A writer who works for a tech company describes how he helps train AI models how to write, by making up pretend responses to hypothetical chatbot questions
Journalists and other writers are employed to improve the quality of chatbot replies. The irony of working for an industry … Mastodon: @michellemanafy@journa.host . X: @timkendall_path , @rosenzweigjane , and @martyswant LinkedIn: Mark Little , Elke Schwarz, PhD , Jason Bissell , and Peter Llewellyn Mastodon: Michelle Manafy / @michellemanafy@journa.host : “Indiscriminately learning from data produced by other models causes ‘model collapse’ - a degenerative process whereby, over time, models forget the true underlying data distribution.” Enter the professional writers, who are feeding AI “gold standard” responses... https://www.theguardian.com/ ... … X: Tim Kendall / @timkendall_path : A clear explanation of AI's voracious thirst for human created content and knowledge. That demand is generating work for human creators, and may never be slaked. Humans are always needed and not replaced. The same may be true of computational pathology. https://www.theguardian.com/ ... Jane Rosenzweig / @rosenzweigjane : “humans in the loop” has quickly moved from an assurance that humans will always play a key role in thinking/writing to euphemism for “humans are required to help AI companies make money” (this, from Guardian piece on writers who are being paid to improve LLMs, link below) [image] Marty Swant / @martyswant : Lots of great thoughts/lines in this @guardian story about the irony of writers improving LLM-powered chatbots. https://www.theguardian.com/ ... [image] LinkedIn: Mark Little : “It is like being paid to write in sand, to whisper secrets into a slab of butter. Even if our words could make a dent, we wouldn't ever be able to recognise it.” … Elke Schwarz, PhD : “Every machine is expansionistic, that is to say, imperialistic; each creates its own service- and colonial empire. … Jason Bissell : “secret sauce” behind these celebrated models relies on plain old human work. — Thought provoking article asking what is the relationship between human and Ai … Peter Llewellyn : In The Guardian today, here's something to think about #medcomms - maybe we'll be seeing some new jobs for #freelance #MedicalWriters 🤔 …
Context & Ripple Effects
Earlier coverage showed AI text generation reducing demand for some human-written work while also creating lower-paid editing work to make machine output sound human. This report identifies a more upstream role: writers are producing the reference-quality material used to shape chatbot behavior.
That reliance matters because the corpus frames indiscriminate reuse of model-generated data as a route to degradation. It also complicates the idea that writing tools simply replace authors: a prior account found a style-generating LLM’s output often felt hollow or approximate, while this work pays humans to supply the missing standard.
First-order effects
- Writers and journalists gain paid work creating hypothetical question-and-answer examples that companies can use to improve chatbot responses.
- AI companies become more dependent on curated human-authored responses rather than relying solely on material generated by models themselves.
Second-order effects
- The new training role extends the job shift already visible in copywriters’ lower-paid work humanizing AI text: writing labor is redistributed from commissioned output toward editing and training inputs.
- Demand for high-quality human material can raise the value of specialized contributors and quality-control workflows, even as broader AI publishing expands the volume of synthetic text.
Third-order effects
- If this pattern persists, model quality will increasingly depend on a managed supply chain of human expertise, making content provenance and evaluation a core operating capability rather than a one-time training task.
- The industry may split further between abundant generated text and scarce trusted human reference material, especially as AI-generated books and personalized articles spread across the internet and create more synthetic input to filter.
The trend: Generative AI is turning human-authored content from a finished product into an ongoing production input for model training, evaluation, and reliability.