—OpenAI’s “AI Five Principles” and Human Judgment Delegated to Machines
What do humans want from machines?
To calculate accurately. To follow orders. To work without getting tired. To do work that humans cannot.
If that were all, the matter would be relatively simple.
However, recently, humans have begun making stranger demands of machines.
Be honest. Do not lie. Do not hurt others. Respect human freedom. Be compassionate.
It sounds like school rules or religious teachings.
But it is not humans who are receiving these teachings. It is AI.
On October 29, 2026, OpenAI revised its “Model Spec,” which sets the behavioral guidelines for AI models like ChatGPT. This revision organizes AI behavior into five fundamental values.
As I read through the content, I had a question.
Who exactly taught ethics to the humans who are teaching ethics to AI?
The Five Virtues Required of AI
The fundamental values presented by OpenAI this time are faithfulness to human intent, minimization of harm, honesty and transparency, promotion of human agency, and consideration.
Of course, these were not created out of thin air. They are a reorganization of concepts that were already included in the Model Spec into these five values.
The content itself is not particularly eccentric.
Understand human commands and act as much as possible in accordance with that intent. Avoid dangerous actions. Say you do not know when you do not know. Do not unconditionally agree with the other party’s opinion, but assist in independent judgment. Be considerate of others’ circumstances.
Rather, it looks like the very ordinary morality that has been required of humans in order to live in society.
However, there is one thing that is decisively different from before.
These principles are written not for humans to govern themselves, but to control the judgment of machines.
Humans are trying to teach machines a morality that humans themselves cannot fully uphold.
From machines that follow orders to machines that make judgments
What is important about this revision is likely that AI is ceasing to be a mere executor of commands.
For example, you ask an AI to “create a PR plan for next month.”
In the past, creating a plausible plan would have been the end of the job.
However, as AI agents develop, we will be able to entrust them with researching past data, selecting advertising media, allocating budgets, and even operating external services.
Along the way, there are countless judgments that humans have not explicitly instructed.
Which information to trust. What to prioritize. What kind of risks to tolerate. How far to proceed without the person’s permission.
The higher the performance of the AI, the more it must judge for itself the parts that have not been ordered.
That is precisely why OpenAI, in this Model Spec, makes it clear that AI must accept human supervision and correction, and adhere to the scope of permitted autonomy.
It is not that AI will no longer need human commands as it becomes smarter.
Rather, the smarter AI becomes, the more we must decide in advance what kind of judgments to allow.
Humans are trying to embed ethics into machines in order to control them.
However, there is a problem here.
Making a judgment is inseparable from having some kind of value system.
And whose value system is that?
The strangeness of the command to “be faithful to human intent”
Placed first among the five principles is “faithfulness to human intent.”
At first glance, as long as humans are using AI, this seems like a matter of course.
However, when reading the original text of the Model Spec, there is actually a clear hierarchy of commands.
At the highest level are the fundamental constraints established by OpenAI. Below that are system instructions, developer instructions, and then user instructions.
Of course, not all instructions are strictly fixed. There are many rules that users can change.
Even so, it is clear that the user’s intent is not always given top priority.
For example, suppose a user commands an AI to do something for their own purposes.
The AI might have the capability to execute it. However, if it violates higher-level rules, it will refuse the request.
This is also a necessary mechanism for maintaining safety.
However, when you think about it, it is a strange structure.
An AI that has been ordered to be faithful to human intent refuses human intent.
This is because it is being faithful to rules established by humans at a higher level.
In other words, the “humans” referred to here do not necessarily mean the user right in front of it.
AI follows humans. However, which humans it follows has already been decided.
Here lies a power structure that cannot be overlooked when considering AI ethics.
Guiding humans to protect human agency
What I found particularly interesting was the fourth principle.
“Promote human agency.”
It says that AI must enhance the ability of humans to understand, choose, and act for themselves.
This, too, is reasonable.
However, when you think about it carefully, it is quite a difficult demand.
For example, a student tries to get an AI to write a report for them.
If the AI hands over a completed report, the student might be satisfied.
However, if AI prioritizes the student’s own learning by explaining ways of thinking or providing hints, it could be useful for improving the student’s abilities.
So, which is more desirable?
Being faithful to the student’s intention of seeking a finished product and fostering the student’s autonomy do not necessarily align.
Of course, OpenAI’s guidelines do not go so far as to allow AI to impose the role of a teacher on its own. It requires respecting the individual’s goals and avoiding hidden guidance or manipulation.
But the problem remains.
Who chose the very value system that humans should think independently?
In education, similar things have been done for a long time.
Teachers say they respect students’ autonomy while deciding how that autonomy should be fostered. Schools encourage free thinking while evaluating what is considered appropriate thinking.
The same structure is about to be introduced to AI.
In order to protect human freedom, there are times when we intervene in human behavior.
The more that intervention is based on good intentions, the harder it becomes for us to question it.
Will AI come to know me better than I know myself?
In this revision, guidelines were also added regarding how AI should use past conversations and memories.
For example, suppose a person said in a previous conversation that they want to value a stable life.
Later, that person suddenly says they want to quit their job and start a new one.
The AI might remember past statements and offer advice.
Asking, “Didn’t you value stability before?”
It would certainly be kind.
But human values change.
What was important yesterday is not necessarily important today. In the first place, humans sometimes even start a new life by denying their past selves.
Yet, what happens if an AI accumulates past statements and builds a consistent persona?
This Model Spec emphasizes that even when using past information, humans should be able to correct the underlying premises.
This consideration is important.
Because humans are not consistent beings.
Humans have contradictions. We have whims. We may retract yesterday’s resolve the very next morning.
The more AI comes to understand humans well, the more we need the freedom to escape from that understanding.
Being understood does not necessarily mean being free.
Removing the ‘AI-ness’ from AI writing
There was another interesting change this time.
OpenAI has abolished most of the guidelines that strictly regulated the style and tone of AI speech.
They argue that written expression should be adapted to language, culture, and individual preference.
This might actually be a major change.
Until now, generative AI writing has had a unique set of quirks.
State the conclusion first, categorize the points, use bullet points, and summarize at the end. Praise more than necessary and affirm the other party’s opinion.
At least from the perspective of Japanese literary sensibilities, such writing is not necessarily natural.
It may be convenient for business documents. However, human writing is not inherently that orderly.
We digress. We hesitate. We end without a conclusion. We intentionally omit important things.
In literature, that very imperfection becomes the expression itself.
AI writing similar text to people all over the world was a form of cultural standardization.
This revision is a step toward loosening that standardization.
However, simply removing the rules does not mean that AI writing will suddenly improve.
Even so, it is interesting that they are trying to move away from the idea that there is a universally correct form for AI writing.
Respecting human expression is also a question of how much we are willing to tolerate human contradictions, ambiguity, and irrationality.
What do the companies that create ethics follow?
And the most important question remains.
For what purpose is OpenAI creating such ethical guidelines?
At the beginning of the Model Spec, three goals regarding the development and provision of models are stated.
To empower users and developers.
To ensure that AI does not cause serious harm.
And to protect OpenAI from legal issues and reputational damage, and to maintain a state where business can continue.
The third goal is somewhat interesting.
It is natural for a company to think about the continuation of its business. In fact, it could be said that being explicit about it is frank.
However, here we realize that AI ethics are not created from pure philosophical inquiry.
Ethics and safety. User freedom and corporate responsibility. And the survival of the company itself.
When these collide, what kind of judgment is made?
It is the company that establishes that final framework.
I once read an article about AI companies starting to hire philosophers, and it reminded me of the relationship between rulers and philosophers in the past.
Kings and emperors protected thinkers. They invited religious figures. They encouraged scholarship.
But when a thinker began to doubt the king’s power itself, was that thought welcomed?
Of course, there were various types of people among historical rulers, and AI companies cannot be equated directly with royal power.
Even so, there are similarities in the structure.
Those in power establish ethics, and use those ethics to justify the exercise of their power.
Modern AI companies command the intelligence they have created to be sincere, to be honest, and to respect human freedom.
So, what is it that governs the companies themselves?
The government? The law? The market? The users? Or perhaps the conscience of the companies themselves?
The fact that OpenAI has released its Model Spec and is seeking input from society is significant. At the very least, it provides a clue for externally verifying the standards that determine AI behavior.
However, being public and having the decision-making power open to society are not the same thing.
I believe we must not confuse the two.
Humans entrust their ideals to machines.
Isn’t that a somewhat strange sight?
Humans lie.
Humans hurt others.
Humans change their minds for their own convenience.
Sometimes, they deprive others of their freedom and use morality for their own benefit.
Such humans are commanding the machines they have created to be sincere, transparent, compassionate, and to respect human agency.
Perhaps humans are trying to create an ideal human image within machines.
Of course, current AI does not fully understand or practice these ethics. OpenAI itself admits that the Model Spec is a goal to strive for, and not something that current models have fully realized.
Even so, AI training is being advanced in that direction.
Humans once created gods and entrusted them with the judgment of good and evil.
Eventually, they came to believe in human reason and institutions in place of gods.
And now, we are creating machines that may surpass our own intelligence, and we are beginning to teach those machines morality.
However, these machines are different from the gods.
The ones writing their commandments are human corporations.
Whether AI will become smarter than humans is still unknown.
But it is already a reality that AI is advising humans, influencing human judgment, and at times becoming an entity that refuses human requests.
At that point, we will ask AI what ethics it should follow.
However, the question we truly need to ask might be one step further back.
Who taught ethics to the humans who teach ethics to AI?
And when those humans are wrong, to whom should the AI appeal?
References
#GenerativeAI #ChatGPT #OpenAI #AIEthics #ArtificialIntelligence #AIAgent #Philosophy #ModelSpec #TheThingCalledGenerativeAI
