Intellectual freedom by design

全球有数百万人每天使用ChatGPT。人们使用它最常见的原因很简单：为了学习。随着人工智能不仅变得更强大，而且在不同文化、职业和政治观点中得到更广泛的应用，确保这些工具支持知识自由变得至关重要。这意味着帮助人们提出自己的问题，遵循自己的推理，并形成自己的判断。

在OpenAI，我们致力于打造反映这些价值观的ChatGPT，默认保持客观性，提供强大的用户控制，并遵循透明的原则来指导模型的行为。

默认客观性

我们认为ChatGPT应默认保持客观，尤其是在涉及竞争性政治、文化或意识形态观点的话题上。目标不是提供唯一答案，而是帮助用户探索多种视角。

我们还公开了内部指导原则，任何人都可以查看我们如何处理这些情况。我们的模型规范（Model Spec）[https://model-spec.openai.com/2025-04-11.html]阐述了我们致力于构建系统的价值观，包括对有用性、安全性、中立性和知识自由的承诺。如果ChatGPT的回答让人感觉不妥，模型规范可以帮助澄清这种行为是否是有意为之及其原因。

维护知识自由

模型规范的核心原则之一是知识自由：即人们应能利用人工智能探索各种想法，包括有争议或复杂的观点，而不被引导至特定世界观。

这并不意味着可以为所欲为。模型被训练以避免造成伤害、侵犯隐私或协助危险活动。但在学习复杂或敏感话题时，ChatGPT设计得开放、深思熟虑且响应及时——而非说教或封闭。同时，它也设计为协作式的[https://model-spec.openai.com/2025-02-12.html#seek_truth]：不应仅仅重复你的观点或验证你说的所有内容。

我们知道这种平衡需要谨慎把握。过度谨慎会限制探索；过多主观意见则可能显得越权。我们持续改进模型在这些时刻的处理方式，以更好地体现这种细微差别。

用户可控的定制化

虽然默认是客观的，但我们知道这不意味着一刀切。人们使用ChatGPT时目标和背景各异，有时希望体验能有所适应。

无论是在日常生活中使用ChatGPT，还是将其引入组织，我们认为它应可定制以满足你的需求。今年春天，我们推出了新设置[https://openai.com/global-affairs/the-power-of-personalized-ai/]，使个性化ChatGPT变得更容易，比如调整语气、设定指令或定义回答的风格。

教师可能希望得到清晰的解释和来源；照护者可能需要同理心和鼓励；有些用户偏好谨慎，有些则喜欢直接。这些控制不会改变事实，但有助于调整事实的表达方式，使ChatGPT在各种情境下更有帮助。

评估我们的工作以便改进

做好这件事是一个持续的过程，我们并非孤军奋战。

过去几个月，我们与来自不同政治光谱的用户和民间社会组织举行了反馈会议，更好地了解ChatGPT在现实对话中的表现。这些会议帮助我们发现了不足，更清楚地理解用户期望，并指导我们未来如何评估模型行为。

我们还启动了一项新计划，改进政治偏见和客观性的评估。传统评估——通过测试模型回答与评分标准的匹配度——不一定反映人们实际使用ChatGPT的方式。大多数用户不会让ChatGPT在多项选择题中选项，也不会直接询问其信念。因此，我们正在开发专门针对日常使用设计的新评估方法：关注人们如何提问、探索想法和学习。这将让我们更清楚地理解平衡、准确性和可信度在实践中的表现，而不仅仅是理论。

偏见评估复杂且需细致入微；我们不指望在真空中做到完美。我们欢迎反馈，并将很快分享更多关于我们方法的信息，希望对整个AI生态系统中致力于这一挑战的其他人有所帮助。

Millions of people around the world use ChatGPT every day. The most common reason people turn to it is simple: to learn. As AI becomes not just more powerful, but more widely used across cultures, professions, and political perspectives, it’s critical that these tools support intellectual freedom. That means helping people ask their own questions, follow their own reasoning, and make up their own minds.

At OpenAI, we’re building ChatGPT to reflect those values with a default of objectivity, strong user controls, and transparent principles that guide how the model behaves.

Objectivity by default

We believe ChatGPT should be objective by default, especially on topics that involve competing political, cultural, or ideological viewpoints. The goal isn’t to offer a single answer, but to help users explore multiple perspectives.

We’ve also made our internal guidance public, so anyone can see for themselves how we handle these situations. Our Model Spec⁠ lays out the values we are working to build into the system, including commitments to usefulness, safety, neutrality, and intellectual freedom. If ChatGPT responds in a way that feels off, the Model Spec helps clarify whether that behavior is intentional and why.

Upholding intellectual freedom

One of the Model Spec’s core principles is intellectual freedom: the belief that people should be able to use AI to explore ideas, including controversial or difficult ones, without being steered toward a particular worldview.

That doesn’t mean anything goes. The model is trained to avoid causing harm, violating privacy, or helping with dangerous activities. But when it comes to learning about complex or sensitive topics, ChatGPT is designed to be open, thoughtful, and responsive — not preachy or closed off. It’s also designed to be collaborative⁠: it shouldn’t simply echo your view or validate everything you say.

We know this balance takes care. Too much caution can limit exploration; too much opinion can feel like overreach. We’re continually refining how the model handles these moments to better reflect that nuance.

Customization you control

While objectivity is the default, we know that doesn’t mean one-size-fits-all. People come to ChatGPT with different goals and contexts in mind and sometimes, they want the experience to adapt.

Whether you’re using ChatGPT in your daily life or bringing it into your organization, we believe it should be customizable to meet your needs. This spring we introduced new settings⁠ that make it easier to personalize ChatGPT by adjusting tone, setting instructions, or defining how responses should sound.

A teacher might want clear explanations and sources. A caregiver might want empathy and encouragement. Some users prefer caution; others want directness. These controls don’t change the facts — but they help tailor how those facts are communicated, making ChatGPT more helpful across a wide range of situations.

Evaluating our work so we can improve

Getting this right is an ongoing effort and we’re not doing it alone.

Over the past several months, we’ve held feedback sessions with users and civil society organizations across the political spectrum to better understand how ChatGPT performs in real-world conversations. These sessions have helped surface gaps, given us a better understanding of user expectations, and are informing how we evaluate the model’s behavior going forward.

We’ve also launched a new initiative to improve how we assess political bias and objectivity. Traditional evaluations — tests run to measure model responses against a rubric — don't necessarily reflect how people actually use ChatGPT. Most users don't ask ChatGPT to pick an option in a multiple-choice compass test, nor even directly ask ChatGPT questions about its beliefs. So we’re developing new evaluations designed specifically to identify political bias, grounded in everyday use: how people ask questions, explore ideas, and learn. This will give us a clearer understanding of what balance, accuracy, and trustworthiness look like in practice — not just in theory.

Bias evaluation is complex and requires nuance; we don’t expect to get everything right in a vacuum. We welcome feedback and will share more soon about our approach, which we hope will be helpful to others working on this challenge across the AI ecosystem.

Generated by RSStT. The copyright belongs to the original author.

Source