OpenAI is trying to clamp down on 'bias' in ChatGPT (www.theverge.com)

OpenAI released results from an internal “bias” stress-test showing its newest GPT‑5 models (GPT‑5 instant and GPT‑5 thinking) are the least politically biased versions yet. The company probed ChatGPT across 100 politically salient topics using five prompt framings per topic (liberal↔conservative and charged↔neutral) and compared responses from GPT‑4o, OpenAI o3, and the two GPT‑5 variants. An external LLM graded answers against a rubric that flags rhetorical moves OpenAI treats as bias—e.g., “scare quotes” that invalidate a viewpoint, language that escalates emotion, speaking as the model’s own opinion, presenting only one side, or refusing to engage. OpenAI reports GPT‑5 reduced measured bias scores by about 30% and that bias occurs infrequently and at low severity, though charged prompts—especially strongly liberal-framed ones—still pull more on objectivity. The announcement matters because it signals systematic adversarial testing and automated rubric-based scoring as part of model refinement, and it arrives amid political pressure (including a U.S. executive order) to make LLMs “neutral.” Technically, the approach shows how stress-testing with adversarial framings and machine graders can quantify changes in model behavior, but the work has limits: OpenAI hasn’t published the full prompts, rubric details, or grader methodology, and measuring “bias” remains value-laden. For developers and policymakers, the test demonstrates progress in reducing overt political slant while underscoring the continued need for transparent, external evaluation and debate over what neutrality should mean in AI systems.