OpenAI contractor claims human trainers fired for AI-like writing
Is this a scandal?
No longer — the story has resolved. Noise 37/100, holding steady, across 1 source.
Expect increased scrutiny of RLHF vendor contracts and potential unionization efforts among data labelers because subjective quality metrics create untenable employment conditions as model outputs converge with human writing styles.
Noise 37/100 — louder than 97% of tracked AI controversies.
Why it matters
Highlights a paradox where AI training labor is penalized for efficiency while models converge on human style, raising questions about sustainable data labeling practices.
Key points
- Mercor quality lead alleges contractors are terminated for 'authenticity' failures including correct em-dash usage.
- AI detection tools like GPTZero are reportedly banned to avoid circular validation conflicts.
- Contractors receive no specific feedback on flags to prevent them from gaming authenticity assessments.
- Over 10,000 contractors work on OpenAI projects under these subjective human-only review standards.
- Workers face a paradox where using AI for efficiency violates the core requirement of human-generated training data.
- OpenAI declined to comment on allegations regarding contractor terminations and evaluation criteria.
The story
A Mercor quality lead working on OpenAI projects alleges that over 10,000 contractors face termination for producing text deemed insufficiently authentic, including correct em-dash usage. The reviewer, identified as gothburz, claims detection tools like GPTZero are banned to prevent circular validation, forcing reliance on subjective human assessment of punctuation and cadence. Contractors reportedly receive no specific feedback when flagged to prevent gaming the system, creating an opaque evaluation standard centered on distinguishing human nuance from model output. The account describes a retired English teacher being removed specifically for grammatical precision that resembled AI generation. This testimony highlights tensions in RLHF workflows where workers are allegedly penalized for efficiency gains derived from the very technology they train. OpenAI declined to comment on these specific allegations regarding contractor management and authenticity standards. Mercor has not publicly addressed the claims concerning its expert vetting processes or termination criteria.
Who's involved
Claims current authenticity enforcement is subjective, punitive, and creates an unsustainable paradox for human trainers.
Positions its platform as hiring experts for judgment and expertise, though has not addressed specific termination claims.
Declined to comment on specific allegations regarding contractor management and authenticity standards.
Noise Level
The timeline
gothburz publishes detailed account
Twitter post outlines subjective enforcement mechanisms and bans on AI detection tools in OpenAI training workflows.
OpenAI declines comment
Company provides brief non-response to inquiries regarding the allegations.
Retired English teacher allegedly flagged
Quality lead reports removing a 31-year veteran educator for correct em-dash usage deemed inauthentic.
The full record
Sources & methodology
- twitter.com — twitter.com
Every claim above traces to these primary items. How we score →
The forecast
Expect increased scrutiny of RLHF vendor contracts and potential unionization efforts among data labelers because subjective quality metrics create untenable employment conditions as model outputs converge with human writing styles.
Forecast, not fact — an editorial estimate we score when this resolves.
That's the complete picture as of — nothing more to know right now. We'll update this page the moment it changes.
Join the Discussion
Discuss this story
Community comments coming in a future update
Be the first to share your perspective. Subscribe to comment.