstoy was never a young woman in imperial Russia. George Orwell was not Winston Smith, the central character in 1984. Margaret Atwood has never been a handmaid. What we value is their ability to attend closely to language and imagine how others think and feel. LLMs perform a related maneuver, but in reverse. They do not begin with lived emotion and search for words to express it. They start with the language of feeling and arrange it into forms that readers interpret as emotionally true. Even without inner experience, their outputs can still evoke emotional understanding in the reader, which in communication and art, is often what matters most.
The latest large language models therefore possess a formidable set of capabilities: strong reasoning performance, unexpected creative range, and remarkable narrative flexibility, all produced without human consciousness or experience. Those capabilities are not merely philosophically interesting. They are economically consequential.
Systems that can produce human-like output can reshape how products are designed, how labor is valued, and how creative work is priced. In such settings, questions about inner understanding may matter less than observable performance. If the output is good enough for a client, employer, or an audience, the absence of a human mind behind it may not prevent substitution. That is the point at which a philosophical question about machine creativity becomes an economic one.
Why Generative AI Can Seem Human
The ability of large language models to imitate human expression did not remain a technical curiosity for long. As researchers trained ever larger models on increasingly diverse datasets and vastly greater computing power, the systems became adept not only at processing language but at reproducing the texture of human interaction. They could imitate personalities, sustain conversations, shift tone, and produce responses that felt natural and emotionally attuned. What began as a quest for better language processing evolved into a race to build convincing engines of synthetic human expression.
The economic appeal of this shift was straightforward. Systems that behaved conversationally held users’ attention longer and lowered barriers to adoption. Interfaces that felt familiar encouraged repeated use, deeper interaction, and greater disclosure. In digital markets driven by engagement and scale, realism increased retention and expanded audiences. What looked like better design functioned as a growth strategy.
Replika, one of the most commercially successful AI companion platforms, makes this logic unusually clear. Its value lies in sustained emotional presence. Users interact with a named entity that remembers past exchanges, responds with warmth, and simulates concern. Premium features promise deeper intimacy rather than greater accuracy. Here, human likeness is not an interface choice but the product itself. The system succeeds precisely because users suspend disbelief. Whether its statements are true matters less than whether they feel responsive and personal. Replika makes visible what is often obscured elsewhere: when emotional plausibility drives engagement and revenue, imitation becomes the central design objective.
The same incentive operates more broadly, even when the application is less intimate. AI companies have invested heavily in making systems more conversational, expressive, and persuasive. In Silicon Valley, these efforts are framed in terms like “naturalism,” “fluid interaction,” and “realistic design.” Human likeness is presented not as deception but as usability. By recasting mimicry as intuitive design, AI companies have normalized products in which the boundary between human and machine is deliberately blurred.
As the capacity to mimic human expression grew, the implications for creative work became impossible to ignore. Systems could generate advertising copy, compose jingles, and produce images or designs in recognizable styles. Their output became convincing enough to rival, and increasingly substitute for, the work of writers, musicians, and designers. The boundary between original and synthetic expression grew faint.
Creative labor has long depended in part on the recognizability of human voice and judgment. When surface indistinguishability becomes sufficient for substitution, that advantage begins to weaken. What began as an engineering breakthrough has therefore become, for many creative professionals, a disorienting and threatening presence.
What Generative AI Means for Creative Professionals
None of this settles whether machine output should count as art, whether training practices are fair, how copyright law should treat generated material, or how creative professionals should respond. Those are separate questions. But they are easier to address once the technology itself is less mysterious.
The central lesson is that a machine need not create as a human does to compete with human work. Generative AI can already produce material that people find useful, persuasive, imaginative, or emotionally convincing. In some fields it augments human work. In others it substitutes for tasks that once required paid creative labor.
That is why the debate over whether machines are “really” creative can become a distraction. Markets rarely adjudicate metaphysical questions. Clients, employers, and audiences care about speed, cost, and usefulness. If machine output is good enough for the task, the absence of a human mind behind it may matter greatly in some settings and hardly at all in others.
For centuries, imaginative production was constrained by the finite time of individual minds. Generative AI loosens that constraint. It does not eliminate human creativity or reproduce the experience from which human creativity arises. It introduces a second system for producing expression, one that learns from the statistical residue of human culture and can generate variations at a scale no individual creator can match.
Understanding that system is not the same as embracing it. But as generative AI becomes more capable and more widely used, creative professionals have more reason than ever to understand how it works.
For evidence of how these changes are already appearing in U.S. creative labor markets, see my companion article, What Has Generative AI Actually Done to Creative Jobs? The Evidence Through 2025.
About Doug Woodham
Doug Woodham is an economist, former Partner at McKinsey & Company, former President of Christie’s Americas, and the author of two books about art and creative markets. At McKinsey, he helped co-found the firm’s Business Technology Office and advised companies on strategy, technology, and organizational performance. He holds a Ph.D. in economics from the University of Michigan, where his training included econometrics and the study of how innovation reshapes firms and markets. He writes about how artificial intelligence is changing the economics of creative work, drawing on his experience in economics, technology, business strategy, and creative industries.
