Table of Contents

"Will AI take my job" is no longer the right question for this story. What a study of Upwork contract data showed was not a change in whether the work exists, but a change in how you get chosen. Track record, credentials and ratings all count for less than they used to. What counts for more is price.

And over the same period, something is moving in exactly the opposite direction. The number of people becoming freelancers is surging.

⚠️ Both of these are happening at once, and that is where we are. People are flowing into a market where the signals that differentiate you have stopped working. The pressure crushing the middle is arriving from the demand side and the supply side at the same time.

📌 About the numbers in this article: I used only peer-reviewed papers and official platform research, and checked every figure against the publisher itself. No second-hand citations from roundup posts, and no in-house data from freelance staffing agencies (those firms profit when more people go freelance, and their numbers cannot be traced to a primary source). Chapter 10 alone is speculation, and says so at the top of the chapter.

1. The conclusion — in a market people are pouring into, the way you get chosen is broken

What happened on the demand side

Track record, credentials and ratings lost weight, and price gained weight. The premium for skill has been eroding faster the more recent the quarter.

What happened on the supply side

Among skilled US knowledge workers, the freelance share went from 28% to 38% in a single year. And 58% of full-time employees are considering the switch.

Look at only one side and you reach a completely different conclusion. Demand side alone says rates are falling; supply side alone says freelancing is booming. Look at both at once and what you see is polarization, accelerating.

2. Where things stand — 5 studies on one page

Source Data Key findings
Siddiq & Zhang
UCLA Anderson (June 2026)		 Upwork, 49,610 freelancers and 2.26 million contracts
January 2021 to March 2026		 Weight on human-capital signals -7.8% / weight on price +1.1% / contract count -7.0%. In the most recent four quarters the effects widen: demand -9.6%, human-capital weight -10.1%, price weight +1.8%
Stanford Digital Economy Lab
Canaries (updated August 2026)		 ADP payroll data
November 2022 to June 2026		 Employment for ages 22 to 25 in AI-exposed occupations is about 19% lower (up from 15% in July 2025) / no comparable gap for experienced workers / it is fewer hires, not more layoffs
Hui, Reshef & Zhou
Organization Science (2024)		 Upwork
Before and after the ChatGPT launch		 Writing work, jobs -2% and earnings -5.2% / image work, -3.7% and -9.4% / ⚠️the drop was largest for the most experienced
Demirci, Hannane & Zhu
Management Science (January 2025)		 Online labor markets
8 months after ChatGPT		 Postings most open to automation (writing plus coding) -21% / image -17% / fewer postings mean tougher competition, while the jobs that remain are more complex and pay more
Upwork (an interested party)
Two official studies		 Its own platform In-Demand Skills 2026 (February) = demand for skills that name AI +109% (AI video +329% / AI integration +178% / AI data annotation +154%), and demand for non-AI skills held steady year over year / Future Workforce Index 2026 (July) = freelance share among skilled workers 28% to 38%, AI work bills +34% hourly (though it states explicitly that not every kind of AI work commands more)

⚠️ Upwork is an interested party. This is in-house research from a platform that benefits commercially when freelancing looks attractive, and it is not independent third-party measurement.

Note that this article carries only the figures I confirmed in the original text of each of Upwork's two reports. Roundup posts sometimes write that "Upwork says hourly rates are up 44%", but the February release says nothing whatsoever about an hourly-rate premium — what it does report is the growth in demand, plus a separate finding that nearly half of business leaders said they would pay a premium for independent talent who are creative and innovative.

3. Paradox 1 — for employees, experience is a shield. For freelancers, it is not

This is the most important finding in this article. Two studies say opposite things about years of experience.

　 Juniors Experienced
Employees
Stanford, through June 2026		 about 19% lower no comparable gap
Freelancers
Organization Science		 — the more experienced, the bigger the drop

A company hires a role; a client buys a deliverable. That difference is what is doing the work here.

The figure on the experienced side comes from this write-up of the Organization Science study. Employers keep experienced people because someone has to make the calls, train the juniors and carry responsibility when something breaks. None of that appears on the quote for a single deliverable. What the client is buying is this piece of work, not that person's working life.

Stanford's analysis names this structure cleanly. Employment falls in occupations where AI substitutes for tasks, and holds flat or rises where AI complements people — and the people benefiting from the latter are the experienced. They earn their living on tacit knowledge rather than codified knowledge.

The problem is that the freelance form of transaction has trouble putting tacit knowledge into a price. What becomes the product is whatever can be written into a proposal, itemized in a quote and signed off at acceptance — in other words, the codified part. So the same experience is a shield in employment and no shield in contract work.

📎 The employee side in detail is covered in Will AI Replace Veterans or Juniors First?. That article is about employment, and it uses an earlier version of Canaries (juniors -13%). The same research keeps being updated — Stanford itself writes that the gap widened from 15% as of July 2025 to 19% in June 2026, and the number will keep moving.

4. Paradox 2 — the work has not disappeared. The middle has

"AI cut the number of jobs" and "AI pushed rates up" are both correct. They are simply looking at different tiers.

The study published in Management Science shows that in the 8 months after ChatGPT launched, postings most open to automation, meaning writing and coding, fell 21%. The comparison is against manual work, so the figure isolates the AI effect.

The same study also writes this. Because there are fewer postings, competition among freelancers intensified, and on top of that, the automatable jobs that remain are more complex and pay more.

So "fewer" and "higher" are two sides of the same phenomenon. The easy work drained away, and the average of what is left went up. Individual rates did not rise; the low-paying work left the market — the two look alike in the numbers and mean completely different things.

Small, routine work

Fewer jobs. The band where the buyer can just hand it to AI themselves

Mid-sized work

The hardest place to be. Not simple enough for AI to replace, yet hard to justify paying a lot for

Large, complex work

This is what remains. Too complex to automate away, and the band that pays more

There is no market-wide answer to "has AI taken your work". The answer changes with which band you are in. And because the middle band holds the most people, the felt experience skews toward "it got harder".

5. What stopped affecting price

This is the most important research to come out in 2026. Siddiq and Zhang at UCLA Anderson followed 49,610 freelancers and 2.26 million contracts on Upwork from January 2021 to March 2026, and ran a difference-in-differences analysis around the ChatGPT launch (Human Capital, AI, and Labor Commoditization).

They turned the free-text part of the profile into numbers with text embeddings, and measured how much each part of a person actually mattered to the client's choice.

What was measured Change
[Full period] Weight on human-capital signals overall (credentials, work history, self-description) -7.8%
Same, weight on price +1.1%
Same, contract count -7.0%
[Most recent four quarters] Demand -9.6%
Same, weight on human-capital signals -10.1%
Same, weight on price +1.8%

The paper names what stopped working. Verified credentials, work history, portfolio, client ratings — every one of them lost its power to predict who wins the contract.

And the important part is that the decline is steeper the more recent you look. The weight on human-capital signals is -7.8% across the full period, and -10.1% in the most recent four quarters (April to June 2025 through January to March 2026). The weight on price widened from +1.1% to +1.8% as well. This is not a one-off shock, but an ongoing trend, and the paper itself writes that it is not confined to the period right after ChatGPT launched and has continued since.

💡 What this means: clients stopped paying the cost of judging whether a given person does good work. With AI at hand, a bad pick can be fixed by the buyer — or they have come to think that the quality gap is simply not as wide as it used to be. Either way, building a track record and being chosen on trust is a strategy that works less well than it did.

6. What started affecting price

Writing only about the losses would be unfair. The same body of data also shows what went up.

The work of using AI itself

In Upwork's July report, people who took on AI work bill 34% higher hourly rates. But the same sentence goes on to say that not every kind of AI work is gaining value — this is not a claim that anything touching generative AI pays

Complexity

In the study by Demirci and colleagues, the automatable jobs that remain are more complex and pay more. As the easy band drains out, the hard band takes up more of the mix

Non-AI skills have not vanished

Upwork's February report says demand held steady year over year and that "even as AI tools spread, businesses are continuing to hire talent at scale"

That last point is an important counterweight to this article's argument. "Track record stopped working" is not "the work is gone". Total demand is growing. What changed is who that demand gets allocated to, and at what price.

7. What is happening on the supply side — the people being pushed out

Everything so far has been the demand side. The supply side is moving faster.

According to Upwork's Future Workforce Index 2026 (July 2026), 38% of skilled US knowledge workers now work freelance. A year earlier it was 28%. That is a 10-point move in 12 months.

On top of that, 58% of full-time employees are considering going freelance — the same report explains this by saying that "more full-time employees are considering freelancing because working conditions remain in flux".

⚠️ This figure needs discounting. Upwork is a company that profits when the freelance market expands, and this is its own research. "Considering" is an intention rather than an action, and it does not mean people will actually make the move.

What makes it hard to ignore anyway is that Stanford's independent data corroborates the pressure on the pushing-out side — employment for young workers in AI-exposed occupations is 19% lower, and it is happening through tighter hiring rather than layoffs. When the front door narrows, people look for another way in.

Lay that over the demand-side finding and the picture sharpens. People are entering a market where track record and ratings have stopped mattering as much. New entrants have no track record, and that is less of a handicap than it used to be. Turn it around and the advantage held by people who spent years building one thins out by exactly that much.

8. Objections and caveats — what the researchers themselves say

An article that skips this part is one you should not trust. There are legitimate objections to the numbers above.

Objection 1: "it is the interest rates"

Economists at Google argued that the fall in young-worker employment can be explained by rising interest rates rather than AI. Stanford answered this in a dedicated rebuttal in February 2026 — occupations with high AI exposure are, on average, less sensitive to interest rates rather than more. Construction and similar sectors, the most rate-sensitive of all, have the lowest AI exposure. Even after splitting the data by rate sensitivity, the decline in high-AI-exposure occupations was larger than average.

Objection 2: "it is just what a low-hiring, low-firing market looks like"

Torsten Slok of Apollo Global Management asks "where is the AI jobs crisis" and argues that the trouble young workers have finding jobs is a sign that the labor market as a whole is in a state of not hiring but not cutting either. The fact that Stanford's data shows fewer hires rather than more layoffs is itself consistent with that reading.

The researchers' own caveats

Stanford attaches explicit caveats to its own results.

  • On causality, they "cannot yet answer definitively"
  • These are not causal estimates but "canaries in the coal mine", early descriptive indicators
  • Under tighter controls (firm-time fixed effects), the effect becomes significant only from 2024, and "some of the timing of this decline is driven by factors other than AI"
  • "We do not consider this paper, or any single study, to be definitive evidence of AI's labor market effects"

At the same time, lead author Brynjolfsson has gone back and re-tested the main objections one by one — excluding tech firms and computer occupations, controlling for rising interest rates and remote-work exposure, including firms that enter and leave the sample, and changing how AI exposure is measured, the divergence survives all of it.

So this article's position is this. The direction looks fairly reliable. The size, and how much of it is AI, can still move.

9. What happens next (3 forecasts and what would falsify them)

📌 Every forecast comes with a confidence label and a "this would prove it wrong" line. No dated numeric predictions, of the "down X% by year Y" kind — if the researchers themselves will not assert causality, I have no basis for asserting a number with a date on it.

✅ Confidence: high — being chosen on track record and ratings will count for even less

The basis is that this has been measured as a trend rather than a one-off shock. The shrinking weight on human-capital signals is -7.8% over the full period and -10.1% over the most recent four quarters, so it is accelerating. As long as AI output quality keeps improving, I can see no reason for this direction to reverse.

What would prove it wrong: the skill premium turns back upward / platforms institutionally raise the weight on verified track record and bring back decision criteria other than price.

🟡 Confidence: medium — rising supply pushes prices down further

The basis is two independent observations overlapping. The inflow behind a freelance share going from 28% to 38%, and the shift in which price gains weight and demand moves toward the cheaper side, are happening at the same time. More entrants plus weaker differentiation signals means stronger price competition.

Why I hold this at medium: the inflow figure comes from Upwork's own research, with no independent verification. The 58% who are "considering" it is intent, not behavior.

What would prove it wrong: median rates turn upward / demand grows faster than supply / the intent to switch never turns into actual switching.

🔴 Confidence: low — buyers do more in house, and outsourcing of "building" shrinks

I can describe the mechanism. If AI agents can carry the implementation, the client can absorb the "building" part they used to outsource. What is left is deciding what should be built and owning the result.

But confidence is low. I could not find data that measures this directly. Unlike the two above, this is inference, not observation.

What would prove it wrong: job counts and rates in the middle band recover / firms bring AI in house without reducing how much they outsource.

10. What to do — everything from here is speculation

⚠️ This chapter is what I thought after looking at the numbers above. It is not measured fact.

There is no research backing its effectiveness, so please read it as one available option. Conditions in this field change extremely fast, and the premises may well have shifted six months from now. We are nowhere near the point where anyone can say "do this and you will be fine."

In fact, if the forecast I labeled "confidence: low" in chapter 9 misses, part of this chapter misses with it. Apply it to your own situation and keep only what fits.

1. Shift toward selling the outcome rather than the deliverable (speculation)

Why I think so: what stopped affecting price was the codified part (credentials, work history, portfolio), and what remained in the high-value contracts was complexity. If that holds, then the closer the acceptance condition sits to "the metric moved" rather than "the deliverable exists", the harder your price is to compare.

Concretely: not "build one landing page" but "measure the signup rate and keep fixing it until it moves". But going to performance-based pay puts the collection risk on you, so that has to be handled as a separate problem.

2. Decouple pricing from time (speculation)

Why I think so: when AI shortens the work, an hourly rate creates the perverse incentive that the faster you get, the less you bill. The productivity gain goes straight to the client and none of it stays with you.

Concretely: a fixed price per project, or a monthly retainer. But a fixed price puts the loss on you when the estimate is wrong, so it does not suit work whose requirements move a lot.

3. Have reasons to be chosen other than your track record (speculation)

Why I think so: what Siddiq & Zhang measured is how people get chosen on a platform. Where profiles and prices are lined up side by side, track record lost weight. But if you are not standing in that lineup, the pressure does not reach you directly.

Concretely: referrals, ongoing relationships, being asked for by name in a specific field. This is not "you no longer have to sell", it is that the shape of selling changes, and it takes time to get going. It does nothing for someone who needs work right now.

4. Move to the side that uses AI (though the premium may shrink)

This is the only one with numbers behind it — AI-related work carries a premium (+34% on hourly rates in Upwork's July report). But the source is Upwork's own research, not independent third-party measurement.

And it needs care. The same report states explicitly that "not every kind of AI work is gaining value". It is not "anything with AI in the name gets a premium". On top of that, this is a premium for AI skills while they are still scarce, and with demand growing +109% year over year, supply will head there too — as entry increases, I would expect it to shrink.

Things that look like a bad idea

Answering with nothing but more track record

You would be putting more into a move that has been measured as working less well. Not wasted, but it will not get you there on its own

Cutting your price to win volume

Competing on price in a market where price already carries more weight puts you on the same ground as the cheaper new entrants

Summary

  • The question is not "will the work disappear" but "how did the way you get chosen change". In an analysis of 2.26 million contracts, track record, credentials and ratings lost weight (-7.8%) and price gained weight (+1.1%). In the most recent four quarters that gap is wider still
  • For employees experience is a shield, and for freelancers it is not, because a company hires a role and a client buys a deliverable
  • The work has not disappeared. The middle has. Postings most open to automation are down 21%, and the jobs that remain are more complex and pay more
  • The supply side is moving faster than the demand side. The freelance share went from 28% to 38% in a year (though this is the platform's own research)
  • The researchers themselves do not assert causality. Under tight controls the effect is significant only from 2024, and they state that some of the timing is driven by factors other than AI
  • The chapter on what to do is speculation. This field changes fast, and we are not at the point where anyone can say "do this and you will be fine"

FAQ

Q1. So will there end up being fewer IT freelancers?

The data points the other way. In Upwork's research, the freelance share of skilled US knowledge workers rose from 28% to 38% in a year. What is shrinking is not the headcount but the pricing power on any one job. That said, this is the platform's own research, so it needs discounting.

Q2. Is it true that the more experience you have, the worse off you are?

In the Organization Science study, the drop was largest among the most experienced. But that measured a short-term effect right after ChatGPT launched, and its subject matter was mainly writing and image work. I think the fair reading is not "experience is worthless" but "experience does not automatically protect your price".

Q3. Am I safe if I pick up AI skills?

For now there is a premium (+34% on hourly rates in Upwork's July report). But the same report also writes that not every kind of AI work is gaining value. And a premium that rests on scarcity normally shrinks as more people enter. Upwork's own demand data has skills that name AI growing +109% year over year, so supply is heading there too. "Currently advantageous" is a more accurate reading than "safe".

Q4. Is the same thing happening outside the US?

Every figure in this article comes from US and global platform and employment data, and none of it measures another national market directly. For individual national freelance markets, staffing agencies do publish their own numbers, but they have a customer-acquisition interest and the figures cannot be traced to a primary source, so this article does not use them. The structure, that a transaction selling deliverables makes it hard to put tacit knowledge into a price, should operate whatever the market, but the size and the timing can differ.

Q5. Why not do the "list of jobs AI will take" angle?

That angle is covered in other articles (Jobs AI will take / Jobs that survive the AI era). What this article covers is not occupations but the shape of the transaction. Within the same occupation, the effect shows up differently depending on whether you sell it as employment or as contract work — which is the subject of chapter 3.

Q6. What happens if the forecasts turn out wrong?

Chapter 9 lists what would prove each one wrong, so if that happens, please call it wrong. Two of them in particular are observable: "the skill premium turns back upward" and "median rates turn upward". It is worth assuming that this field moves fast and that the premises can change within six months.

Related articles