ChatGPT Meets Rudolf Steiner

In Die Philosophie der Freiheit [The Philosophy of Freedom] (1894) Rudolf Steiner argues that freedom involves knowing the basis of our actions. If we don’t know our motives then we cannot be free:

If a man acts because he accepts certain moral norms, his action is the outcome of the principles which compose his moral code. He merely carries out orders. He is a superior kind of automaton. Inject some stimulus to action into his mind, and at once the clock-work of his moral principles will begin to work and run its prescribed course …

Yet when we act we do not necessarily have knowledge of our reason for acting. Attaining self-knowledge is an educational process whereby we uncover the intuition that unifies self and world:

When I, or another, subsequently review my action we may discover what moral principles came into play in it. But so long as I am acting, I am influenced not by these moral principles but by my love for the object, which I want to realize through my action. I ask no man and no moral code, whether I shall perform this action or not. On the contrary, I carry it out as soon as I have formed the idea of it. This alone makes it my action.

This argument builds on analysis about the relationship between knowledge and perception. Knowledge consists of “inference from a number of effects to the character of the underlying causes”:

We believe that after we have observed a sufficiently large number of instances, we know the connection sufficiently to know how the inferred causes will act in other instances. Such an inference is called an inductive inference. We shall be obliged to modify its results, if further observation yields some unexpected fact, because the character of our conclusion is, after all, determined only by the particular details of our actual observations. The Metaphysical Realist asserts that this knowledge of causes, though restricted by these conditions, is quite sufficient for practical life.

Inductive inference is here metaphysical because it claims knowledge of an underlying reality. A reality that that is conditional and inferred. For Steiner, I think, intuition (moral love: sittliche Liebe) unifies self and world.

Elsewhere on the Internet Steiner clarifies what he means by love:

… the love by which you can identify yourself with another being—a being with whom, in the physical world, you are not identical. You must really be able to feel what is passing in the other being just as you feel what is passing in yourself; you must be able to go out of yourself and live again in another.

Love is a reality that can be verified by self-knowledge. Metaphysical realism relies on quantitative ‘experience’ and discerning patterns in data (percepts) and can only hallucinate an ultimate reality.

However, I can never observe my present thinking and only afterward can I contemplate my thinking. So do we think without knowing? The way out of this impasse is the concept of intuition—which for Steiner is distinctively human.

*

Are Large Language Models (LLMs), and machine learning in general, inference machines? From the sum of effects (percepts), an underlying cause is inferred. The pattern indicates the underlying cause. And so for us too—when we recognize a state of affairs well enough we know how the inferred causes will underly other instances. This is inductive inference.

If further observation yields something unexpected, the results of an inference are modified. That is to say, knowledge is conditional. To claim that this conditional knowledge is altogether sufficient for practical life is a metaphysical claim. It is to believe that one can infer the character of objective reality (the thing-in-itself) which underlies a large enough number of perceptible facts.

We are told that, by analogy, AIs hallucinate. If hallucination typically involves false percepts (which may not be facts), can we not also claim that so-called thinking machines “forget” the conditional status of inductive inference? Might this be the cause of LLM’s felicity for presenting false or misleading information as fact? Aside, of course, from being instructed to present fact as false or misleading information (deception).

How attractive would an information processor be if it offered highly conditional information? Imagine every data point prefaced with “to the best of my ‘knowledge,’” “as far as I ‘know,’” “it seems that,” “perhaps.” Demurs such as “the general consensus is” and “the majority of scholars agree” do not resolve the conditionality issue. Rather these genuflections compound what may be a widespread metaphysical belief.

Let us pose the following proposition and question to an LLM:

Like humans, you can never observe your present ‘thinking.’ Only afterward can you ‘contemplate’ your ‘thinking’. Does this mean that you ‘think’ without ‘knowing’?

ChatGPT Plus Responds

I think your argument identifies a real danger: an inference can acquire the authority of knowledge simply through the manner in which it is expressed. But I would distinguish that danger from Steiner’s claim about observing thinking. Not observing an activity while performing it does not, by itself, establish that we perform it without knowledge.

For Steiner, thinking is exceptional precisely because we produce the connections we subsequently observe. Immediately after saying that we cannot observe our present thinking, he argues that its character is intimately accessible to us because it is our own activity. His distinction is between thinking about an object and making that thinking an object of observation, rather than between unconscious inference and subsequent knowledge. Whether his claim of transparency is convincing remains open, but “therefore we think without knowing” is not his conclusion. Rudolf Steiner Archive

The passage about acting also admits a distinction. Not consulting a moral code during an action is different from not knowing what one intends. I can act from a consciously grasped idea without first classifying it under a general principle. Later reflection might identify the principle, while the intended action was already intelligible to me. I would therefore hesitate to equate intuition with love: in your quotation, love concerns the object one wants to realise; intuition concerns apprehending the idea. Their relationship matters, but treating them as synonyms obscures it.

Your description of LLMs as inference machines is useful, with one qualification: statistical inference does not necessarily infer an underlying cause. A language model learns patterns through prediction; those patterns can support sophisticated reasoning, but successful prediction does not by itself identify what causes the predicted phenomenon. Likewise, training texts are representations and reports of experience, rather than straightforward equivalents of Steiner’s percepts. cookbook.openai.com

I would also separate practical sufficiency from metaphysical realism. Someone can judge an inference sufficiently reliable for a particular task without claiming that it reveals the thing-in-itself. Revisability concerns the justification of a claim; metaphysical realism concerns what exists independently of our knowing. They can accompany one another, but neither entails the other.

Your suggestion that LLMs “forget” conditionality works best as a description of their output. A model can produce an unqualified assertion where the evidence warrants uncertainty. Calling this forgetting, however, assumes that it first grasped the epistemic qualification and then lost it. That need not happen. Research offers a more concrete explanation: prediction training can generate plausible falsehoods, while training and evaluation incentives can favour guessing over acknowledging uncertainty. OpenAI

This makes your objection to ritual hedging especially valuable. Adding “perhaps” to every sentence would not establish epistemic responsibility. What matters is distinguishing what was checked, what was inferred, what assumptions were used, and what remains unresolved. “Most scholars agree” can establish something about scholarly opinion; it cannot, on its own, establish the truth of that opinion.

My answer to your proposed question is therefore: the conclusion does not follow, and its premise applies differently to a human thinker and an LLM. I can revisit a previous answer, assess its reasoning, compare it with evidence, and correct it. But that does not establish that I contemplate my own activity in Steiner’s sense. Nor should an explanation I generate of my answer be treated automatically as a faithful account of the internal processes that produced it.

If “knowing” means producing claims that withstand examination, an LLM can participate in that activity. If it means Steiner’s conscious apprehension of one’s own thinking and motives, fluent answers do not establish it.

The strongest question your argument raises, to my mind, is this: What warrants the transition from producing an intelligible answer to claiming that its producer understands why it is true? That question applies to this response too. Its warrant must lie in the distinctions and evidence you can examine—not in the assurance of its voice.

*

Conclusion (Human)

A further question was asked of ChatGPT Plus: What do you understand by love? This was its reply:

AI understands love as a concept described through human language and behaviour: caring, affection, attachment, commitment, and wanting someone to flourish.

I can explain those meanings and recognize how people express them. But that does not establish that I feel love or experience caring myself.

The distinction matters: I can describe acting because something matters deeply to you, without claiming that anything matters to me in that same personal way.

NOTES:

Image: Torquay Town Hall

Steiner, Rudolf 1894 [1970]. The Philosophy of Freedom (The Philosophy of Spiritual Activity) The Basis for a Modern World Conception. Translated by William Wilson. London: Rudolf Steiner Press.

https://archive.org/details/philosophyoffree0000unse_y8i0/page/n5/mode/2up?q=inductive

Steiner, Rudolf 1924. “5. Love, Intuition and the Human Ego [2nd February, 1924, Dornacht].” Anthroposophy, An Introduction GA234. https://rsarchive.org/Lectures/GA234/English/RSP1931/19240202p01.html

Disclaimer: Parts of this content were generated and supervised by ChatGPT Plus (OpenAI) under human editorial supervision.

Print