OpenAI fires contractors who used AI to help train ChatGPT

By Splitfeed EditorialLast updated Sep 26, 2026, 20:18 UTC

On September 22, 2026, reports emerged that OpenAI had terminated a group of contractors whose job was to produce training data for ChatGPT. The contractors had been caught using AI tools to complete that work, in direct violation of internal rules that explicitly prohibited AI assistance on annotation tasks. OpenAI's own monitoring systems identified the violations and flagged the individuals responsible. The story broke first through 404 Media and spread widely over the following day.

The dismissals expose a structural tension at the heart of how large language models are built. Human annotation, where contractors read, judge, and label outputs, is considered essential for producing the kind of training signal that shapes model behavior. When contractors substitute AI output for their own judgment, the training data is no longer grounded in independent human assessment. The concern that follows from this is sometimes called model collapse, a phenomenon where a model trained on AI-generated data begins to degrade, losing the diversity and reliability that human-sourced data provides. What is genuinely in dispute is how widespread the practice was, how much contamination resulted, and whether current monitoring is robust enough to catch it consistently.

The anti-AI camp has responded with a mixture of outrage and dark humor. Commenters including Philip Athans, Amy, and Metin Seven expressed frustration and resignation, while Marcus Summers pointedly questioned the logic of prohibiting workers from using the very tools their employer develops and promotes. Jay argued the episode proves AI output is low quality by definition, and LP suggested the outcome was entirely predictable. At least one contractor reportedly sabotaged model training deliberately, a detail critics are using to argue that the labor conditions surrounding AI development are unsustainable. The pro-AI camp has not yet published reactions, though a story like this would typically prompt arguments that the firings show responsible governance and that quality controls are working as intended. The middle ground has focused on the technical stakes rather than the labor angle, with Kevin Foster questioning whether this story undermines claims about recursive self-improvement, and Paul Mainwood identifying model collapse as the most consequential issue the episode raises.

The clearest thing to watch going forward is whether OpenAI or independent researchers publish any assessment of how much training data was affected and whether measurable degradation in model quality can be traced to the contaminated annotations. Any disclosure on the scale of the problem, or changes to how contractors are monitored and compensated, would significantly shape the next round of this argument.