Skip to main content

AI Skills Grew 109%. The Top-10 List Barely Moved.

Updated 12 min readMarkdown
AI Skills Grew 109%. The Top-10 List Barely Moved.

Upwork’s sixth annual skills report is the best public data we have on what AI actually did to hiring in 2025. It’s also a clean lesson in how a growth number can point you the wrong way.


Upwork published its In-Demand Skills report on February 4, 2026. Most reports in this genre follow a script: announce that some capability is exploding, then tell you to go learn it. This one breaks character in the summary.

The most sought-after skills — full stack development, general virtual assistance, data analytics, and graphic design — have remained consistent year over year.

That line lands harder coming from Upwork than it would from almost anyone else. The company’s business rests on the premise that firms need new kinds of talent, so it has every incentive to sell an everything-changed story. Its own data tells a narrower one: the work changed shape, not substance.

I went through the five figures, six categories, and fifteen growth rates, plus the outside research the report leans on. Four findings hold up. Five things the report can’t tell you — and those may matter more.

The leaderboard didn’t move

The least newsworthy finding is the most important one. Across the six categories, most of the top 10 skills by earnings are the same names in roughly the same order as a year earlier. Companies are still paying people to run SEO, generate leads, cut corporate video, and design logos. They are still paying for data entry and spreadsheet analysis.

The report is pointed about the skills that were supposed to be first in line for automation:

  • “Vibe coding” has not reduced hiring for web design and development.
  • Generative image and video tools have not reduced demand for video editing, image retouching, or illustration.

The conclusion is blunt: in AI’s first real year of deployment, buyers paid for the same work.

That is not the same as saying nothing changed. It changed downstream. The same job now gets broken into more steps, and the human moves toward the end of that chain.

What the data measures, and what that rules out

The methodology runs two paragraphs. It’s the foundation for everything else, so it’s worth unpacking.

Dimension How it’s defined
Source Actual earnings of freelancers on Upwork — not survey responses, not job post counts
Window Jan 1 – Dec 31, 2025, compared with the same period in 2024
Buyer side US-originated demand only
Scope Six categories: coding & web dev, customer service & admin, data science & analytics, design & creative, sales & marketing, consulting
Eligibility A skill needs $100,000+ in category earnings over 12 months to be ranked
Cleaning Completed contracts only, which strips out price-shopping posts
“AI skill” tag Six skills that name AI explicitly: AI integration, AI chatbot development, AI video generation & editing, AI image generation & editing, AI data annotation & labeling, generative AI modeling

Using earnings instead of postings is the report’s best decision. It measures where money went rather than what people said they wanted. It also means every reading has to carry a qualifier:

Revenue = demand × price

That equation turns into a knife further down.

Growth lives in the prefix

This is the densest set of numbers in the report. Skills that explicitly name AI grew 109% year over year. Everything else grew 23%.

The ratio is the headline. The composition is the story. Every skill on the fast-growth list is an old job wearing a new prefix.

  • AI video generation and editing (+329%) is not a new profession. It’s editors shipping with AI.
  • AI integration (+178%) is not model training. It’s wiring a model into a client’s CRM and support stack.
  • AI data annotation (+154%) is the manual labor downstream of models.
  • AI chatbot development (+71%) is a support rebuild with a new label.

Actual AI building — generative AI modeling — grew 21%, and inside data science it ranked below knowledge representation.

Which punctures an assumption the industry has been carrying for two years: the shortage isn’t people who train models. It’s people who can put them inside an existing business.

Flag one detail. The third-fastest-growing coding skill was firmware development, at +14%. In a category supposedly being rewritten by AI, everything outside the two AI skills is growing in the low teens. The tail is thin.

Data science’s growth engine is labeling

The three fastest-growing data skills: AI data annotation (+154%), knowledge representation (+31%), generative AI modeling (+21%).

The most knowledge-heavy category is being pulled by its least knowledge-heavy task. Piecework, low barrier to entry, easy to staff at volume. Meanwhile modeling grows at 21%.

That’s a barbell. A small number of well-paid model and architecture people on one end, a large pool of low-cost labeling and cleaning labor on the other, and the middle — feature engineering, experiment design, classical ML — getting squeezed out of the frame.

For anyone planning a move, it’s an uncomfortable signal: the AI jobs that are easiest to get into are the ones most exposed to the next round of automation.

The human premium is real and unmeasured

These are the numbers that got quoted:

  • 77% of business leaders say AI is increasing their need for fractional talent with specific, specialized skills, rather than traditional full-time roles.
  • 47% would pay a premium for independent talent that is innovative.
  • 45% would pay a premium for talent that is creative.
  • 92% plan to expand their use of AI agents in the next year.
  • 53% describe the macro environment as challenging.

All from one survey: October 2025, 349 business leaders.

349 is a thin base for a narrative. At 95% confidence, the margin on a proportion near 50% is about ±5 points. The gap between 47% and 45% is two points. That isn’t a finding. It’s noise.

It also measures intent, not behavior. The question was “would you pay more,” not “did you pay more.” Nowhere in the report is there wage, rate, or contract-price data to check that 47% against.

The qualitative framing holds up better than the survey. The report’s useful idea is recomposition rather than replacement: the work doesn’t vanish, it gets reassembled. That concept earns its keep because it holds two true things at once — aggregate demand is stable, and specific people still get replaced.

Five things this report can’t tell you

1. Rising revenue can mean fewer people

The core metric is total freelancer earnings, which decomposes three ways:

Revenue = headcount × rate × volume

Turn 100 junior annotation contractors into 10 senior reviewers, triple the rate, and total revenue rises while 90 people lose work.

The dataset sees money, not people. So “recomposition, not replacement” is a claim this data can neither confirm nor refute. It might be right. This report isn’t the evidence for it.

2. The $100k floor tilts the growth list toward small skills

A skill qualifies for the fast-growth list at $100,000 in trailing-12-month earnings. So +329% could mean $100k became $429k — a handful of contracts.

Full stack development, the top skill in its category, is plausibly in the tens of millions. Ten percent growth on that base is a seven-figure gain. Three hundred and twenty-nine percent on a $100k base is $329,000.

The growth list and the volume list are different measurements. The report puts them on the same page, and readers will remember the percentage.

3. Keyword tagging undercounts AI, by construction

A skill counts as “AI” only if the name says AI. A contract titled “video editing” that was delivered entirely through AI tools lands in the 23% bucket.

So 109% is a floor, not a total. The report says as much: other categories “may contain an AI component but do not call it out directly.”

The same logic runs in reverse. The 23% for traditional skills is also inflated by AI. These two numbers aren’t a treatment and a control. They’re a comparison of two naming habits.

4. Twelve months can’t see the entry-level rung disappear

The report honestly cites the counter-evidence: more than half of global business leaders expect AI to eliminate significant numbers of jobs; entry-level roles are already declining; the WEF expects 39% of workers’ skills to be transformed or obsolete by 2030.

Its own data can’t test any of that, for the reason above. Counting earnings by skill category can’t separate a junior data-entry contractor from a senior one. They occupy the same cell.

Replace five junior roles with one senior plus AI, and category revenue can hold flat while five career entry points disappear.

Stable totals and a hollowed-out middle are entirely consistent with this dataset. The report reports the half it can see.

5. The publisher benefits from the conclusion

Upwork sells on-demand access to specialized freelancers. The report concludes that companies need more specialized fractional talent, that core work hasn’t gone away, and that buyers should keep hiring.

That doesn’t make the numbers wrong. It does mean the conclusion and the business point the same direction.

Credit where due: the report doesn’t chase the “AI replaces everyone” narrative that would pull more traffic. A report willing to say “less changed than you think” is usually worth more than one saying “everything did.”

Nine months later, the curve is bending

The report is from February. Upwork’s own monthly hiring data for July 2026 makes a useful correction:

  • AI video generation surged over the past year, then cooled month over month in July.
  • Video animation and social media design grew year over year and accelerated month over month in the same period.
  • Data annotation, up 154% on the annual measure, was down year over year in July, with a sharp monthly rebound — consistent with project-based training cycles.
  • Accuracy verification, error detection, and web testing were down both year over year and month over month.

That pattern supports the thesis sitting above it.

Tool-level demand spikes. Workflow-level demand stays.

The +329% in February was the sound of a tool being bought, not a job being redefined. What persists is the unglamorous part: someone has to turn what the model produced into something shippable.

The three declines are the part I’d watch. Accuracy verification, error detection, and web testing are quality gates. When they shrink while generation grows, teams are shipping model output as if it were finished work. That’s a cheap quarter and an expensive year.

What to do with this

If you freelance. Add AI to what you already sell instead of switching lanes. Every high-growth item on the list is an old skill with a new prefix. Rewriting “video editing” as “AI video ad production at volume” is the cheapest way to move from the 23% bucket into the 109% one. Just don’t bet the business on one tool. July already showed the spike fading.

If you’re employed. The sleeper line in the report is that talent recruitment and sourcing was the fastest-growing consulting skill. Companies are short on people, not tools. Inside an organization that translates to: whoever can decide who to hire and how to split work between people and agents is scarcer than whoever can operate the tools.

If you manage a team. 92% of leaders plan to expand agent use. Upwork’s own Human+Agent Productivity Index says completion rates improve by up to 70% when a human stays in the loop. Read together: agents are a capacity tool before they’re a headcount tool. Cut people with them and you keep the savings and lose the quality, which is what those three declining quality-gate skills look like at macro scale.

If you’re starting out. This report isn’t kind to you. The fastest-growing skills are low-barrier, easily substituted, and cyclical; annotation was already negative year over year by July. The durable edge sits where the report doesn’t measure: judgment, reliability, and the ability to turn a vague request into something a client can sign off on.

The one-line version

AI didn’t replace the work. It added a prefix to it.

“+109%” measures how popular the prefix became, not how unimportant the old work was.

What actually shifted is where human hours sit inside a job. From producing the thing to judging whether the thing is right.

The report’s real value may be that it makes the gap between revenue and headcount visible enough that you have to fill it in yourself.


Sources

Note: growth rates follow the report’s published methodology. The revenue-versus-headcount argument is my analysis, not a conclusion stated in the report.

5 views

No comments yet. Be the first to comment.