Why AI Problems Are Becoming Philosophical Problems

KunYuan

Jun 22, 2026

When I say that AI problems are becoming philosophical problems, I am not trying to make a technical issue sound more abstract.

I am also not saying that every AI engineer must first study the history of philosophy, or that every AI product must be translated into philosophical language. That would be too heavy, and too vague. The real situation is simpler: many AI problems first appear as engineering problems, product problems, governance problems, or safety problems, but the deeper we go, the more they force us to ask questions that engineering alone cannot define.

These questions are not appearing because philosophy wants to enter AI. They are appearing because AI has entered the things philosophy has always cared about.

When AI only answers one-off questions, it is still easy to understand it as a tool. It produces a piece of text, and we check whether that text is correct, useful, or aligned with what we asked for. At that stage, many problems can still be treated as output problems.

But when AI begins to remember, advise, plan, call tools, participate in action, generate explanations, influence decisions, and enter work, education, organizations, governance, and everyday life, the question is no longer only whether the output is good. We begin to face more basic things: memory, action, responsibility, reality, judgment, and meaning.

These words may sound philosophical, but they are already showing up inside products and systems.

A long-term memory system forces us to ask: who does memory belong to? Is it the user’s record, the system’s asset, part of a relationship, or a structure that can be called, edited, forgotten, and transferred? If this question is not clearly defined, long-term memory is no longer merely a feature. It becomes a problem of identity, boundary, and trust.

An agent system forces us to ask: who is acting? When AI calls tools, modifies files, writes code, triggers workflows, or executes tasks, we cannot only ask whether it achieved the goal. We also have to ask whose action this is. Is it the model’s action, the user’s action, the system’s action, or a chain of delegated action?

An explanation system forces us to ask: is explanation the same as evidence? AI can produce an explanation that sounds reasonable, but does that explanation actually correspond to the process by which the system arrived at the result? Is it showing a real cause, or generating a narrative that makes the user feel reassured? If an explanation merely sounds plausible, it cannot automatically function as audit.

A human-in-the-loop process forces us to ask: does the presence of a human mean the presence of judgment? A person can confirm, authorize, review, and approve, but these actions do not automatically mean that the person understood the reasons, set the boundaries, recognized the risks, or took responsibility for the consequences. This is no longer only a workflow design issue. It is a question of judgment, responsibility, and agency.

A generative system forces us to ask: how is reality verified? When text, images, voices, video, data, and evidence can all be generated, reality is no longer only about whether there is content. It is about whether the content still has a traceable, testable, and accountable relationship to the world.

Education and work force us to ask: how are human capacities formed? If AI can help students write, researchers summarize, employees decide, and managers generate plans, then we must ask not only whether efficiency has improved, but whether humans are still going through the processes through which ability is formed.

These questions are not staying inside philosophy books. They have become real problems in AI system design, product interaction, safety boundaries, governance responsibility, and everyday use.

This is why I say AI problems are becoming philosophical problems.

Not because philosophy is higher than engineering, but because engineering must first know what it is trying to implement. If a system is going to handle memory, it must know what counts as memory. If an agent is going to act, it must know what counts as action. If a governance system is going to allocate responsibility, it must know what counts as responsibility. If an explanation mechanism is going to support trust, it must know the relationship between explanation and evidence. If a human-in-the-loop design is going to preserve human agency, it must know what it means for judgment to truly remain present.

If these concepts are not defined clearly, engineering can still move forward, but it moves forward inside ambiguity. Features increase, capabilities grow, workflows become smoother, but we may not know what we are amplifying.

This is how I understand philosophy-first.

It is not a posture, not a style, and not a way to make AI sound profound. It simply means that in certain places, philosophy must help define the problem before engineering can truly test the answer.

Philosophy defines the problem. Engineering tests the answer.

This does not mean philosophy gives all the answers. Philosophy cannot replace engineering, experiments, data, models, products, or institutional design. Its role is more specific: before we optimize, it asks what we are optimizing; before we deploy, it asks what we are allowing; before we measure, it asks what we are treating as success.

If an AI system makes conversation feel more natural, we need to ask whether naturalness means trustworthiness. If an agent makes tasks more automatic, we need to ask whether automation makes responsibility clearer. If a model makes expression more fluent, we need to ask whether fluency means understanding is more real. If a system keeps a confirmation button for the human, we need to ask whether confirmation means judgment remained in the formation of the work.

These questions cannot be answered only by stronger models. They require conceptual clarification.

Here, philosophy first appears as a capacity for clarification. It helps us understand what a term actually refers to, where its boundary lies, how it differs from neighboring concepts, and how it can be misused, diluted, or replaced inside real systems.

For example, explanation is not audit. Confirmation is not judgment. Personalization is not understanding. Efficiency is not capacity formation. A feeling of safety is not safety evidence. These distinctions may look small, but if they are blurred, system design, user understanding, and institutional governance all begin to drift.

Philosophy also helps with reality diagnosis. It does not remain only at the level of concepts; it has to look at how those concepts fail inside real systems. A concept may be clear in a paper, but inside a product it may be reshaped by buttons, default settings, interface language, workflow pressure, and organizational responsibility. If philosophy cannot enter these realities, it remains abstract.

Further, philosophy must help form standards of judgment. We cannot only say “this is complicated,” nor can we always remain at the level of “more research is needed.” In certain places, we need standards that can actually be used: what kind of explanation is strong enough to support trust? What kind of permission reflects real understanding of boundaries? What kind of AI assistance strengthens human capacity, and what kind merely helps humans avoid the formation of capacity?

Finally, philosophy must be translated into action. Otherwise, it remains only language. Executable philosophy is not philosophy turned into slogans. It means letting conceptual clarification, reality diagnosis, and standards of judgment enter design, governance, education, writing, research, and everyday AI use.

This is what I mean by executable philosophy.

It is not “philosophers instructing engineers.” It is closer to building usable conceptual tools across engineering, product, governance, and human judgment. It helps us ask, when facing an AI system, not only “Can it be done?” but “What is this, exactly?” Not only “Does it work well?” but “What does this performance sacrifice?” Not only “Was a human involved?” but “In what way was the human involved?”

These questions are not entirely new. Philosophy has long dealt with agency, action, responsibility, reality, knowledge, meaning, and common life. But AI makes them more urgent because it pushes them out of books and classrooms into system design, workflows, and public institutions.

In the past, “who is acting?” may have been a question in the philosophy of action. Now it is also an agent-system question.

In the past, “what is reality?” may have been an epistemological question. Now it is also a question of generative media, evidence chains, and information environments.

In the past, “how should responsibility be assigned?” may have been a question for ethics and legal philosophy. Now it is also a question for AI deployment, organizational governance, and automated decision-making.

In the past, “how do humans form judgment?” may have belonged to education, psychology, and philosophy. Now it is also a daily question for anyone using AI to write, research, learn, or decide.

That is the change.

AI has not made philosophy fashionable. It has made many questions that could once be postponed impossible to postpone.

If we do not define memory, we cannot build trustworthy long-term AI systems. If we do not define action, we cannot understand what agents are doing. If we do not define responsibility, we cannot know who carries the consequences. If we do not define reality, we cannot preserve the world inside generated content. If we do not define judgment, we cannot know whether humans remain in the loop as agents rather than interfaces. If we do not define meaning, we cannot understand why humans still matter when AI approaches or exceeds many human capacities.

These questions eventually return to a larger point: AI is not only changing tools. It is changing some of the conditions under which human life is formed.

How memory is formed. How action happens. How responsibility is carried. How reality is verified. How judgment is preserved. How meaning is created.

This is why I do not think philosophy-first is decoration. It is not a way to make research appear deeper, and it is not an elegant wrapper added around technology. It is necessary because if we do not define the problem first, more powerful systems may simply amplify undefined problems.

Of course, defining the problem does not solve the problem. After philosophy defines the problem, engineering must test answers, products must endure real use, institutions must establish boundaries, research must provide evidence, and humans must practice judgment in concrete life. Philosophy is not the destination. It only helps us avoid moving efficiently in the wrong problem.

This is why I say AI problems are becoming philosophical problems.

Not because AI is no longer technical, but because AI technology has entered the conditions that shape how humans become human. Technology remains important, engineering remains important, governance remains important, but none of them can bypass these premises.

This is especially important when we talk about AI risk.

Many people begin with the intuitive question: will AI harm humanity? That question matters. But before it, there is a more basic question: when we say humanity is at risk, what exactly do we mean by humanity?

If humanity is only a biological species, the boundary of risk looks one way. If humanity also includes beings capable of judgment, responsibility, institutional life, truth-seeking, and meaning-making, then the boundary of risk becomes deeper.

This is not a word game. It shapes how we define safety, how we understand harm, how we design governance, and whether we are able to see slow risks that do not appear as disasters.

Later, in the Paper Guide, I will take you int

[truncated for AI cost control]