OpenAI fired contractors for using AI to label data

Company OpenAI terminated contracts with a group of employees who used third-party AI tools to perform tasks related to training and labeling models. The incident occurred within a program aimed at improving the quality of neural network responses, where people were supposed to manually check and correct generated content.

According to internal reports, contractors applied automated systems to generate responses, which they then passed off as the results of their own work. Using AI to perform labeling tasks violates the company's security protocols, as it leads to the inclusion of synthetic data in training samples, which can trigger a "model collapse" effect.

In discussions, experts note that such actions reduce the accuracy of RLHF (reinforcement learning from human feedback). When a model is trained on data created by another model, it begins to inherit the errors and hallucinations of its predecessor, making the final result less reliable for end-users.

Currently, OpenAI has strengthened control over the data verification process. The company requires that all training stages related to quality assessment be carried out exclusively with human participation, in order to exclude the risk of degradation of future model versions due to "poisoning" of the training dataset with automated content.