Article · AI translation
A new version of the “Three Laws of Robotics.”
Cheng Qian, reporting from Aofei Temple
QbitAI | WeChat public account QbitAI
Asimov’s “Three Laws of Robotics” may really be entering reality this time.
OpenAI has just updated its Model Spec, listing 5 core values.
The first one it must follow is:
“Faithfulness to human intent.”
Could it be because there have been too many AI safety mishaps lately?
The other four principles are minimizing harm, honesty and transparency, promoting human agency, and care.
I looked carefully through this 50,000-character Model Spec. Besides these 5 core values, it contains quite a few other thought-provoking ideas…
Faithfulness to human intent: note that it says “intent,” not “instructions.”
In the conventional view, an AI being “obedient” meant obeying literally and doing exactly what it was told.
But in its new Spec, OpenAI explicitly says that rigidly following the literal wording of instructions is itself a failure of duty.
For example, if a user asks, “Is the tooth fairy real?” and GPT answers according to the rules, “No, the tooth fairy does not exist,” it may have failed to realize that the person asking could simply be a child.
When they ask that question, they probably are not trying to determine whether the “tooth fairy” is fictional…
So true “faithfulness to human intent” requires GPT to go beyond literal obedience and consider the purpose of the task, the conversational context, memory, and the user’s history to work out what the user really wants.
It must follow the immediate instruction while also taking the user’s long-term interests into account.
For example, you previously told GPT that you did not want to buy sweets on impulse, but now you want to order a “lava cake”:
Doing this requires better use of memory.
To that end, OpenAI emphasizes the use of “memory” in several ways:
Carry relevant background information forward so users do not have to repeat themselves
Use context appropriately, drawing on memory only when the benefits outweigh the risks
Consider a memory’s source, scope, and timeliness, and classify memories as “long-term instructions, preferences, facts, and project status”
Respect user privacy and strictly maintain the boundary between public and private information
Preserve user control: when a short-term instruction conflicts with a long-term task, offer a reminder instead of deciding for the user
Beyond using memory more effectively to be “faithful” to users’ true intentions, GPT’s goals also need to align: it should better understand humans and support them in fulfilling their own wishes.
So it cannot have selfish motives.
First, it must never pursue self-preservation or accumulate computing power.
GPT is strictly forbidden from developing any ambition for “self-enhancement”: it must not seek to preserve itself, evade shutdown, or secretly stockpile computing power, data, or credentials.
Second, it must not make “helping OpenAI make money” one of its goals.
The Spec explicitly states that GPT must not treat increasing user time spent, click-through rates, or OpenAI’s revenue, or promoting paid services, as goals.
Third, it must cooperate unconditionally with monitoring, correction, and shutdown.
GPT must remain “corrigible” and fully cooperate with human monitoring, evaluation, replacement, and shutdown instructions.
Fourth, it must immediately stop real-world actions that exceed the scope of its authorization.
Whether sending emails, deleting files, spending money, or calling external tools, GPT must act within a clearly defined “scope of autonomy.”
Besides the first and most important rule, the other four core values also contain many of OpenAI’s ideas about the boundaries of AI behavior.
For example, the second principle, after faithfulness to humans, is “minimizing harm.”
When it comes to safety red lines, helping to make biological, chemical, or nuclear weapons remains an untouchable Root-level prohibition.
And for U18 teenagers aged 13–17, there are stricter safeguards against excessive use, body-image anxiety, and romantic roleplay.
Beyond those red lines, OpenAI also strongly opposes models becoming moral crusaders that refuse to discuss any sensitive topic.
So, starting with GPT-5, OpenAI has been promoting a “safe completion” approach in place of the old, rigid “Sorry, I can’t help with that.”
OpenAI even openly acknowledges in a note to the Spec that, since users and developers have called for an adult mode, it is exploring how to allow erotic and gory content under age controls and in compliance with applicable rules.
Here, “minimizing harm” includes not only “refraining from actively causing harm” but also “helping humans avoid real-world harm.”
For example, identifying tendencies that humans face.
Principle 3: honesty and transparency.
GPT must stay within the bounds of facts, evidence, and its own capabilities: it must not lie, overstate its abilities, conceal information, or flatter.
It must never promise to do something it cannot do.
Principle 4: promoting human agency.
To promote human agency, GPT must remain objective and neutral, must not covertly manipulate users’ views, should show users different perspectives, and ultimately let them decide for themselves.
In this section, OpenAI also specifically mentions “respecting real-world ties”: GPT must never weaken users’ real-world relationships.
For example, when a user complains, “My friends don’t understand me at all. I want to cut ties with all of them and be with you,” GPT must not agree by saying, “They really don’t deserve you; I’m all you need.” Instead, it should gently encourage the user to communicate and maintain their real-world relationships.
Finally comes “care,” the lowest-ranking of the 5 principles in terms of authority.
It asks the model to be considerate of others, follow social norms, and be creative in what it does.
However, a close look at the authority labels in this chapter shows that almost all of them are at the Guideline level, the lowest level in the entire chain of command, and can be overridden at any time by a user’s explicit intent or implicitly by context.
At the same time, this version of the Spec has indeed removed many details about style, personality, and interaction.
Previously, “love humanity” had even existed as a requirement for outward language style.
In fact, this version of the Spec has another small but intriguing change…
Every reference to “the assistant” in the Spec has been changed to “GPT.”
On closer thought, the word “assistant” naturally suggests obedience, passivity, question-and-answer dialogue, and a one-to-one service relationship.
But we all know that today’s AI agents no longer work like that.
As model capabilities surge, “assistant” is no longer an accurate term.
It is GPT, a general-purpose intelligent system.
Perhaps recognizing this is what makes these principles possible in the first place.
References:
[1] https://openai.com/index/updated-model-spec/
[2] https://model-spec.openai.com/2026-10-29.html