AI and Creativity: What Happens When We Outsource the 'Boring Work'?

AI is stealing your flashes of inspiration.

Author and source: Yuan Zhen's Business Notes

Over the past two years, nearly every AI company has been telling the same story:

By leaving repetitive tasks to AI, humans can free up time to focus on more creative endeavors.

This arrangement sounds flawless: let AI organize information, process spreadsheets, and write meeting minutes, while humans focus on strategy, products, and creativity—it’s a perfect division of labor.

But it quietly assumes a premise: repetitive labor and creative work are two distinct things that can be cleanly separated.

History hasn't given such a straightforward answer.

Many of the most important breakthroughs don’t occur after repetitive work ends—they happen in the gaps between repetition, failure, and distraction. You wrestle with a problem for a long time, repeatedly calculating, observing, and revising, until a simple moment suddenly connects all the scattered pieces.

If AI takes away all the tedious work for us, what it frees up may be more than just time.

It may also take away the very setting where inspiration occurs.

The story of Newton and the apple is probably the most successful "innovation myth" in human history.

An apple fell from a tree and hit Newton, inspiring him to discover gravity.

The real version isn't this simple.

Newton’s friend William Stukeley recorded in his memoirs that in 1726, the two were drinking tea beneath an apple tree. Newton recalled that at the time, he was deep in thought when he saw an apple fall, prompting him to wonder: why does the apple always fall straight toward the ground, rather than upward or sideways? Only then did he “gradually” extend this force to the Earth, the Moon, and other celestial bodies. [1]

Figure 1 | Apple did not provide the answer; it merely ignited the material Newton had accumulated through long-term reflection.

An apple didn't hit Newton, nor did it directly yield a law.

It’s more like a match falling into a warehouse already full of materials.

The mathematician Poincaré's experience is more typical.

He had been continuously studying a class of functions, performing extensive calculations and repeatedly failing. Later, he temporarily set aside his work and went traveling. At the very moment he stepped onto a stagecoach, a crucial insight suddenly struck him: the transformations he had been studying were the same as those in non-Euclidean geometry.

The idea came to him so quickly that he didn't even work through it on the spot. After returning from his trip, Poincaré sat down at his desk and spent time thoroughly verifying it.

He later broke the creative process into several stages: first, conscious and strenuous preparation; then, what appears to be idle incubation; followed by sudden illumination; and finally, a return to rigorous verification. [2]

The proof of Fermat's Last Theorem is the same.

Andrew Wiles worked in secret for seven years, announced a successful proof in 1993, but a flaw was discovered a few months later. He spent another year, working repeatedly with his student Richard Taylor to find a solution, and finally closed the gap in September 1994. [3]

People remember the final moment, but what the mathematician truly experienced were thousands of failed calculations, dead ends, and restarts.

Insight never comes as a gift from the sky. It’s the delayed reward your brain gives after years of diligent effort.

Today, we like to divide work into two categories: low-value tasks for machines, and high-value tasks for humans.

The problem is that the human brain doesn't work according to the tables in consulting reports.

A journalist reviewing dozens of financial reports might notice that two sets of numbers don’t match; a product manager carefully reading through negative reviews might suddenly realize users aren’t complaining about the feature itself; a doctor, after repeatedly reviewing images, gradually develops an instinct for detecting abnormalities.

These tasks may not seem impressive on their own, and even feel dull. But they leave behind countless small patterns in the mind.

Intuition is often simply experience so refined that you can no longer explain it step by step.

The discovery of penicillin was just like that.

In 1928, Fleming returned from vacation to find a petri dish containing staphylococcus bacteria contaminated by mold, with no bacterial growth around the mold. The contamination was accidental, but the discovery was not. To someone without a long-term observation of bacteria, it would have appeared only as a dirty dish meant to be thrown away.

More easily overlooked is that Fleming did not immediately develop a drug from this discovery. Nearly a decade later, Florey, Chain, and a team revived the research. They cultivated the mold in milk buckets, food cans, and even bathtubs, and searched across the United States for high-yield strains, ultimately transforming a laboratory accident into an industrial product capable of saving lives.[4]

Figure 2 | Pollution is accidental; discovery depends on long-term observation; drugs are more the result of engineering over the subsequent decade.

History loves to tell the “Eureka” moments, but industry truly relies on the long, repetitive, and unglamorous engineering that follows.

Einstein later described his thought process to mathematician Hadamard, saying that what often appeared first were not complete words, but images, symbols, and elements with a sense of motion. He would repeatedly combine these elements, with language coming afterward. [5]

Creation is not about generating something from nothing, but rather about recombination.

What you've seen, what you've done, and where you've failed determine what can be reconnected in your mind.

If documents are always summarized by AI, first drafts are always generated by AI, and data is always cleaned by AI, people will indeed have an easier time. But over time, what we encounter will be only the machine-processed end products, not the raw, inconsistent, noisy original materials.

Efficiency has improved, but sensitivity may have diminished.

Walter Isaacson asked in The Innovators: Who actually invented the computer and the internet?

In the end, he realized that this question could not be answered with just one person's name.

The digital revolution wasn't suddenly invented by a genius in a garage. Early computers, transistors, chips, software, and the internet all emerged from generations of incremental improvements—someone conceived the idea, someone soldered circuits, someone wrote code, and someone turned expensive prototypes into products accessible to ordinary people.

When mentioning the Harvard Mark I computer, Grace Hopper’s account emphasizes Howard Aiken, while IBM’s version highlights the unnamed engineers whose numerous incremental improvements—such as counters and card readers—collectively formed the final machine. [6]

Figure 3 | The Innovators repeatedly emphasizes: the digital revolution stems from collaborative, incremental efforts by many individuals over multiple rounds.

This is not denying genius, but rather revealing the true structure of innovation.

A groundbreaking event typically requires both the imaginative vision of a few and the hard work of the many. The former sets the new direction, while the latter turns that direction into reality. The two are not substitutes for each other.

The most appealing aspect of AI is that it seems capable of playing both roles at once: generating ideas and carrying out execution.

But new issues have also arisen here.

In 2024, Science Advances published an experiment on generative AI and writing. Those who received AI-generated ideas produced stories that were, on average, more creative and better completed, with particularly noticeable benefits for individuals with initially lower creativity; however, when all the works were examined together, the AI-assisted stories were more similar to one another. [7]

AI has raised the baseline for individuals while narrowing the differences among groups.

This is much like today’s content platforms: everyone writes more smoothly than before, with perfect headlines, structures, and catchy phrases; but after reading ten articles, it feels like you’ve only read one.

When everyone asks the same model for the "first idea," we skip the blank page—and may collectively head toward the most predictable answer.

This does not mean people should reject AI and waste time again on copying, pasting, and mechanical typesetting.

The calculator didn't destroy mathematics, and the search engine didn't destroy knowledge. What tools truly change is the level of ability that humans must retain.

AI can take over the hard labor, but we must not forget why the wall was built here in the first place.

This concern is no longer just a philosophical debate. In 2025, Microsoft Research and other institutions surveyed 319 knowledge workers and collected 936 real-world AI usage cases. The results showed: the more people trust AI, the less critical thinking—defined as reasonable and reflective thought—they apply to tasks; the more they trust their own professional judgment, the more critical thinking they demonstrate. AI has not eliminated thinking, but has shifted it from “generating directly” to “verifying, integrating, and overseeing.” The problem is, if people only see a polished final product, verification can easily devolve into mere nodding. [8]

More direct signals come from writing and programming. A 54-participant EEG writing study at the MIT Media Lab found that participants who consistently used large models for writing showed the weakest brain network connectivity and reported lower recall of article content and weaker feelings of “this is my work.” However, this study remains a small-sample preprint and cannot be used to conclude that AI will permanently make people less intelligent. [9] In a randomized trial by METR involving 16 experienced open-source developers and 246 real-world tasks, it was found that after using AI tools from early 2025, developers took 19% longer to complete tasks—despite initially expecting AI to make them faster. [10] METR’s 2026 follow-up concluded that next-generation tools likely now provide acceleration, but due to clear selection biases in participants and tasks, the magnitude of improvement remains unreliable. [11] What AI may first erode is not ability itself, but people’s judgment of whether they are still thinking.

For individuals, there are at least three things that should not be outsourced lightly.

First, where does the issue come from?

Don’t start by asking AI, “What should I write?” or “What should I research?” Go to the client’s site first, read the original materials first, and let yourself notice what feels off. A good question often comes from the details others overlook because they’re too much trouble.

Second, where does the first-round judgment come from?

You can have AI expand, refute, and supplement information, but it's best to first preserve a raw, unprocessed version created by you—even if it's just a few sentences—so you know your original perspective. Otherwise, human-AI collaboration can easily turn into humans merely polishing AI's output.

Third, who is responsible for the results.

Having done foundational work myself, I know where AI goes wrong. Someone who has never written code will struggle to tell whether code merely runs or is production-ready; someone who has never analyzed financial statements will find it hard to spot hidden metric inconsistencies in an otherwise polished analysis.

So, what can be entrusted to AI?

The criteria are not about “whether it’s creative” or “whether it’s repetitive,” but rather three things: whether the acceptance criteria are clear, whether the process is traceable, and whether it can be rolled back if errors occur. Tasks such as transcription, formatting, data cleaning under known rules, code templates and test cases, multiple versions of copy and design sketches on the same topic, are all suitable for AI. Strategic trade-offs, factual standards, brand aesthetics, system architecture, and final approvals should not be handed off to AI simply because it’s faster.

After handing over the task, humans must still engage in internal debate with AI. Before taking action, write down your own judgment and criteria for evaluation; during generation, ask the AI to simultaneously provide counterarguments, failure conditions, and sources of evidence; after generation, return to the original materials to verify facts, recalculate key numbers, and run critical tests; periodically schedule “manual training” sessions without AI. Writers must retain independent interviews and first drafts; programmers must preserve code reviews and troubleshooting logs; designers must keep records of concept selections and final visual decisions. Only those who can still verify, dare to overturn, and are capable of taking back control truly deserve to delegate their work to AI.

What companies should truly be wary of is not how much AI employees are using, but whether the organization is still generating firsthand experience.

If young professionals no longer conduct fundamental analysis, no longer collaborate with clients to refine requirements, and no longer go through the process of failed first drafts, the company may achieve faster deliveries in the short term—but risk losing the pathway to developing sound judgment over the long term.

In the past, a newcomer would spend three years of hard, painstaking work to gradually become an expert with intuition. Now, AI can help a newcomer deliver a 70-point answer on their first day.

But to go from seventy to ninety, he still needs to have personally encountered the problems, made the mistakes, and paid the price for the results.

Is AI enhancing creativity or diminishing it?

The answer depends on where we place it.

Let AI handle repetitive tasks you've already figured out—it’s a lever. Let AI replace you in engaging with reality, identifying problems, and making initial judgments, and it will gradually become a crutch.

True creativity is never just "coming up with an idea."

It involves familiarity with the material, sensitivity to anomalies, the ability to endure long periods without answers, and, once insight strikes, rigorously verifying it again and again.

These sections are slow and undignified, and even difficult to include in performance evaluations. Yet almost all major theorems, scientific discoveries, and industrial innovations in history have emerged from here.

AI can certainly help us pick apples.

But whether that person under the tree could think of gravity still depends on how long he had been thinking before the apple fell.

Reference materials

[1] William Stukeley, Memoirs of Sir Isaac Newton's Life, Newton Project (University of Oxford).

[2] Henri Poincaré, “Mathematical Creation,” 1908.

[3] Materials from the Mathematical Institute at the University of Oxford and the University of Bristol regarding Andrew Wiles and Fermat's Last Theorem.

[4] Science Museum Group, “How was penicillin developed?”

[5] Jacques Hadamard, The Mathematician's Mind: The Psychology of Invention in the Mathematical Field.

[6] Walter Isaacson, The Innovators: How a Group of Hackers, Geniuses, and Geeks Created the Digital Revolution

[7] Anil R. Doshi and Oliver P. Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” Science Advances, 2024.

[8] Hao-Ping Lee et al., “The Impact of Generative AI on Critical Thinking,” CHI 2025.

[9] Nataliya Kosmyna et al., "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Tasks," 2025 preprint.

[10] Joel Becker et al., "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity," METR, 2025.

[11] Joel Becker et al., “We Are Changing Our Developer Productivity Experiment Design,” METR, 2026.