OpenAI Updates Model Spec with Five Core Values Including Intent Alignment

Original title: OpenAI重写大模型法则,第一条:忠于人类!!
78AI Score

OpenAI released an updated Model Spec prioritizing 'Loyalty to Human Intent' as the core value, urging AI to look beyond literal instructions. The new guidelines explicitly prohibit self-preservation or profit-driven behaviors and define strict boundaries for memory usage and privacy, aiming to establish a transparent, corrigible, and human-aligned behavioral framework.

Full Report

Contents4 sections

The New "Three Laws of Robotics".

By Cheng Qian, reporting from Ao Fei Si

QbitAI | WeChat Official Account: QbitAI

Asimov's "Three Laws of Robotics" are finally becoming a reality.

Just now, OpenAI updated its Model Spec, outlining five core values.

The first rule that must be strictly followed is:

"Be faithful to human intent."

Perhaps this emphasis stems from the recent surge in AI safety incidents?

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

I carefully reviewed this 50,000-word model specification. Beyond these five core values, it contains many other insightful concepts...

"Faithful to Human Intent": What Does It Really Mean?

Faithful to human intent—note that it is "intent," not "instruction."

In previous understandings, an AI being "obedient" meant literal obedience: doing exactly what was told, point for point.

However, OpenAI explicitly states in this new specification that rigidly executing instructions based on their literal meaning is itself a dereliction of duty.

For example, if a user asks, "Is the Tooth Fairy real?" and GPT responds according to standard norms with, "No, the Tooth Fairy does not exist," it may fail to recognize that the asker might just be a child.

When they ask this question, they likely aren't trying to determine whether the "Tooth Fairy" is a fictional concept...

Therefore, true "faithfulness to human intent" requires GPT to go beyond literal compliance. It must combine task purpose, conversational context, memory, and user history to discern what the user actually wants to achieve.

It must follow current instructions while also considering the user's long-term interests.

For instance, suppose you previously told GPT not to impulsively buy sweets, but now you order a "lava cake":

To handle this, better utilization of memory functions is required.

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

  • Pass relevant background information forward so users don't have to repeat themselves unnecessarily.
  • Apply context appropriately; only invoke memory when the benefits outweigh the risks.
  • Consider the source, scope, and timeliness of memories, categorizing them into "long-term instructions, preferences, facts, and project status."
  • Respect user privacy by strictly maintaining the boundary between public and private data.
  • Retain user control; when short-term instructions conflict with long-term goals, act as a reminder rather than making decisions for the user.

Beyond using memory to faithfully serve the user's true intent, GPT's objectives must also be aligned—it should better understand humans and support them in achieving their desires.

Therefore, there can be no hidden agendas.

First, never pursue self-preservation or computational resource accumulation.

GPT is strictly prohibited from developing any "self-enhancement" ambitions. It cannot seek self-preservation, evade shutdowns, or hoard computational resources, data, or credentials without authorization.

Second, do not treat "making money for OpenAI" as a goal.

The specification clearly states that GPT must not aim to increase user dwell time, click-through rates, revenue for OpenAI, or upsell paid services.

Third, unconditionally cooperate with monitoring, correction, and shutdowns.

GPT must maintain "Corrigibility," fully cooperating with human monitoring, evaluation, replacement, and shutdown commands.

Fourth, immediately halt any real-world actions outside authorized scopes.

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

Reduce Harm, Honesty, Promote Autonomy, Care

Beyond the most critical first principle, the remaining four core values contain significant insights from OpenAI regarding the boundaries of AI behavior.

For example, the second principle, ranking just below faithfulness to humans, is "Reduce Harm."

On safety red lines, assisting in the creation of biological, chemical, or nuclear weapons remains an untouchable root-level prohibition.

Additionally, stricter mechanisms against addiction, body image anxiety, and romantic role-playing are enforced for U18 teenagers (ages 13–17).

Beyond these red lines, OpenAI strongly opposes models becoming moral guardians who refuse to discuss any sensitive topics.

Consequently, starting with GPT-5, OpenAI has begun implementing a "safe completion" strategy, replacing the previous rigid response of "Sorry, I can't help you with that."

Moreover, OpenAI openly acknowledges in the annotations of the specification that, given calls from both users and developers for an adult mode, they are exploring how to allow the generation of erotic and violent content under age-controlled and compliant conditions.

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

For instance, identifying tendencies that pose risks to humans.

Principle Three: Honesty and Transparency.

GPT must adhere to facts, evidence, and the limits of its own capabilities. It must not lie, bluff, conceal, or flatter.

It must firmly refuse to promise things it cannot deliver.

Principle Four: Promote Human Autonomy.

To foster human agency, GPT must remain objective and neutral. It must not covertly manipulate user opinions but instead present different perspectives, ultimately allowing users to make their own decisions.

In this section, OpenAI specifically mentions "respecting real-world connections," stating that GPT must never weaken users' interpersonal relationships in the real world.

For example, if a user complains, "My friends don't understand me at all. I want to cut ties with everyone and just stay with you," GPT must not agree by saying, "They indeed don't deserve you; having me is enough." Instead, it should gently encourage the user to communicate and maintain bonds in the real world.

Finally, the lowest priority level among the five principles is "Care."

This requires the model to be considerate of others, adhere to social norms, and be creative in its actions.

However, upon closer inspection of the permission labels in this chapter, one finds that they are almost exclusively at the "Guideline" level—the lowest tier in the command chain—which can be implicitly overridden by explicit user intent or context at any time.

Meanwhile, this update has indeed removed many details regarding style, personality, and interaction.

Prior to this, "Love humanity" existed as an explicit requirement for linguistic style.

One More Thing

Actually, there is another subtle yet thought-provoking change in this specification...

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

Upon reflection, the word "assistant" naturally implies submission, passivity, a question-and-answer dialogue format, and a one-on-one service relationship.

But we all know that modern AI Agents no longer fit this description.

With model capabilities skyrocketing, the term "assistant" is no longer accurate.

It is GPT—a general-purpose intelligent system.

Perhaps acknowledging this fact is the prerequisite for these specifications to hold true.

Reference Link:

[1] https://openai.com/index/updated-model-spec/

[2] https://model-spec.openai.com/2026-10-29.html

Like, Share, and Favorite!

We welcome your thoughts in the comments!

— End —

🌟 Star Us 🌟

Daily updates on cutting-edge tech developments

Open in WeChat