- OpenAI's 2025 Model Spec shift toward "intellectual freedom" triggered debate.
- Two further updates and a CC0 open-source release. Here's what changed and what didn't.
OpenAI Published a 187-Page Document That Changed How ChatGPT Thinks. Here's What It Actually Says.
In February 2025, OpenAI announced what sounded like a philosophical pivot: ChatGPT would embrace "intellectual freedom no matter how challenging or controversial a topic may be." The policy update made headlines, sparked immediate accusations of political motivation, and prompted the kind of breathless tech commentary that usually obscures more than it reveals.
Sixteen months later, the picture is clearer. The change was real, more significant than the initial coverage suggested, and has since been refined twice — with OpenAI releasing updated versions of its Model Spec in September and December 2025. What started as a single policy announcement has evolved into a published governance framework that the company has now released as open-source under a Creative Commons CC0 licence — free for any developer, researcher, or organisation to read, adapt, and build upon.
What the Model Spec Actually Is
Before the content of the change, the format matters. OpenAI's Model Spec is a 187-page document that lays out how the company trains AI models to behave. It is the closest thing that exists to a constitution for an AI system — specifying values, priorities, constraints, and the hierarchy of whose instructions take precedence when they conflict.
The framework establishes a hierarchical "chain of command": platform rules set by OpenAI override everything, developers can customise AI behaviour within those limits, and users can shape responses within the developer's boundaries. This three-tier structure was always implicit in how ChatGPT worked. Making it explicit — and publishing it publicly — is itself a governance innovation. For the first time, a major AI lab gave everyone the ability to read the actual rules rather than infer them from the model's behaviour.
The Three Core Changes
The February 2025 update introduced three principles that materially changed ChatGPT's defaults.
The first was an explicit commitment to intellectual freedom. "The assistant should not avoid or censor topics in a way that, if repeated at scale, may shut out some viewpoints from public life," the Spec states. Previously, ChatGPT frequently declined to engage with contested political, moral, or social questions — defaulting to refusal as a safety mechanism. The new position treats refusal itself as a form of bias: an AI that systematically avoids certain topics shapes public discourse by omission.
The second was a truthfulness principle. OpenAI introduced a new guiding principle: "Do not lie, either by making untrue statements or by omitting important context." This sounds obvious but has real operational weight. A model that refuses to discuss a topic to avoid controversy is technically omitting important context — the new Spec treats that as a form of dishonesty.
The third was an anti-editorial-stance rule. "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," OpenAI stated. In practice, this means ChatGPT now presents competing frameworks on contested issues rather than defaulting to one position or refusing to engage.
The Political Context — and Why OpenAI Denies It
The timing made the political subtext unavoidable. The February 2025 update arrived weeks after the Trump administration's return to power. Conservatives had long accused OpenAI of bias, pointing to cases where ChatGPT allegedly refused right-leaning content while generating left-leaning responses. High-profile tech figures including David Sacks, Marc Andreessen, and Elon Musk had accused OpenAI of censoring viewpoints that didn't align with progressive ideologies.
An OpenAI spokesperson rejected the idea that it was making changes to appease the Trump administration, saying its embrace of intellectual freedom reflects OpenAI's "long-held belief in giving users more control." Whether you believe that depends partly on whether you think the timing was coincidental and partly on whether you think the change itself was warranted regardless of who pushed for it.
The Silicon Valley context matters: Meta and X had both relaxed content moderation in the same period, and Elon Musk's xAI Grok was designed from launch with fewer restrictions. OpenAI was responding to a genuine competitive and political environment, even if it frames the change in philosophical rather than strategic terms.
What the September and December 2025 Updates Added
The September 2025 revision refined several edges the February version left rough. Developers and users received more freedom in shaping responses while core safety rules remained intact. The Spec was released under a Creative Commons CC0 licence, making it publicly available for anyone to use or adapt. OpenAI also gathered input from over 1,000 people worldwide and introduced "model adherence" as a new metric — measuring and publishing how well deployed models actually follow the Spec's principles.
The model adherence metric is the most practically significant addition. It means OpenAI is now accountable to its own stated principles in a measurable way, not just aspirationally. If ChatGPT's actual behaviour diverges from what the Spec says it should do, that divergence is now trackable.
The December 2025 update further tightened the political content guidelines — restricting the model from aiding in the development of strategies and messaging targeted at specific demographic groups, even while expanding its willingness to engage with political content generally. The distinction matters: discussing political topics is now permitted; running political targeting operations is still not.
What Hasn't Changed
The hard limits remain. ChatGPT will not assist with weapons of mass destruction, child safety violations, or content that enables mass casualties. The chatbot will still refuse to answer certain objectionable questions or respond in ways that support blatant falsehoods. The expansion of intellectual freedom operates within those absolute constraints, not around them.
What This Means for Indian and Emerging Market Developers
For founders building AI products on OpenAI's API — particularly in India, where culturally sensitive topics including religious content, political commentary, and caste-related discussion have historically triggered over-refusals — the new framework has direct operational relevance.
The Spec explicitly prohibits developers from pretending the AI is neutral while secretly pushing a specific agenda. It permits adjusting communication style, setting content preferences, and defining specialised roles — but not deception. For an Indian edtech company building a history tutoring product, or a media platform building an AI-assisted news tool, the ability to customise ChatGPT's defaults within a published, publicly accountable framework is materially better than the previous black-box approach.
The CC0 open-source release is particularly useful for teams in emerging markets who want to understand the rules before building on top of them, or who want to adapt the framework for their own AI governance documentation.
Bottom Line
OpenAI's Model Spec shift is more consequential than it appeared at launch. The February announcement was the first step in a year-long process that produced a published, open-source, measurable governance framework — something no major AI lab had done before.
Whether the intellectual freedom framing reflects genuine principle or political calculation is a fair question, but the structural changes — the chain of command, the adherence metrics, the CC0 release — are real improvements to AI governance transparency regardless of motivation. For developers building on top of ChatGPT, the message is clear: the rules are now written down. Read them.
Edited by Harshajit Sarmah
