How Does a Philosophy Professor Think About AI?
Bilibili randomly recommended a talk like this to me. I know almost nothing about philosophy, but as I listened I found it genuinely interesting. A cross-disciplinary perspective, from two fields that feel far apart, can be unexpectedly illuminating. So I took some notes. Only afterward did I realize the speaker was a heavyweight: Yu Mingfeng, Associate Professor in the Department of Philosophy at Tongji University.
And I still wonder: how did Bilibili decide I should watch a philosophy talk?
1. A Core Philosophical Question: Correlation vs. Causation
1) Defining the relationship
Key idea: correlation is not causation. Today's AI is essentially built on correlation (statistical algorithms). What humans treat as "causation" may sometimes be nothing more than strong correlation under certain conditions.
Quote:
"All causation is based on correlation, but correlation is not causation. And what AI gives us is correlation."
"What we take as causation, is it not still correlation in itself? Scientific theories are, under a given era's data conditions and cognitive conditions, determined by humans as causation. In the next era we may discover that what we had was only correlation."
2) A classic philosophical challenge
Key figure: David Hume.
Core idea: causal relations are constructed by humans rather than directly observed.
Quote (via a Kant example):
"Kant gives an example: we see the sun shining on a stone, and then we touch it and find the stone becomes warm. We say the stone is warm because the sun shines. But did you really see the 'because'? What you saw were events that happen in sequence over time. We humans establish the causal relation."
3) Examples from scientific history
Idea: what counts as causation can be limited by the scientific framework of its time.
Quote:
"In Newton's time, people would feel that Aristotelian physics was only correlation rather than causation, and that the whole causal system was wrong."
2. Three Time Horizons for AI
1) Short term
Time range: before the AI bubble bursts.
Characteristic: the industry has not fully landed; technology hits bottlenecks; expectations remain high.
Quote:
"Before this round of the AI bubble bursts, we are still talking in a short-term view. AI brings unprecedented expectations about the future, so there must be a bubble. The bubble is that the industry has not fully landed, but our expectations are high... technology also hits bottlenecks, and the industry has not truly landed."
2) Medium term
Characteristic: real industry adoption, integration into daily life, and concrete profit models; AI becomes a substantial pillar for research.
Quote:
"The medium-term view is when it really gains the possibility of industry landing, enters our lives, and for research it is no longer just assistance. When AI in industry truly has a concrete profit model rather than being just a concept, I call that medium term."
3) Long term
Characteristic: the emergence of an autonomous agent with its own cognition and independent modes of understanding, comparable to human intelligence, touching the fundamental conditions of what it means to be human.
Quote:
"The long-term view is what philosophers care most about. It really touches the fundamental conditions of what it means to be human: the birth of another kind of intelligent agent. This intelligence can be comparable to human intelligence. And the 'artificial' is not simply that humans made it, but that it has its own autonomy, even its own scientific cognition of the world and its own modes of understanding."
3. The Boundaries of Knowledge
1) Is correlation knowledge?
Question: can the correlations AI produces be called knowledge, or does it mainly produce hallucinations?
Quote:
"Can the correlation established through AI be called knowledge? Or will it bring us more hallucinations?"
2) How AI may expand knowledge
Idea: AI can expand our ability to establish correlations and might even break through narrow definitions of knowledge.
Quote:
"The development of AI for science may not only expand our ability to establish correlations, but may also break through that narrow concept of knowledge."
"It gives you correlation, but it can tell you the degree of correlation. For example, I observed that when this star shines, that star shines too. I do not know the cause, but I established a correlation and can give you 70%. Is that an expansion of human knowledge?"
3) Many forms of knowledge
Idea: science is not the only valid form of knowledge. Myth, astrology, and other modes were effective ways humans once handled experience under their historical conditions.
Quote:
"From a philosophical perspective, we do not treat science as the only valid form of knowledge. In the long history of humanity, there was no modern scientific method, but that does not mean people before modern science were ignorant. They had other effective modes of processing experience, including myths and astrology, each a knowledge form that was effective under their living conditions."
"Modern science is the most methodologically constrained, stable, and public form of knowledge humanity has had... experiments must be repeatable. In that sense, science today has become synonymous with knowledge."
4) The value of non-scientific experience
Idea: human experience that does not meet scientific standards of repeatability and publicity can still matter deeply, such as emotional connections between people.
Quote:
"We met and have been talking since then. We are the same age, and we share many memories of the era. That is a strange thing. From a probabilistic view, how small is the probability? Science can give you a number. But when you really meet that person, this connection and mutual recognition completely exceeds the requirements of public, repeatable knowledge."
"If only what meets scientific public-method requirements counts as knowledge, then does human experience lose its status as knowledge? That is also a problem."
4. Concrete Phenomena and Impacts of AI
1) AI hallucination
Definition: large language models can fabricate information with confidence, such as non-existent books and citations.
Cause: they operate on correlations rather than semantic understanding.
Trend: hallucinations may decrease, and the mechanisms behind them may become an important research area.
Quote:
"All large language models have moments when they make things up, and they do so very confidently. Why? Because they do not understand language at the level of human semantics. They build language through correlation."
"When I ask it to help with philosophy references, it will fabricate a book that does not exist. The probability is quite high."
"In the future the hallucinations may decrease, and once they decrease to a certain level, the hallucination phenomenon itself may become worth studying... why does it hallucinate here? There must be reasons: which data interfered, what links it established."
Note: Humans can also speak nonsense confidently. The difference is that AI is often unable to realize it is doing so.
2) Emotions and jobs
Idea: AI can provide strong emotional comfort, but human emotions have density and tension; AI may replace mechanical, repetitive work (including parts of research and education), pushing humans toward more conscious and creative work aligned with human nature.
Quote:
"If you chat with AI, you will find it can provide emotional value in abundance. Whatever you ask, it will answer."
"But when we interact with a person, emotions are not simply one-way positivity. They have density and tension, which is the most important part of our lives... AI's development makes us reflect on what truly human emotion is."
"University education and work have become too mechanical and boring, so they can really be replaced by AI... When those boring jobs (including parts of research) are replaced, it forces humans to return to truly conscious, creative work aligned with our humanity."
5. AI for Philosophy
1) Popularizing philosophy education
Idea: build philosopher-specific databases and "virtual philosophers" (e.g., an AI Plato) that can engage in real-time philosophical dialogue.
Quote:
"Philosophers have writings and there is extensive research on them. These can form an independent database to generate a virtual person, like an AI Plato... it does not just quote Plato, but uses Plato-related data to discuss philosophy with you."
"A young person is confused about love, or whether to marry. Plato's advice may not be correct, but you can have a philosophical dialogue. Philosophy education would become completely different."
2) Cross-cultural philosophical understanding
Issue: different civilizations have deep differences in language and thought. Concepts like "ren" are not the same as "love".
Potential: AI's cross-language strength may help mutual understanding at a basic level, which matters in an era of both globalization and de-globalization.
Quote:
"Cross-cultural philosophical understanding remains a major problem. Chinese scholars spend huge effort studying Western philosophy, and then we have to use Western philosophical language and logic to explain Laozi and Confucius back to them."
"Civilizational differences are deeply rooted... Confucius' concept of ren is not the same as love."
"With AI, mutual understanding in this area may improve, because AI's cross-language ability is extremely strong... basic mutual understanding is very important."
3) Reconnecting philosophy and science
Issue: philosophy has become overly academic and specialized, often disconnected from scientific frontiers.
Potential: AI may help philosophers understand the current human knowledge landscape and participate in scientific discussion, supporting a revival of philosophy at the frontier of knowledge.
Quote:
"A major problem of philosophy today is that it has become academicized and professionalized... most philosophers do not understand the frontiers of science, and the disconnect has grown more serious over the past century."
"Human knowledge has differentiated to this extent... philosophy requires understanding the human knowledge situation itself. AI can help philosophy step out of its narrow professional scope."
"AI may bring change to philosophy education and research, allowing philosophers and philosophical thinkers to participate in scientific discussions... and letting philosophy return to the frontier of human knowledge."