Is AI replacing writing jobs? What's shrinking and what's paying

No, AI isn’t replacing writing jobs across the board, but it is shrinking the cheapest, most generic part of the market and making entry-level work harder to get. Studies of freelance platforms found fewer writing gigs and lower earnings after ChatGPT launched. The Bureau of Labor Statistics expects writer employment to stay flat and says AI is likely to dampen demand. The freelance jobs that remain tend to be more complex and better paid, so your side hustle should be built around those.

What AI speeds up here

Task	AI helps?	Why
First drafts of routine copy (product blurbs, simple blog posts)	Yes	Buyers most often hand this work to a chatbot, and it’s where freelance demand has dropped
Research and outlining	Partly	Good for structure and a first pass, but it still makes up sources and facts
Editing and restructuring your own draft	Yes	It handles tightening, reordering and changing tone well
Fact-checking and original reporting	No	It can’t call a source, check a document or vouch for a claim
Matching a client’s voice and domain	Partly	It needs real examples from you, and it drifts back to generic phrasing
Pitching and keeping clients	Partly	It can draft the email but can’t build the relationship

What the research actually measured

Freelance marketplaces took the first hit

Researchers at WashU Olin and NYU studied Upwork data. After ChatGPT launched in November 2022, monthly jobs for writing-related freelancers fell 2% and monthly earnings fell 5.2%. Higher-skilled freelancers saw larger drops, though the authors call that finding “suggestive.” The study only covers short-run effects.

A second study, “Who Is AI Replacing?” (Demirci, Hannane and Zhu, Management Science, 2025), looked at a large global freelancing platform. About eight months after ChatGPT launched, postings for automation-prone writing and coding jobs had fallen 21% compared with manual-skill jobs. That figure covers writing and coding together, with no separate number for writing. Competition went up, and the jobs that remained tended to be more complex and better paid. You may see a “30%” figure credited to this study elsewhere. It doesn’t match the published abstract.

Traditional writing jobs are flat

The BLS counts 140,300 writer and author jobs, with median pay of $76,910 (May 2025). It projects little or no change from 2025 to 2035, with about 11,900 openings a year, and says the use of AI for writing “is projected to dampen demand.” Technical writers are projected to grow 1% over the same period, and BLS says AI tools may slow that growth. Editors are projected to decline 1%, but the BLS editors page doesn’t mention AI, so that decline can’t be pinned on AI.

Young workers feel it first

A Stanford Digital Economy Lab paper (August 2026 revision, ADP payroll data) found that employment for workers aged 22 to 25 in the most AI-exposed occupations is about 19% below where it would be if it had kept pace with less-exposed peers. Experienced workers show no similar gap. Most of the effect comes from less hiring rather than layoffs. Employment falls where AI automates the work and stays flat or rises where AI helps people do it. The authors say the results are descriptive, not causal, and the paper doesn’t single out writers.

Microsoft Research also measured how applicable AI is to different work activities, using Bing Copilot conversations. Fortune reported that writers and authors ranked fifth on its top-40 list. The paper measures applicability rather than job loss, and the lead author said it doesn’t mean AI can fully do any job.

Sort your own work by exposure

AI is mostly taking over tasks, and some jobs that were nothing but those tasks are disappearing. A rough way to sort what you sell:

Under pressure: high-volume SEO filler, basic product descriptions, “spin this article” rewrites, and entry-level roles where the main job was a clean first draft.
Holding up better: work that needs interviews, original data, regulated or technical accuracy, a strong brand voice, or a byline people trust.
Growing in importance: editing, fact-checking, and turning a rough AI draft into something a client can publish.

For each service, ask why a client would pay you if they could get it from a chatbot in 30 seconds. If you have no good answer, change what you sell.

The writing work AI created, and its fine print

Training and evaluating chatbots is a newer kind of writing work. We cover the platforms in more depth in AI writing jobs from home: the platforms, the pay, the catch. The short version:

DataAnnotation: its FAQ lists $25 to $50+ per hour for evaluating chatbot responses. The homepage quotes “up to” figures as high as $75 to $125+ an hour for experts. “Up to” is a ceiling, not typical pay. The Starter Assessment takes about an hour and you can only take it once. Pay goes out through PayPal.
Outlier (owned by Scale AI) doesn’t publish its rates. General onboarding takes 30 to 90 minutes, and pay starts once you’re working on a project.
Handshake AI is only open to US-based workers with work authorization. Its AI Evaluation Specialist role is listed at up to $17 an hour, and a Philosophy Expert role at up to $120, so pay follows expertise.
Alignerr’s site lists Prompt Writer roles at $40 to $120 an hour. That’s the company’s own claim.

This work isn’t stable. Scale AI cut about 500 contractors in July 2025. xAI laid off about 500 of its roughly 1,500 data annotators in September 2025 and moved toward specialist tutors. More than 200 contractors who rated Gemini output for Google through GlobalLogic were laid off. In November 2025, Mercor ended a Meta project and offered a similar one at $16 an hour, $5 less. In March 2026, Mercor disclosed a data breach, and Meta then paused its contracts with the company.

These platforms verify your ID, and the Mercor breach shows what can happen to that data. Apply only through the official site, and treat the income as a bonus rather than your rent.

Use AI the way the MIT study did, then add what it skipped

The best-known productivity result comes from an MIT experiment published in Science. 453 professionals did 20 to 30 minute writing tasks. The half given ChatGPT finished about 40% faster, and evaluators rated their work 18% higher. Weaker writers gained the most.

The tasks were simulated, participants had no client context, nobody fact-checked the output, and only ChatGPT-3.5 was tested. In real client work, writing prompts and fact-checking eat into the time you save. Here’s a workflow that allows for that (for which tools to use, see AI tools for freelance writers: which ones earn their keep):

Write the brief before the prompt: audience, goal, the three points that must land, and banned phrases.
Feed it real material, such as your interview notes, the client’s past posts and product specs. Thin input gets generic output.
Draft with AI, then rewrite the opening and closing yourself, because that’s where voice shows most.
Check every claim against a primary source before the piece goes to the client.
Read it aloud once. Clunky, chatbot-sounding phrases are easier to hear than to see.

Prompt you can copy

This prompt audits your current services and shows where you’re exposed:

prompt2 [placeholders]

You are a blunt freelance business advisor. Below are the last [5] jobs
from a freelance writer, with client type, task, word count and fee.

[PASTE JOBS]

For each job:
1. Label it AUTOMATE (a client could get an acceptable version from a
   chatbot alone) or AUGMENT (needs the writer's expertise, sources,
   interviews or judgment).
2. Explain the label in one sentence.
3. For every AUTOMATE job, suggest one way to turn it into AUGMENT work
   (e.g. add original quotes, data, editing, strategy) and a revised
   service description under 40 words.

Do not invent market rates or statistics. If you are unsure, say so.

Where AI falls over

Made-up facts and sources. Chatbots produce confident quotes, statistics and citations that don’t exist. Check everything before you sell or publish it.
Generic output. If the input is thin, the result reads like everyone else’s draft, and clients have learned to spot it.
Voice drift. After a few paragraphs the client’s style fades and generic phrasing creeps back, so edit long pieces in sections.
Ownership questions. The US Copyright Office’s January 2025 report on copyrightability concluded, as summarized by the Copyright Alliance, that prompts alone don’t give enough human control to make you the author. Human selection, arrangement and creative edits can still be protected, and that’s decided case by case. If a client expects to own the copyright, your editing has to be substantial.

Platform rules on AI content

Google’s guidance says generative AI can be useful for researching a topic and adding structure to original content. It also warns that generating many pages without adding value for users may violate its spam policy on scaled content abuse, and it suggests adding information on how content was created. Ecommerce sites must label AI-generated product titles and descriptions as AI-generated, so if you write for a store, your client has to follow that rule.

Freelance marketplaces and individual clients set their own rules on AI use, and those rules change. Read the platform’s current terms and the client’s contract, and ask the client if you’re unsure. Never present AI output as fully hand-written when the client’s terms forbid AI use.

The human-only part

The Demirci study’s finding that the remaining freelance jobs were more complex and better paid points to what clients still pay people for:

Interviewing a founder and finding the story they didn’t know they had.
Knowing a field well enough to spot the claim that’s subtly wrong.
Signing off on a piece and standing behind it.
Knowing what to cut.

AI makes drafts cheap, so put your rates on the judgment around them.

The money side: taxes and scams

In the US, freelance writing and AI-training income is self-employment income. For tax years after 2025, clients only have to send a 1099-NEC once they pay you $2,000 or more. Payment platforms only have to file a 1099-K if you receive more than $20,000 and more than 200 transactions. The income is taxable either way: the IRS says gig income must be reported even without a form. It goes on Schedule C, and self-employment tax (15.3%) applies once net earnings reach $400. If you expect to owe $1,000 or more, you may need to make quarterly estimated payments. This is general information, not tax advice, so check your own situation with a tax professional.

On scams, the FTC reports that job-scam losses grew from $90 million in 2020 to $501 million in 2024, and fake offers to pay you to rate or “like” things are a common pattern. The FTC’s rule is “Never pay anyone to get paid.” DataAnnotation and Mercor both say they charge no fees. Ignore unexpected recruiter texts and WhatsApp messages, and report scams at ReportFraud.ftc.gov.

Try it in 20 minutes

Open the free version of a chatbot such as ChatGPT or Claude (our comparison of ChatGPT, Claude and Gemini covers the differences, and best free AI tools for freelancers covers where free tiers stop). Paste in your last five gigs, or five you’d like to get, and run the audit prompt above. Pick one job it labels AUTOMATE and rewrite your offer to add one thing a chatbot can’t supply, such as a short client interview, an original data point or a fact-check pass. Then compare the old and new offers side by side. If the new one is clearly worth more, you’ve found the part of your work AI isn’t replacing.