noworkleft.com

Home › Jobs › Proofreaders and Copy Markers

Office and Administrative Support · SOC 43-9081

Proofreaders and Copy Markers

Read transcript or proof type setup to detect and mark for correction any grammatical, typographical, or compositional errors. Excludes workers whose primary duty is editing copy. Includes proofreaders of braille.

Will AI replace this job?

The data cannot say, and anyone who gives you a percentage for “replaced” is guessing. What it does show is a gap between three things, for the 11 tasks O*NET lists for this occupation:

Could an LLM do it?

18%–91%

of tasks: an LLM alone could do the task or cut its time by half. Human raters say 18%, GPT-4 says 91%.

Counting tasks that need extra software around the model: 100%–100%.

Is it happening?

1 of 11 tasks

show measurable use in Claude conversations (Anthropic Economic Index). Anthropic’s occupation score, which counts only tasks that were also rated feasible, is 18%. Observed is a subset of “could” by construction.

Claude only, mostly US users, sampled periods in 2025. Not all AI use, and not employment.

What is left?

0 of 11 tasks

were rated “no LLM speed-up” by every rater (so, by Anthropic’s method, they cannot show up as observed use either). That is the part of the job these studies see no sign of AI touching.

It does not mean those tasks are safe. It means these methods found nothing there.

Task by task

Sorted with the tasks seen in use first. The three coloured squares are the three raters, in this order: human raters, GPT-4, GPT-4 with an alternative rubric. “Usage score” is Anthropic’s published task_penetration value; 0 means none seen.

LLM aloneneeds extra softwareno speed-up
Task (O*NET 31.0)Could an LLM?Claude usage score
Mark copy to indicate and correct errors in type, arrangement, grammar, punctuation, or spelling, using standard printers' marks. all raters 0.84
Read corrected copies or proofs to ensure that all corrections have been made. all raters none seen
Correct or record omissions, errors, or inconsistencies found. all raters none seen
Compare information or figures on one record against same data on other records, or with original copy, to detect errors. all raters none seen
Route proofs with marked corrections to authors, editors, typists, or typesetters for correction or reprinting. all raters none seen
Consult reference books or secure aid of readers to check references with rules of grammar and composition. all raters none seen
Consult with authors and editors regarding manuscript changes and suggestions. all raters none seen
Archive documents, conduct research, and read copy, using the internet and various computer programs. all raters none seen
Write original content, such as headlines, cutlines, captions, and cover copy. all raters none seen
Typeset and measure dimensions, spacing, and positioning of page elements, such as copy and illustrations, to verify conformance to specifications, using printer's ruler or layout software. (supplemental) all raters none seen
Read proof sheets aloud, calling out punctuation marks and spelling unusual words and proper names. (supplemental) all raters none seen
How to read this. These are averages over task lists written for the US labour market, rated by people and by GPT-4 using early-2023 model capabilities, next to a count of what one company’s chatbot users actually ask for. A task being “exposed” says nothing about whether employers will cut jobs, shift work, or hire more of you. See how to read an exposure score.

Try it with your own task list

Sources: task list from O*NET® Database (U.S. Department of Labor, CC BY 4.0; modified by noworkleft.com); “could” labels from Eloundou et al., GPTs are GPTs (OpenAI/University of Pennsylvania, MIT licence, labels from early-2023 models); “seen in use” from the Anthropic Economic Index (Claude usage only, CC BY). Dates, versions and limits.

Other jobs in Office and Administrative Support

Data built 2026-11-06. O*NET 31.0. Task match rates: 99.4% with the OpenAI/Penn labels, 99.2% with Anthropic usage.