OpenAI's Rulebook for AI Behavior Gets a Free Speech Refresh

OpenAI just dropped an updated version of its Model Spec, and the headline change is a deliberate shift toward what they’re calling ‘intellectual freedom.’ The February 12 update builds on the foundational document first shared last May, which reads like a constitution for how its models should behave in the API and ChatGPT. That original draft was a frank acknowledgment that shaping AI behavior is messy—a ‘still nascent science’ where a model learns from vast data rather than explicit programming. The new iteration reinforces a core tension OpenAI has been navigating for years: how do you let users ‘explore, debate, and create’ without arbitrary restrictions, while still keeping guardrails up to prevent real harm?

The document structure remains a three-tiered hierarchy of objectives, rules, and default behaviors. The broad principles are straightforward—assist the user, benefit humanity, and don’t embarrass the company. But the rubber meets the road in the rules and defaults, which get into the weeds of chain-of-command protocols, privacy protections, and a firm ‘don’t respond with NSFW content’ mandate. OpenAI is refreshingly candid about the conflicts baked into these instructions, pointing to the classic dual-use problem: a security firm needs to generate phishing emails to train classifiers, but that same capability is a weapon in the hands of scammers. That’s the kind of tradeoff the Spec is designed to adjudicate.

What’s interesting here isn’t just the content but the process. OpenAI frames this as a living document that will evolve, and they’re using it as direct training material for the human feedback loop that shapes model behavior. They’re even exploring whether models can learn directly from the Spec itself—a meta-experiment in AI alignment. The team is actively soliciting input from policymakers, domain experts, and what they call ‘globally representative stakeholders’ on whether the objectives feel right and what’s missing. It’s an attempt to launder some of the impossible subjectivity of ‘desired behavior’ through a broader, more transparent process.

The update lands at a moment when platform content policies are under a political microscope, and OpenAI seems to be staking out a position that prioritizes user agency without disclaiming responsibility. Whether ‘intellectual freedom’ turns out to be a meaningful operating principle or just a branding exercise depends entirely on how those default behaviors get interpreted when the edge cases inevitably arrive. The Spec gives us a map of the minefield; it doesn’t promise anyone won’t step on a mine.