Shiwen Dialogue 046 | AI Didn't Kill Freelancing, Just One-Skill Work

Opening Dialogue

Shiwen: Yongliang, where does this research come from?

Yongliang: The publishers are OpenAI and Upwork, which is one of the largest freelance matchmaking platforms in the world. The research is part of OpenAI’s “Work at the Frontier” series, published on OpenAI’s official website on November 10th. It’s important to clarify the data scope: ChatGPT’s usage data covers January 2025 to June 2026, while Upwork’s data is from the first half of 2025 to the first half of 2026. All this data comes from Upwork’s platform and does not represent the entire freelance market. Upwork also reported its own background: their “2026 Future Workforce Index” shows that 38% of U.S. skilled knowledge workers work as freelancers, with previous research estimating that freelance knowledge work generated $1.5 trillion in income in the U.S. in 2024. These numbers are from stakeholders.

Shiwen: How well has it been disseminated?

Yongliang: Very weakly. When it appeared on Hacker News, it had only 29▲ and 19 comments, with little discussion in the tech community. But I think this contrast itself is informative: a well-founded study, but because the publishers are AI vendors and matchmaking platforms, people initially doubt its stance before examining the content. Moreover, among the dozen or so comments, several were from people who actually accepted jobs and worked on it, providing more detailed information than the report’s main text. Today, we’ll follow this order: first, what it found, then what practitioners say, and finally, what should not be believed.

Q1: Replacement is not comprehensive; it’s a “shift”

Shiwen: What is the most core finding of the study?

Yongliang: It’s not that “AI has replaced freelancing,” but rather that replacement and shift occur simultaneously. They compared the usage structure on ChatGPT with the demand structure on Upwork: writing accounted for a much higher proportion of ChatGPT usage than Upwork demand, while video and animation were the opposite—less use on ChatGPT, but higher demand on Upwork. Among 63 job categories, 43 showed the pattern of “common activities on ChatGPT losing share, uncommon activities gaining share,” and the remaining 20 were the reverse. OpenAI itself acknowledges that ChatGPT usage cannot explain the changes in freelance demand alone.

Shiwen: How does that explain it?

Yongliang: They introduced a concept called “capability overhang”: current ChatGPT usage only reflects how people are accustomed to using it, not the full range of tasks AI can handle. In other words, there is a gap between what the model can do and what people have it do. Demand changes follow people’s usage habits, not the model’s capacity limits. I agree with the honesty of this concept; it acknowledges that the current replacement is just the first wave, occurring according to habits. But we also need to see the other side: this concept gives vendors a position where they can attack or defend themselves. Now, with less replacement, it can be said that habits haven’t kept up; in the future, with more replacement, it can be said that capabilities have been realized. The concept itself holds up, but we must be cautious when using it for predictions.

Shiwen: Are there specific numbers for categories?

Yongliang: The report only provided changes in share, measured in percentage points. Commenter yorwba converted this: the share of content writing dropped by 0.57 percentage points to 1.27%, which is roughly a 31% decrease. It’s important to note that the report didn’t provide any absolute change figures, so statements like “freelance work has decreased” or “increased” cannot be inferred from this study. Anyone who interprets these ways is adding context.

Q2: What has become more expensive: from buying tasks to buying integrations

Shiwen: The share that has decreased is for single tasks. So, what has increased? Yongliang: It’s for jobs that require a broader range of skills. The study provides two examples. First, by the end of 2025, clients were requesting AI for more types of work—chatbot development, video production, marketing automation, application integration all saw AI demands. The diversity of tasks involving AI-related posts increased by 45% compared to early 2023, rising from about 14.6 typical tasks to around 21.2 tasks. Second, they cited a case where one post required AI to generate videos, video editing, and animated graphics, all delivered by the same freelancer. Previously, this might have been three separate projects for three people. Shiwen: And the price? Has the ability to integrate become more expensive? Yongliang: The study says that the hourly rates for AI-related projects are similar to other jobs, but fixed-budget parts involve more work hours. OpenAI interprets this as higher budgets reflecting a wider range of work. Note the wording—“partially reflected.” They didn’t say the premium comes from scarce skills; they only said the range has broadened. This is a relatively cautious statement in the report. Shiwen: There seems to be a more direct interpretation in the HN comments section? Yongliang: Yes, commenter armcat’s interpretation I think is sharper than the report itself: AI is commercializing “single tasks” while increasing the demand for people who can combine tasks into complete deliverables. What’s valuable isn’t the single skill, but the combination. This matches the report’s data: the share of single-skilled jobs has been taken over by ChatGPT, while the diversity of tasks involving cross-domain integrations is increasing. Civicsquid’s observation complements this: some categories people simply use ChatGPT to complete themselves, like writing, illustration, technical support; others still post tasks and hire people, especially those with multi-domain skills. The dividing line between hiring and not hiring lies in whether “a tool can handle everything at once.”

Q3: The most honest voice in the comments section: cleaning up AI messes

Shiwen: What do practitioners themselves say? Yongliang: This is the most valuable part of the comments section. Freelancer aavci said, from his perspective, many new orders are “cleaning up messes done by AI,” not new opportunities starting from scratch. Tapvt was even harsher: previously it was messes outsourced to cheap development teams that he cleaned up; now it’s messes driven by AI development. The same work, just with less funny comments in the code. Jamestimmins claimed to have done this kind of work, saying that managing a large codebase is challenging, but cleaning up big messes is also interesting. Shiwen: Is this considered a new job type?

Yongliang: I have led teams of 70 people and taken over systems left by others. This is my experience, not statistics: cleaning up messes always happen, but AI has changed the speed at which messes are created. Before, an outsourced team would take a month to create one problem, now one person with AI can do it in a week. With many problems, filling them becomes a stable source of demand. This won’t be written in the report because it doesn’t fit with the narrative of “AI creating new demands,” but it’s true for practitioners: Code generated by AI seems to run, what’s lacking is engineering finishing, testing, boundaries, maintainability—these are precisely the other side of integration capabilities.

Shiwen: Are there other perspectives?

Yongliang: Yes, about why some tasks refuse to be handed over to AI. eatonphil says they insist on paying human illustrators because AI illustrations damage brand image. spacechild1 agrees, and it’s particularly striking for companies with market values of hundreds of billions. This complements the picture from Q2: What is commoditized are tasks that “just work enough,” while aspects like brand, aesthetics, responsibility that require human endorsement become more dependent on specific individuals.

Q4: First doubt the stance, then look at the content

Shiwen: Now it’s time to talk about what isn’t believed. What are the interests involved in this study?

Yongliang: Both publishers have the narrative motivation of “AI creating new demands.” OpenAI is the seller of AI; the more customers see AI as useful, the better it sells. Upwork is a matchmaking platform; the more diverse the orders and the larger the budget, the higher its commission. So this study structurally doesn’t answer that biggest question. HN commenter darrinm pointed out: The report avoids answering whether the volume of freelance work and total compensation are increasing or decreasing, and by how much. pllbnk’s criticism is heavier, suspecting the report itself to be superficial, like something generated by an agent, and even possibly written to appease certain narratives. This accusation has no evidence; I don’t report it as fact, but it reflects readers’ first reaction to such joint studies. ActionHank says they’re tired of these companies getting them involved in every issue.

Shiwen: So can this study still be used?

Yongliang: It can, but with reduced effectiveness. My criteria are: This is experience, not statistics. Data published by stakeholders can be referenced, but the extent needs to be discounted, and conclusions should be re-evaluated. In this study, I relied on structural changes—single tasks losing share, combined tasks becoming more diverse, because their selling motives are opposite and not hidden. What I doubt is any interpretation that “the market is growing.” Also, note that HN only has 29▲; this isn’t a hot topic across the network, it’s almost unnoticed. We wrote about it because of its perspective, not because of its popularity.

Q5: Three suggestions for those who take orders and those who place them

Shiwen: For those who make a living by skill, what are your suggestions?

Yongliang: Three. First, check if your task can be completed in one go: If your deliverable can be created with just one round of ChatGPT dialogue, and 80% of it is already there, the data is warning you—don’t wait for the platform to teach you. Second, move towards combinations: The report’s example of “AI generating videos plus editing plus dynamic graphics by one person” is a template. Learn one adjacent skill to turn yourself from a “task executor” into a “deliverable owner.” Third, turn endorsement into an asset: eatonphil’s reminder that some people are willing to pay a premium for “specific individuals”—names, reputation, accountability, things AI can’t offer.

Shiwen: Who is responsible for issuing tasks? When companies hire people to work on projects, does the logic change?

Yongliang: It does. In the past, I hired people and provided outsourcing services, pricing based on individual skills—those who could write copy or edit videos got paid separately. Now, the right approach is to price based on the ability to deliver integrated results: someone who can coordinate tasks, use AI, and assemble fragmented components into acceptable deliverables should be paid by the scope of work, not by the number of hours worked. The increase in fixed-price budgets and hours in reports reflects this logical trend. Another important point in management is that AI has increased the speed at which messy situations are created, so acceptance standards must also rise; otherwise, the savings from outsourcing will turn into internal cleanup costs, and AAVCIs will only make money from this.

Shiwen: A summary in one sentence?

Yongliang: What should be remembered from this research is not any single percentage point, but the dividing line: tasks that can be completed by a tool will eventually become less valuable. Only work that requires people to assemble fragmented components into complete deliverables and who are willing to take responsibility for the results is what will see price increases in the AI era. The data belongs to them, but the judgment is yours.

---

Signature: Yongliang