AI Labor Tracker / Foundations

Knowledge is free. Creativity isn't.

The price of machine thinking is collapsing faster than any technology on record, which makes original ideas, taste and the nerve to bet the scarce things worth building.

Knowledge is free. Creativity isn’t.

I believe that more than almost anything else I write about. In this piece I want to test it against the best research I can find, including the studies that cut against it.

The first half isn’t really an opinion anymore. It’s a price chart.

The price of thought

When OpenAI released its o3 model, it scored 75% on GPQA Diamond, a hard graduate-level science test, at a cost of about 30 cents per question. Less than eighteen months later, a newer model matched that score for about four-hundredths of a cent.

That is a 725-fold drop, according to a September analysis by Epoch AI. The researchers compare it to a $50,000 car falling to $69.

Across benchmarks, Epoch estimates that the cost of reaching a given level of AI performance falls about 47% per quarter, or roughly 13 times per year. Its earlier work found declines ranging from 9 to 900 times a year depending on the task.

To put that in context, Epoch lined it up against other famous cost curves. Electricity got about 5% cheaper per year from 1892 to 1973.

Computing got about 1.5 times cheaper per year from 1940 to 2001. DNA sequencing, the poster child for collapsing costs, got about 1.8 times cheaper per year this century.

Epoch is careful to flag the caveats. Benchmarks are not the same as useful work, and labs tune models to score well on them.

Fair. But even if the true number is a fraction of 13, nothing in economic history has gotten cheaper this fast.

When something gets that cheap, it stops being the thing people pay for. Nobody bills for arithmetic anymore. Soon, very few people will bill for knowing things.

Who gains first

I want to be clear that this is mostly good news.

In a Science experiment, economists Shakked Noy and Whitney Zhang gave 453 college-educated professionals writing tasks. The group with ChatGPT finished about 40% faster, and quality rose 18%. The weakest writers gained the most, which narrowed the gap between workers.

A study of more than 5,000 customer support agents found the same pattern. Access to an AI assistant raised the number of issues resolved per hour by about 15%, with gains around 30% for the least experienced workers, in the Quarterly Journal of Economics.

And in the best-known field experiment, 758 Boston Consulting Group consultants using GPT-4 completed 12.2% more tasks, 25.1% faster, with quality more than 40% higher, on work that sat inside what the researchers called AI’s “jagged frontier,” per the Harvard Business School paper.

Cheap knowledge makes ordinary people better at their jobs. That is a big deal, and I’m for it.

But look closely at who gains. In the customer support study, the tool had “minimal impact on experienced and highly skilled workers,” according to the NBER version of the paper. The researchers suggest the AI was spreading the best practices of the top agents to everyone else.

Think about what that means. The know-how that used to separate a great agent from an average one got bottled and handed out. The best agents’ expertise became a free input.

That is wonderful for the new hire. It is a warning for anyone whose edge was simply knowing more than the person next to them.

This is what I mean by knowledge being free. Facts were never the scarce part; libraries have been free for a long time. What’s new is that applied know-how, the kind you used to get from ten years on the job, now ships inside the software.

When knowledge is the product

Now the hard part for workers. If what you sell is knowledge, information, or a competent first draft, your price is falling with the chart above.

That is already showing up in freelance markets, where prices move first. After ChatGPT launched, writers on Upwork saw a 2% drop in monthly jobs and a 5.2% drop in monthly earnings, according to research in Organization Science by Xiang Hui, Oren Reshef and Luofeng Zhou. Image-generation tools brought a similar hit to design freelancers.

A larger study of job posts found that demand for automation-prone freelance work fell about 21% within eight months of ChatGPT’s release. Writing jobs fell 30%, and software and web development fell 21%, according to Demirci, Hannane and Zhu in Management Science.

These are not catastrophic numbers. They are a signal. The work that got hit first is the work that looks most like retrieving and rearranging what’s already known.

What stays scarce

So if knowledge is getting cheap, what isn’t?

My answer is four things. Original ideas. Taste, meaning the ability to tell what’s good.

Judgment, meaning the ability to tell when the machine is wrong. And the nerve to bet on something before the evidence is in.

The research has a lot to say about the first three. The fourth one you’ll have to take on faith from me for now.

Original ideas, and the sameness problem

The most useful study here comes from Anil Doshi and Oliver Hauser, published in Science Advances. They had people write short stories, some with access to AI-generated ideas and some without.

Individually, AI helped. Writers who got five AI ideas produced stories rated 8.1% more novel and 9.0% more useful. The least creative writers gained the most, up to 26.6% on “well written.”

But collectively, the AI-assisted stories looked more like each other. The researchers call it a social dilemma: each writer is better off using AI, but the pool of stories as a whole gets less diverse.

That finding keeps replicating. Lennart Meincke, Gideon Nave and Christian Terwiesch found that ChatGPT raises the creativity of individual ideas in brainstorming but shrinks the diversity of the overall pool, in Nature Human Behaviour.

In visual art, a study of artists on a large art-sharing platform found that those who adopted text-to-image tools became about 25% more productive, but the average novelty of their work declined, per PNAS Nexus. The peak novelty went up.

That last detail is the whole thesis in one sentence. The average gets more average.

The best gets better. If everyone has the same machine, the thing that separates you is what you bring that the machine didn’t.

Taste and judgment

The consultant study has a second half that gets less attention. On a task deliberately placed outside AI’s frontier, consultants using GPT-4 were 19 percentage points less likely to reach the right answer than those working alone, according to the HBS paper.

Read that again. The tool made skilled people worse when they trusted it on the wrong problem. Knowing which problem is which is a judgment call, and judgment is not in the price chart.

Where the evidence cuts against me

I promised honesty, so here’s the part I don’t love.

At Wharton, researchers pitted GPT-4 against students at an elite university, enrolled in a product design course, in generating new product ideas. GPT-4’s ideas had higher average purchase intent, 46.8% versus 40.4%, and 35 of the 40 top-rated ideas came from the machine, in a paper aptly titled “Ideas are Dimes a Dozen”.

In another study, more than 100 NLP researchers blind-reviewed research ideas, and they rated the AI-generated ones as more novel than the human experts’ ideas, with slightly lower feasibility, per Si, Yang and Hashimoto.

And in a classic test of divergent thinking, asking for unusual uses of everyday objects, AI chatbots beat the average human in a Scientific Reports study.

So if your definition of creativity is “comes up with lots of decent ideas,” the machine wins. I’ll concede that. The idea stage is getting cheap too.

But the same studies carry the counterpunch.

In the divergent-thinking study, the best humans still matched or beat the best AI. The title of the paper is literally “Best humans still outperform artificial intelligence.”

And the NLP researchers ran a follow-up. They had 43 people actually execute the ideas, spending more than 100 hours each. After execution, the AI ideas’ scores dropped significantly more than the human ideas’, and on several measures the ranking flipped in favor of the humans, according to “The Ideation-Execution Gap”.

My read: AI is very good at producing ideas that sound new. It is much worse at producing ideas that turn out to be good once someone tries them.

The scarce skill is picking the one that works and seeing it through. A demo is not a product, and anyone who has shipped software knows the gap between the two is where all the work lives.

The nerve to bet

Nobody has run a randomized trial on courage.

But think about what actually creates value in an economy. Somebody decides to start the company, launch the product, take the case, open the restaurant, write the book nobody asked for. That decision happens under uncertainty, with someone’s money and reputation on the line.

An AI can give you a hundred ideas and a thousand-word analysis of each one. It cannot want one of them badly enough to quit its job.

It doesn’t carry the risk, so it doesn’t carry the conviction. When knowledge is free, the person who will commit becomes more valuable, because the analysis everyone else hides behind costs nothing.

That is why I keep calling this relevant to ordinary workers and not just artists. The nurse who notices the pattern, the electrician who knows the drawings are wrong, the manager who kills the project nobody else will. All of that is creativity in the sense that counts.

How to build it

If this thesis is right, creativity is a skill to train, not a gift you have or don’t. Here’s what the research suggests.

Generate first, then prompt. The homogenization studies show that when everyone starts from the AI’s ideas, everyone ends up in the same place.

Write your own rough ideas before you open the chat window. Then use the machine to push on them.

Ask for variance on purpose. Meincke, Ethan Mollick and Terwiesch found that prompting techniques such as chain-of-thought can make AI-generated idea pools much more diverse, in a working paper on prompting diverse ideas. Don’t accept the first list.

Build your taste deliberately. Study great work in your field and be able to say why it’s great. Taste is pattern recognition about quality, and you get it the old-fashioned way, by looking at a lot of examples and being honest about what’s mediocre.

Execute more than you ideate. The ideation-execution gap tells you where the value is. Ship small things, see which ones survive contact with reality, and learn from the ones that don’t.

And take a real bet now and then. Small, survivable, but real. Nerve is a muscle.

Where to spend your time

Stop competing with the machine on what it does for nearly free. Don’t build your career on knowing things, summarizing things or producing the competent first draft. That price is falling 13 times a year.

Hand all of that to the tools, and spend the time you get back on the parts that stay scarce: original ideas, taste, judgment, and the willingness to commit. We’ll be measuring how fast machine thinking is growing in our Gross Intelligence Product series. The faster that number grows, the more those four things are worth.

Knowledge is free. Creativity isn’t. Put your next free hour into the second half of that sentence.

Sources

1. Emberson, L. and Roodman, D., “The plunging price of thought,” Epoch AI, Sep 22, 2026. https://epoch.ai/publications/the-plunging-price-of-thought
2. “LLM inference prices have fallen rapidly but unequally across tasks,” Epoch AI, Mar 2025. https://epoch.ai/data-insights/llm-inference-price-trends
3. Noy, S. and Zhang, W., “Experimental evidence on the productivity effects of generative artificial intelligence,” Science, Jul 2023. https://www.science.org/doi/10.1126/science.adh2586
4. Brynjolfsson, E., Li, D. and Raymond, L., “Generative AI at Work,” Quarterly Journal of Economics, 2025. https://academic.oup.com/qje/article/140/2/889/7990658
5. Brynjolfsson, E., Li, D. and Raymond, L., “Generative AI at Work,” NBER Working Paper 31161, 2023. https://www.nber.org/papers/w31161
6. Dell’Acqua, F. et al., “Navigating the Jagged Technological Frontier,” Harvard Business School, 2023 (published in Organization Science, 2025). https://www.hbs.edu/faculty/Pages/item.aspx?num=68729
7. Hui, X., Reshef, O. and Zhou, L., “The Short-Term Effects of Generative Artificial Intelligence on Employment: Evidence from an Online Labor Market,” Organization Science, 2024. https://pubsonline.informs.org/doi/10.1287/orsc.2023.18441
8. Demirci, O., Hannane, J. and Zhu, X., “Who Is AI Replacing? The Impact of Generative AI on Online Freelancing Platforms,” Management Science, 2025. https://pubsonline.informs.org/doi/10.1287/mnsc.2024.05420
9. Doshi, A. R. and Hauser, O. P., “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” Science Advances, Jul 2024. https://www.science.org/doi/10.1126/sciadv.adn5290
10. Meincke, L., Nave, G. and Terwiesch, C., “ChatGPT decreases idea diversity in brainstorming,” Nature Human Behaviour, May 2025. https://www.nature.com/articles/s41562-025-02173-x
11. Meincke, L., Mollick, E. and Terwiesch, C., “Prompting Diverse Ideas: Increasing AI Idea Variance,” SSRN, 2024. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4708466
12. Zhou, E. and Lee, D., “Generative artificial intelligence, human creativity, and art,” PNAS Nexus, Mar 2024. https://academic.oup.com/pnasnexus/article/3/3/pgae052/7618478
13. Girotra, K., Meincke, L., Terwiesch, C. and Ulrich, K., “Ideas are Dimes a Dozen: Large Language Models for Idea Generation in Innovation,” Wharton Mack Institute, 2023. https://mackinstitute.wharton.upenn.edu/wp-content/uploads/2023/08/LLM-Ideas-Working-Paper.pdf
14. Si, C., Yang, D. and Hashimoto, T., “Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers,” arXiv, Sep 2024. https://arxiv.org/abs/2409.04109
15. Si, C., Hashimoto, T. and Yang, D., “The Ideation-Execution Gap: Execution Outcomes of LLM-Generated versus Human Research Ideas,” arXiv, Jun 2025. https://arxiv.org/abs/2506.20803
16. Koivisto, M. and Grassini, S., “Best humans still outperform artificial intelligence in a creative divergent thinking task,” Scientific Reports, Sep 2023. https://www.nature.com/articles/s41598-023-40858-3