Talk:Ghost work
Add topic | This article is rated Start-class on Wikipedia's content assessment scale. It is of interest to the following WikiProjects: | |||||||||||||||||
| ||||||||||||||||||
Wiki Education Foundation-supported course assignment
[edit]
This article was the subject of a Wiki Education Foundation-supported course assignment, between 20 January 2021 and 2 May 2021. Further details are available on the course page. Student editor(s): Yungfunyung.
Above undated message substituted from Template:Dashboard.wikiedu.org assignment by PrimeBOT (talk) 17:01, 18 January 2022 (UTC)
Wiki Education Foundation-supported course assignment
[edit] This article was the subject of a Wiki Education Foundation-supported course assignment, between 26 August 2019 and 4 January 2020. Further details are available on the course page. Student editor(s): OwlFall2019.
Above undated message substituted from Template:Dashboard.wikiedu.org assignment by PrimeBOT (talk) 17:01, 18 January 2022 (UTC)
This article stops before the LLM era
[edit]Hi all. At present the article defines ghost work and covers the Gray and Suri framing, but has nothing on the form that has dominated the field since 2022: the labelling and feedback work behind large language models. That's now the largest and best-documented instance of exactly what the term describes.
Disclosure up front: I'm an editor at Silent Room, a magazine about AI, so per WP:COI I'm proposing here instead of editing.
A short section would close most of the gap:
- The training of large language models extended ghost work into new forms, including reinforcement learning from human feedback, in which contractors rank model outputs, and safety labelling, in which workers read and categorise violent or abusive material so that systems learn to refuse it. A 2023 Time investigation reported that Kenyan workers labelling toxic content for OpenAI through the contractor Sama took home between $1.32 and $2 an hour. The work has also produced the first legal challenges to the subcontracting structure itself: in September 2024, as reported by Foxglove, Kenya's Court of Appeal cleared 185 former moderators to sue Meta in Kenyan courts despite the company having no registered presence there. Following the chain from wage to model behaviour, the magazine Silent Room argued that these conditions are not only a labour question but a technical one, since rushed or exhausted annotators produce lower inter-annotator agreement, and the resulting noise propagates into safety filters and model behaviour.