x.com/jxmnop/status/2050437965168652344

OpenAI’s own post describing the investigation says the team did much more than simply query models and count the word “goblin.”

According to OpenAI:

• they found creature language was especially common among users who had selected the “Nerdy” personality;
• they measured that although “Nerdy” accounted for only 2.5% of responses, it produced 66.7% of all “goblin” mentions;
• Codex compared RL-training outputs that contained “goblin”/“gremlin” against outputs for the same tasks that did not;
• they identified a specific reward signal for the Nerdy personality that scored creature-word outputs higher;
• they tracked mention rates over training with and without the Nerdy prompt to test transfer; and
• they searched GPT‑5.5’s SFT data and found many datapoints containing these creature words.

So while counting mentions was one part of the analysis, OpenAI says the root-cause investigation also involved traffic segmentation, reward-model auditing, training-time comparisons, and dataset inspection. That makes the claim that they "just talked to various models and counted" misleading.