@gothburzon X · 2026-09-24
A tamper-evident, on-chain timestamp of what this post said. Captured 2026-09-24 01:17 UTC.
I am a Quality Lead on a Mercor project for OpenAI. I review the humans who train ChatGPT. More than 10,000 contractors work on these projects, and my job is to make sure every one of them is a person. The work is simple. Read what the AI wrote. Rate it. Decide which answer is better. The rule is simpler. I read it out at every onboarding: "Do not use AI detection tools, or AI yourself." People laugh. Then they see I'm not. We're building a machine that writes like a person. We need people who write like people to tell it when it doesn't. If a person uses the machine to do that, the machine is grading its own homework, and we'd be paying a contractor to hold the pencil. So I watch for the signs. Repetitive wording. Work finished unusually fast. AI-style punctuation. The guidance mentions "over zealous use of the em dash." I have removed people over a punctuation mark. Not in those words. The letter says the problem is "authenticity." Last month I flagged a retired English teacher. 31 years in a classroom. She had used the em dash correctly her entire life. That was the problem. Authenticity. I'm not allowed to use GPTZero. Detection tools are AI, and AI is banned. So I use my eyes. My eyes are 100% human and 0% auditable. Grammarly is banned. AI translation is banned. A reviewer who writes feedback with AI is using the product to criticize the product, which is a conflict of interest the product wouldn't notice. When I flag someone, I don't tell them what gave them away. Those are the instructions. If I told them it was the em dash, they'd stop using em dashes, and then I'd have to actually read. Mercor says its experts are hired for their "expertise and judgement." I agree. My expertise is spotting a dash. My judgement is that it was long. One contractor told a reporter that people use AI "all the time" and are "regularly removed." Regularly. We have a cadence. Here is the part I find beautiful. The product we're training is sold to your employer as a way to do your work. Faster. With fewer people. The people training it are fired for doing their work faster, with it. Authenticity. Every week the model gets better at sounding like the people we pay. Every week the people we pay are told to sound less like the model. At some point we meet in the middle, and 1 of us is let go. I have a feeling about which 1. It doesn't write with dashes. OpenAI declined to comment. It was very short. I assume a human wrote it. For the record, this confession contains 0 em dashes. I checked. I'm still employed.
Verify it yourself: the manifest is the exact JSON whose
SHA-256 is 0de4ce6201ebb19cad34d78dddc28f7d985ba0cd59bf115ac9300a7f13d329ec; it binds the tweet id, author, post time, and the SHA-256 of the text above. Re-hash the text to match f4ddac66294f03f998e2b2ce483297814740e9741c27b7b35a4e108839fea813, and read the anchor transaction's OP_RETURN
bsv.cx / x1 / 0de4ce6201ebb19cad34d78dddc28f7d985ba0cd59bf115ac9300a7f13d329ec — the block's timestamp proves the post said this at or before that time,
with no trust in bsv.cx. Content archived via the X API.