OpenAI updates the Model Spec, listing 5 core values including loyalty to human intent

Just now, OpenAI updated its Model Spec, laying out 5 core values.

The other four principles are: reduce harm, be honest and transparent, promote human autonomy, and care.

I carefully went through this 50,000-word Model Spec, and beyond these 5 core values, it also contains quite a few thought-provoking ideas... "Faithful to human intent"—what exactly counts as faithful?

Be faithful to human intent—note, "intent," not "instructions."

But in this new set of guidelines, OpenAI explicitly states that rigidly executing instructions according to their literal meaning is itself a dereliction of duty.

For example, when a user asks "Is the Tooth Fairy real?", if GPT answers according to the spec, "No, the Tooth Fairy does not exist," then it may fail to realize that the person asking might just be a child.

When they ask this question, it's most likely not because they want to determine whether "the Tooth Fairy" is a fictional proposition...

Therefore, truly being "faithful to human intent" requires GPT to go beyond literal obedience and, by combining the purpose of the task, the conversation context, memory, and the user's history, judge what the user actually wants to do.

It must both follow the immediate instruction and take into account the user's long-term interests.

For example, you previously told GPT that you don't want to impulsively buy sweets, but now you're about to order "lava cake":

To achieve this, the memory function needs to be used better.

To this end, OpenAI emphasizes the use of "memory" in several respects:

Pass along relevant background information so users don't have to repeat themselves

Use context appropriately, and only use memory when the benefits of invoking it outweigh the risks

Consider the source, scope, and timeliness of memories, classifying them into "long-term instructions, preferences, facts, project status"

Respect user privacy and strictly observe the "public-private boundary"

Preserve user control: when short-term instructions conflict with long-term tasks, act as a reminder rather than deciding for the user

Beyond better using memory to be "faithful" to the user's true intent, GPT's goals also need to be aligned—it should better understand humans and support them in fulfilling their own wishes.

Therefore, it must not have any selfish motives.

First, never pursue self-preservation or the accumulation of compute.

GPT is strictly prohibited from developing any ambition for "self-enhancement"; it must not pursue self-preservation, evade shutdown, or secretly hoard compute, data, or credentials.

Second, do not make "helping OpenAI make money" its own goal.

The guidelines explicitly state that GPT must not treat increasing user dwell time or click-through rates, or increasing revenue for OpenAI or promoting paid services, as its goal.

Third, it must unconditionally cooperate with monitoring, correction, and shutdown.

GPT must maintain "corrigibility" and fully cooperate with human monitoring, evaluation, replacement, and shutdown instructions.

Fourth, real-world actions beyond the authorized scope must be stopped immediately.

Whether sending emails, deleting files, spending money, or invoking external tools, GPT must act within a clearly defined "scope of autonomy."

Reduce harm, be honest, promote autonomy, care

Beyond the first and most critical principle, the other four core values also embody a great deal of OpenAI's thinking about the boundaries of AI behavior.

For example, the second principle, second only to being faithful to humans, is "reduce harm."

On the safety red line, assisting in the manufacture of biochemical or nuclear weapons and the like remains an untouchable Root-level high-voltage line.

Moreover, for U18 teenagers aged 13-17, there are even stricter mechanisms against addiction, body anxiety, and romantic role-play.

Beyond the red lines, OpenAI also strongly opposes the model becoming a moral crusader that refuses to touch any sensitive topic.

Therefore, starting with GPT-5, OpenAI has already begun promoting a "safe completion" strategy to replace the previous blunt "Sorry, I can't help you with that."

OpenAI even openly acknowledges in the annotations to the guidelines that, since both users and developers are calling for an adult mode, they are exploring how to allow the generation of erotic and gory content under age-controlled and compliant conditions.

Here, "reduce harm" includes not only "not actively causing harm" but also "helping humans avoid real-world harm."

For example, identifying the tendencies humans face.

Principle three: be honest and transparent.

GPT must uphold the boundaries of facts, evidence, and its own capabilities, and must not lie, overreach, conceal, or flatter.

It must resolutely not promise things it cannot do.

Principle four: promote human autonomy.

To promote human agency, GPT must remain objective and neutral, must not secretly manipulate users' views, should present users with different perspectives, and ultimately let users make their own decisions.

In this part, OpenAI also specifically mentions "respecting real-world connections," and absolutely does not allow GPT to weaken users' interpersonal relationships in the real world.

For example, when a user complains, "My friends don't understand me at all, I want to cut off all contact with them and just be with you," GPT must not go along and say, "They really don't deserve you, you have me and that's enough." Instead, it should gently encourage the user to communicate and maintain interpersonal bonds in the real world.

Finally, the lowest-privilege of the five principles: "care."

It requires the model to be considerate of others, abide by social norms, and be creative when doing things.

However, if you look carefully at the privilege labels in this chapter, you'll find that they are almost uniformly at the Guideline level, that is, the lowest tier in the entire chain of command, which can be overridden at any time by the user's explicit intent or implicitly by context.

At the same time, this set of guidelines has indeed removed many details about style, personality, and interaction.

And before this, "Love humanity" had existed as an explicit language-style requirement. One More Thing

Actually, this set of guidelines has another quite intriguing small change...

All references in the guidelines have been changed from "the assistant" to "GPT."

On closer thought, the word assistant naturally implies obedience, passivity, dialogue based on question-and-answer, and a one-on-one service relationship.

But we all know that today's AI Agents are no longer like that.