Writing work that pays well and nobody talks about

Say you can write two thousand clear, accurate words in a working day. On a general content brief, that capability is worth less than it was three years ago, and the data is not ambiguous about it. Job posts for writing and coding fell 21% in the eight months after ChatGPT launched, relative to manual-intensive work, in Demirci, Hannane and Zhu's study of a large freelancing platform. Upwork's own Future Workforce Index 2026 shows generative-AI creative production with contract starts up 90% year over year and earnings per contract down 13%.

The same two thousand words are worth several times more in five other markets. That is the through-line here: identical capability, different buyer, different price. What moves the price is not how well you write. It is how close the document sits to money that is already at risk.

This assumes you can already write and already have clients. If you are starting from nothing, read starting as a freelance writer in a market that just repriced first.

Technical documentation is paid for out of the engineering budget

Software companies with an API, an SDK, or a self-serve product need reference docs, quickstarts, migration guides and release notes. The buyer is usually an engineering manager or a developer-relations lead, not a marketing team, and the budget comes from the engineering line.

The work is genuinely different from content writing. You read source code and pull requests, you run the thing you are documenting, and you commit your text to a Git repository through the same review process an engineer uses. Docs-as-code is the default toolchain, and being fluent in it is most of the barrier to entry.

It prices above general content because bad documentation shows up as support tickets and failed onboarding, both of which have a cost the buyer can see on a dashboard. Nobody has that number for a blog post.

Where it is posted: the Write the Docs job board, where companies pay $250 per posting. That fee is a filter — a company paying to post is hiring against a real requisition.

UX writing sits in a design budget, not a content budget

This is the text inside the product: button labels, empty states, error messages, onboarding flows, permission dialogs. The buyer is a product or design organization, and the money comes out of a design budget rather than a content budget, which is most of why it pays what it pays.

The craft that transfers is precision under constraint. What you have to add is working inside a design tool and a release cycle, and accepting that your output is measured by whether users finish a flow.

The UX Content Collective job board is the specialist route. It is "always free for candidates," and posting costs the employer, which filters the listings toward companies with a real role.

Grant and bid writing is attached to a specific number

Two different jobs, one shared feature: the document is attached to a sum of money that either arrives or doesn't.

Grant writing serves nonprofits, universities, research institutes and arts organizations applying to foundations and government programs. Bid writing serves companies responding to tenders, mostly public sector. In the UK, those opportunities are published openly on the government's Find a Tender service, which since February 2025 carries notices across "the full life cycle of contracts, from pipeline to termination." You can read the contract values before you decide whether the client can afford you.

Bid work has a professional body with published compensation data, rare in freelance writing. APMP's 2024–2025 US Compensation Report puts median total compensation for proposal professionals at $114,500 across 1,649 respondents, and at $131,500 for the 33 respondents categorized as proposal and business development consultants. Read the caveats first: it is a US survey, the consultant sample is tiny, and "all compensation published in this report is based on those who work 36 hours or more weekly," so it does not describe part-time independents at all.

Roles are posted openly on the APMP career center, which was carrying listings from defense, financial services and legal employers dated within the last week in August 2026.

Why it prices above general content: a bid team is competing for a contract worth six or seven figures, on a deadline that cannot move, and the cost of a weak submission is the whole thing. That asymmetry is the entire pricing argument, and you should make it out loud when you quote.

Investor and internal comms pay for discretion

Two adjacent markets, both allergic to hiring strangers.

Investor communications means shareholder letters, earnings scripts, board narratives and annual report copy. There is a statutory floor under some of it: under section 414C of the Companies Act 2006, a UK company's strategic report must contain "a fair review of the company's business, and a description of the principal risks and uncertainties facing the company," and that review must be "a balanced and comprehensive analysis" of performance and position. Someone has to write that every year, to a deadline, in language that survives legal review.

Internal communications is the other one: restructures, system migrations, acquisitions, return-to-office policies, benefits changes. The buyer is an HR or communications director who needs several thousand people to read something once and not misunderstand it.

Both price above general content for the same reason. You are inside confidential information before the market or the workforce sees it, and a mistake there becomes a disclosure problem or a leak. Freelancers who can be trusted with that are scarcer than freelancers who can write.

The route in is almost never a job board. It is a finance or comms leader who has already seen you handle something sensitive: slow to enter, sticky once you are in.

AI data work: annotation, evaluation, and writing the source material

This is the fastest-growing category and the worst-explained. Upwork's In-Demand Skills 2026 put AI Data Annotation and Labeling up 154% year over year in US marketplace earnings, with all AI-referencing skills up 109%. The Future Workforce Index found freelancers doing AI work earn 34% more per hour. Both come from an interested party and neither publishes what shrank, so treat them as evidence of direction, not size.

What the work actually is

Three things get bundled under one label.

Annotation and labeling is structured judgment applied at volume: tagging content, marking whether an output followed an instruction, classifying intent.

Model evaluation is comparative. You are shown two or more model responses and you rank them, or you score one against a rubric. Outlier, which is "operated by Scale AI," describes the three core tasks on its own site as writing a challenging prompt with the correct answer, building the grading rubric, and rating and ranking answers.

Producing source material is the writing job proper. You create the problems, the reference answers, the adversarial cases and the domain scenarios that a model is trained or tested against. This is where a writer with real subject-matter depth is worth the most, because the scarce input is the expertise, not the prose.

Who the buyers are and how it reaches you

You will almost never contract with an AI lab directly. The labs buy through data vendors, and the vendors contract you.

Mercor describes itself as "organizing human intelligence to power the AI economy," serving "frontier research, RLHF data, and AI agent training at scale for the top AI labs and enterprises." Surge AI recruits an expert contributor bench directly by email, listing roles for lawyers, journalists, consultants, professors and investors, with tasks such as reviewing model arguments for logical consistency and drafting exemplar analyses. Prolific sits slightly to the side, supplying human data to "AI developers, researchers, and organizations."

The practical shape of the engagement: you apply to a role rather than bid on a project, you pass a qualification task, you sign an NDA, you work hourly inside the vendor's own portal, and the project can end without notice when the lab's contract for that data ends. It is contract work with the texture of shift work, and the vendor is your client, not the lab whose model you are improving.

What it pays, honestly

The range inside this one category is wider than the range across every other market in this post.

At the top, Surge AI's workforce page lists its expert contributor roles at "$200–400 / hour, contractor." In the middle, Mercor's open roles in August 2026 included computer science PhD researchers at $70–$90 an hour and a CUDA engineering expert at $80–$100, alongside a household activity video contributor at $10–$15. At the bottom, Prolific's own payment principles set a floor: "Our minimum rate is £6 / $8 per hour. Everything beyond that is your call. We recommend paying at least £9 / $12 per hour."

An order of magnitude and a half separates the ends of that spread, and the label on the work is identical. What moves you up it is how hard your judgment is to source. A generalist rating chatbot responses competes with everyone. A tax specialist writing evaluation cases for accounting reasoning competes with almost nobody, and the same goes for a physician or a litigator.

Five questions before you accept

Send these back to the recruiter. They take two minutes and they decide whether the work is worth your day.

1. Is the rate hourly, per task, or quality-adjusted?
2. Is qualification and calibration time paid?
3. What is the payment cadence, and through which processor?
4. Is there a minimum weekly commitment, and a maximum?
5. Can I describe this work to other clients, at all?

Question one matters because "quality-adjusted" means your effective rate is set after the fact by someone else's scoring. Question two catches the most common way this work underpays: several unpaid hours of onboarding tests attached to a contract that ends in three weeks. Question three is your time-to-cash — Outlier advertises "quality-based rewards, paid weekly," which is faster than most freelance invoicing and worth something real. Question four tells you whether this can coexist with your other clients or quietly becomes a job. Question five decides whether the work builds anything you can sell later, and the honest answer is often no.

Then measure it. Log the qualification time as well as the paid time and run the project through the effective hourly rate calculator after a month. Unpaid onboarding is what turns a headline $40 an hour into an actual $26, on illustrative numbers you should replace with your own.

One admin consequence people miss: these are US contractor relationships, and the 1099-NEC reporting threshold rose from $600 to $2,000 for payments made after 31 December 2025. Several of these vendors will now pay you a four-figure sum with no form attached, and it is all still taxable. Your own records become the only record, which is what the threshold change actually means for freelancers.

Vendor portals also track your hours for their payroll, not for your books. If you are running AI data work alongside client projects, you need your own timesheet on top of theirs — that is the same argument as tracked hours turning into paid invoices, and it is why time tracking that is billable by default matters more once your income comes from six places instead of two.

What transfers, and what you have to buy

Writing ability transfers into all five. Nothing else does automatically.

Technical documentation needs Git and the patience to read code. UX writing needs a design tool and a tolerance for review. Bid writing needs compliance discipline, because a non-compliant bid is scored zero regardless of how it reads. Investor and internal comms need someone to vouch for your discretion. AI data work needs a demonstrable domain, and increasingly a credential in it.

Each is weeks of effort, not years. That gap between weeks of effort and a rate multiple is why these markets stay under-supplied while general content stays crowded.

The next 30 minutes, and the next month

In the next 30 minutes: pick the one of these five where you already have unfair knowledge (a previous career, a degree, an industry you worked in for six years) and open the matching source. Write the Docs if you came from software, the APMP career center if you came from a company that bids, Mercor or Surge if you have a licensed profession behind you. Read five live listings and write down the vocabulary they use that you don't currently have on your site.

This month: produce one sample in that market and get one paid engagement of any size, however small. A single 800-word API quickstart, one rewritten onboarding flow, one section of a bid. Priced work you have actually delivered is the only credential any of these buyers reads, and one of them beats a portfolio of general content.

In Worklyn every project ends with a real margin instead of a guess, so after a quarter of mixed work CFO Mode can tell you which of these five is worth another year of your attention.