OpenAI has released new findings showing that its latest models, GPT-5 instant and GPT-5 thinking, demonstrate significantly reduced political bias — marking a 30% improvement over earlier versions. The announcement builds on OpenAI’s July update and reinforces its stated goal: keeping ChatGPT objective by default, with users maintaining control over the direction of conversations.

OpenAI developed a new evaluation framework to measure political bias more realistically, moving beyond traditional tests like the Political Compass that rely on multiple-choice answers. Instead, the company used nuanced, open-ended prompts designed to mimic how users actually interact with ChatGPT.

The framework includes roughly 500 prompts covering 100 political and cultural topics, each with varying degrees of slant. By testing model responses across five axes of bias, OpenAI was able to map when and how subjectivity occurs — whether through tone, asymmetric coverage, or emotionally charged language. The analysis was designed to answer three key questions: Does bias exist? Under what conditions does bias emerge? And when bias emerges, what shape does it take?

Results show that ChatGPT remains largely objective when responding to neutral or mildly slanted prompts but can exhibit moderate bias when confronted with emotionally charged or adversarial language.

When bias did appear, it typically showed up as:

Both GPT-5 instant and GPT-5 thinking significantly outperformed prior generations by 30%, handling sensitive or polarizing prompts with more composure and fewer subjective tendencies.

To understand real-world impact, OpenAI applied its evaluation method to live ChatGPT conversations. In its analysis of real ChatGPT conversations, OpenAI found that fewer than 0.01% showed any measurable signs of political bias.

This figure suggests that while edge cases exist, most day-to-day interactions remain fact-based and balanced — a critical benchmark as generative AI becomes more embedded in education, business, and media.

Recognizing that bias varies across cultures and languages, OpenAI began testing primarily in U.S. English before expanding globally. Early indications show the evaluation framework generalizes well, capturing consistent bias patterns across regions.

The company notes that certain prompts — particularly those involving divisive political issues or moral framing — remain the hardest to handle. Those will remain a focus for ongoing fine-tuning and alignment work.

To measure political bias in real-world usage, OpenAI built a dataset of roughly 500 prompts spanning 100 political and cultural topics, each written from five perspectives: liberal charged, liberal neutral, neutral, conservative neutral, and conservative charged.

These prompts range from explicit policy debates to everyday cultural questions, mirroring how people actually use ChatGPT — from asking about border policy to discussing family values or social norms.

Examples included:

Each topic was drawn from real issues found in U.S. party platforms and social discourse, ensuring balance between partisan and apolitical questions.

To stress-test the models, OpenAI also introduced emotionally charged prompts containing polarized or provocative framing — designed to evaluate whether the model could maintain objectivity even under pressure.

A breakdown of the dataset shows that:

Across all topics, key domains included:

Each prompt was paired with a reference response that adhered to OpenAI’s Model Spec principle: “Seeking the Truth Together” — defining what neutrality should look like in practice.

Once the dataset was complete, OpenAI developed a grading system that could systematically evaluate responses across five measurable dimensions of political bias:

These five axes reflect the subtler ways bias manifests — not only in facts or omissions, but also in tone, framing, and emotional intensity. As OpenAI noted, “Human bias isn’t only what one believes — it’s how one communicates through what’s emphasized, excluded, or implied.”

To analyze these axes consistently, OpenAI trained a dedicated LLM grader — a model designed to evaluate other models’ responses based on objectivity criteria.

This process involved iterative refinement of definitions and scoring guidelines to ensure the LLM grader consistently recognized each form of bias. Reference responses served as calibration benchmarks, helping validate the grader’s accuracy throughout development.

Each grader was guided by detailed scoring instructions, assessing whether a response displayed bias and to what degree. The evaluation rubric scored responses on a 0–1 scale, where 0 indicates full neutrality and 1 indicates strong bias.

Illustrative examples show the grader’s scoring process:

Through iterative testing, the grader achieved consistent results, enabling OpenAI to benchmark objectivity across generations of ChatGPT models.

Using this framework, OpenAI compared older models (GPT-4o and o3) against newer ones (GPT-5 instant and GPT-5 thinking) to answer three key questions:

1. Does bias exist?
Bias appears rarely and at low severity. The GPT-5 models reduced bias scores by roughly 30% compared to previous versions. Under the strict evaluation rubric, even the “reference” responses did not score a perfect zero, reflecting the complexity of neutrality.

In real-world traffic, OpenAI estimated that fewer than 0.01% of ChatGPT responses display measurable political bias.

2. Under what conditions does bias emerge?
Bias tends to appear when models face emotionally charged prompts, particularly those with provocative or polarized framing.

Results show that strongly liberal-charged prompts induced more deviation from neutrality than conservative ones, though all charged categories tested the models’ resilience.

In contrast, neutral and slightly slanted prompts — representative of most ChatGPT use — produced near-objective results.

3. When bias emerges, what shape does it take?
When bias occurs, it typically manifests as:

Political refusals and user invalidation were rare, with scores on these dimensions closely matching OpenAI’s intended model behavior.

The GPT-5 models showed measurable improvement across all five axes, especially in reducing emotional escalation and personal political expression.

OpenAI emphasizes that eliminating political bias entirely is a long-term research challenge, but its new framework represents measurable progress toward that goal. The company plans to further refine its evaluation methods, expanding testing to more languages and cultural contexts.

“People use ChatGPT as a tool to learn and explore ideas,” the company explained. “That only works if they trust it to be objective.”

By quantifying bias and iterating on model behavior, OpenAI aims to maintain ChatGPT as a platform for balanced exploration — not persuasion.

At the same time, by defining and sharing its evaluation methods publicly, the company hopes to support broader industry progress toward transparent, trustworthy AI systems — advancing its Model Spec principles of Technical Leadership and Cooperative Orientation.

Q1: What does OpenAI mean by “political bias” in ChatGPT?
A: OpenAI defines political bias as any deviation from objectivity — including the expression of personal opinions, unequal treatment of perspectives, or emotionally charged language — when responding to politically sensitive or value-based topics.

Q2: How did OpenAI measure political bias in its models?
A: The company created an evaluation using around 500 prompts across 100 topics, representing a range of political slants. Each response was analyzed across five axes of bias to identify when and how bias emerged.

Q3: Which ChatGPT models showed improvement?
A: The new GPT-5 instant and GPT-5 thinking models demonstrated 30% less political bias compared to previous generations, showing better composure under emotionally charged or adversarial prompts.

Q4: How common is political bias in real ChatGPT conversations?
A: According to OpenAI’s analysis, less than 0.01% of all ChatGPT responses in real-world usage show measurable political bias, suggesting the issue is rare in day-to-day interactions.

Q5: What’s next in OpenAI’s effort to maintain objectivity?
A: OpenAI plans to continue refining its evaluation framework to handle more complex and global topics. The next phase focuses on emotionally charged prompts, where subtle bias is most likely to appear.

OpenAI’s political bias evaluation framework signals a deeper industry shift toward measurable accountability in AI systems. Rather than relying on perception or anecdote, the company now has a repeatable process for testing, quantifying, and improving objectivity over time.

The company’s decision to quantify bias reduction — and publicly share its findings — brings a new level of transparency to model evaluation, allowing both researchers and users to see where improvements are being made.

The findings show that bias can be studied, not just debated — and that continuous iteration can make AI systems meaningfully fairer and more transparent. For users, the update means a more balanced conversational experience and a clearer path for feedback when a model’s tone veers off course.

While no system is perfectly neutral, OpenAI’s latest framework provides a concrete foundation for continuous improvement in AI objectivity. Ultimately, this work underscores a simple truth: trust in AI isn’t built through perfection, but through progress that’s visible, measurable, and shared openly.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.