OpenAI disclosed on 10 June 2021 that it had modified its language model training approach to address problematic outputs. The company stated it was now training models on a curated dataset to improve behaviour, acknowledging that prior versions produced content users found objectionable [source].

The announcement followed reports of models generating toxic, biased, or otherwise harmful text when prompted. OpenAI indicated the new training methodology involved human feedback to guide model responses toward more helpful and less harmful outputs. The company described this as "training on a curated dataset" rather than relying solely on internet-scale text corpora.

No specific failure examples were detailed in the disclosure. OpenAI framed the change as a proactive improvement rather than a response to particular incidents, though the timing suggested accumulated user feedback had prompted the shift. The company did not quantify how frequently problematic outputs occurred under the previous training regime.

The disclosure marked an early acknowledgement by a major provider that default training methods could produce unreliable or harmful model behaviour at scale. OpenAI stated the curated approach would be applied to future model releases, implying existing deployed models retained the earlier training methodology until updated.

The announcement provided no technical specifics on dataset composition, curation criteria, or how human feedback was weighted during training. OpenAI indicated the work was ongoing and that further refinements would follow as the company gathered more data on real-world model performance.

Users with concerns about specific outputs were directed to OpenAI's feedback channels, though no commitment was made regarding retrospective fixes to already-deployed models.