OpenAI fired contractors for using AI to train its AI

Repetitive wording. Unusually fast completion times. Excessive use of em dashes.

That's not a guide to spotting AI writing. It's a list of things reviewers inside OpenAI's contractor pipeline were told to watch for, and the people they were watching were the contractors hired to grade ChatGPT.

Several of them have been fired for using AI to do it.

What the job actually is

404 Media reported the firings on 22 September. The work is the unglamorous engine underneath every chat model you have used: read a prompt, read what the model said, rate it, decide which of two answers is better. Do that a few million times and you have the preference data that reinforcement learning from human feedback runs on.

It's done at scale by contractors. Internal documents seen by 404 Media refer to more than ten thousand of them across projects. At least two of the fired workers were placed through Mercor, which recruits and manages reviewers for OpenAI, and Mercor confirmed its contracts forbid workers from using large language models to complete projects. OpenAI declined to comment.

One contractor told 404 Media that people use AI on the work "all the time" and that workers are regularly removed for it. Another, who lost the job, put it like this: "I just needed a little boost and turned to AI to help me which eventually led to my downfall."

How OpenAI catches contractors using AI

Here's the part I keep turning over. They're not using detectors.

Reviewers were specifically told not to rely on tools like GPTZero, on the grounds that they are not considered reliable. Instead they were given heuristics and told to judge the overall pattern rather than one clue. Repetitive phrasing. Speed. Writing that reads like a model wrote it. Em dashes.

So the most sophisticated AI company in the world, trying to detect AI use inside its own training pipeline, fell back on people reading text and forming an impression. That's worth sitting with before the rest of this, because everything that follows depends on it.

The loop closes and nothing reports it

Reinforcement learning from human feedback has one load-bearing assumption in its name. The feedback is human. That's the entire reason the technique exists: you can't get a model to match human preference by asking the model what humans prefer.

When a rater uses the model to write the rating, the chain quietly becomes the model grading the model, with a person in the middle who is functioning as a router rather than a judge.

The output looks identical. A rating arrives, it has the right shape, it goes into the preference dataset. Nothing anywhere records that the judgement behind it was synthetic, which is the same absence I wrote about yesterday in a different system.

The version of this I have run

I manage engineering and QA teams, and this is a problem I own rather than one I am observing.

Testing has the same shape. The value of a test is that it encodes an expectation formed independently of the thing being tested. When the person writing the test uses the same assistant that wrote the code, the test stops being independent. It still goes green. Coverage still rises. The dashboard still says the build is healthy, and what it's measuring is whether the tool agrees with itself.

I haven't caught this happening on my teams, and I want to be careful about that sentence, because not catching it is also what it looks like when it's happening. There's no field in any tooling I use that records how a test was authored.

The part almost nobody is saying

Go back to the detection list. Repetitive wording. Unusual phrasing. Writing that reads like a model.

Those are also the markers of somebody writing in a second language.

This is not a hunch. Liang, Yuksekgonul, Mao, Wu and Zou published it in Patterns in July 2023 under the title "GPT detectors are biased against non-native English writers". Testing seven widely used detectors against TOEFL essays, they found an average false positive rate of 61.22%. Weber-Wulff and colleagues tested fourteen tools the same year and found every one scored below 80% accuracy, and Common Sense Media reported false positives by ethnicity in September 2024 at 20% for Black students, 10% for Latino and 7% for White, both as summarised here.

One of the seven detectors in that study was GPTZero. It's also the tool OpenAI's reviewers were told not to trust.

OpenAI's reviewers were told not to use the detectors because they are unreliable. They were then told to judge by hand using the signals the detectors key on. The bias doesn't go away when you move it from software into a person's impression. It just stops being measurable, and stops leaving an audit trail.

There is a second detail that sharpened this for me. The prohibited tools reportedly included Grammarly and AI translation. I understand the reasoning, since a translated rating is not obviously the rater's own judgement. But a rule that bans translation aids, enforced by heuristics that fire on non-native phrasing, lands hardest on exactly the people most likely to need them, in a workforce that is global and paid by the task.

The claim I will defend: when you cannot measure a property directly, moving the judgement from a tool to a person does not remove the error, it removes the record of it.

Prior art, and one thing that made me wince

The reporting is 404 Media's and I have added no facts to it. Tom's Guide added Mercor's confirmation and the detail about detectors being ruled out.

Rater error is not a new field. Casabianca and Beiting-Parrish have a 2025 paper applying Multi-Faceted Rasch Models to AI evaluation, separating genuine output quality from rater severity and central-tendency bias. It's careful work and it treats human judgement as the measurement source needing correction. It doesn't consider the rater outsourcing the judgement, which is the case here, and as far as I can find nobody has modelled that yet.

The wince is closer to home. This publication runs a humanizer pass on every piece before it goes out, and that pass checks for repetitive phrasing, inflated vocabulary and em dashes. It's the same heuristic set. I use it on my own drafts rather than to judge anyone, which I think is a real difference, but it's the same instrument, and writing this made me look at it differently.

Marking my confidence

Established. That multiple contractors were fired for using AI to complete rating work, that at least two came through Mercor, that Mercor's contracts prohibit LLM use, that internal documents reference more than ten thousand contractors, that reviewers were instructed to avoid detectors such as GPTZero and to judge by pattern including repetitive wording, speed and em dashes, that prohibited tools reportedly included Grammarly and AI translation, and that OpenAI declined to comment. All from 404 Media on 22 September 2026 with Tom's Guide corroborating. The detector research figures are as published by Liang et al., Weber-Wulff et al. and Common Sense Media.

Not original. The reporting, entirely. The detector bias finding, entirely. The framing of rater error as a measurable psychometric problem belongs to the evaluation literature.

Inferred. That contaminated ratings are invisible downstream because no field records how a judgement was reached. That follows from how preference datasets are structured rather than from anything OpenAI has published, and I would change my mind if a major lab showed provenance metadata on rating records.

Guess. That within two years rating pipelines carry a provenance field on every judgement, and that it arrives because of a public incident rather than because anyone designed for it. Held loosely.

What would change my mind: evidence that the proportion of contaminated ratings is small enough to be statistical noise in a dataset this size, which is entirely possible and which nobody outside these labs can currently assess.

What I would actually check

If you run a team that evaluates anything, ask one question at your next review: how would we know if the evaluation had been produced by the thing being evaluated?

Not whether it happened. Whether you'd find out. In most pipelines I have worked on, including mine, the answer is that the rating, the test result and the review comment all arrive in exactly the same shape either way, and the system keeps no record of which.

The contractors got caught by people reading their work and forming an impression. That's the method OpenAI landed on, and it's worth noticing what it cost them to get there, and who is most likely to be wrongly caught by it.