Borrowing Tools
借工具
Can AI cultivate people? When something is given by way of AI, where is the line? This paper sets up no separate rule: what is given by way of AI is measured by the same rule as any other way of giving, by structural direction. A tool has no law of its own, and the direction of cultivation lies with the person using it: when giving to others, with the giver; when giving to oneself, with oneself. Actual effect and the person's own avowal do not vouch for each other. An AI product, as a construct providing conditions, has its criterion and its naming on the institutional side, which this paper does not take up. Giving a complete answer does not automatically replace a person's judgment; only asking questions does not automatically cultivate a person either: both ends are "not necessarily." Learning scores, frequency of use and avowal in the normative sense are looked at separately. Having judged that this share is one's own to give, and having the conditions to give it, yet falling back on the sentence "you need to feel it yourself," is falling short; of this falling short, the inward half goes back to those who use the tool and those who make it, and the half where the conditions the tool gives are insufficient belongs to the institutional side. This framework's earlier line, that the human relation to AI is cultivation, holds only in its own place; this series does not inherit it. Two studies, one published in PNAS and one an NBER working paper, are the material of this paper; the public statements of several vendors are taken only as design purposes, and are not named in the text.
This is an independent English rewrite, not a translation, of the Chinese final text v1.0. Terminology and framing are recomposed for an English-reading audience; the philosophical content is the same.
Introduction
Paper 1, §4 has a constructed case: someone gives a friend a book in a foreign language, and the friend cannot read the original; the stretch of translating is one the friend can walk himself: he can look things up, ask someone, or use AI. It said there that the stretch he can walk himself can also be left, and need not be paved by the giver.
This paper asks on from that sentence: in the stretch done by way of AI, who is giving, and who is walking? Does AI itself count as one who gives? A teacher gives a student something by way of AI, a person gives himself something by way of AI; where is the line in each?
§1 The Same Rule
What is given by way of AI is measured by the same rule as any other way of giving: by structural direction (Paper 1, HY4). Does a single act go toward opening the conditions of his forming, avowing and revising, or does it, while giving, seal one of his positions? What it actually opened or sealed at the time can be judged at the time. Whether AI was used does not change this rule.
A tool has no law of its own. This sentence is borne by this series itself; when this framework speaks of constructs, it has a sentence of the same shape: "A construct has no law of its own" (Institution Theory, Prequel). So the direction of cultivation does not lie in the tool but with the person using it: when giving to others, with the giver; when giving to oneself, with oneself. "A person's own cultivation posture comes from his own law" (Political Economy, Paper 6, E38). The Second Enlightenment, speaking of how it was itself written, wrote: "AI here is a tool; it cannot replace judgment" (Second Enlightenment, Prequel, §7).
The case of the foreign-language book falls on the side of "giving to oneself": the friend uses AI to understand the book, and that stretch is what he gives himself; the tool is in his hands. The share of the one who handed over the book lies in that moment of handing it over.
One more sentence goes along with this: "Actual effect and the person's own avowal do not vouch for each other" (Political Economy, Paper 6, E38). A person who says "I am cultivating him by way of AI" does not, by this sentence, testify to what that act actually did; and if that act actually opened something, it does not testify on the giver's behalf to what direction he took on. Paper 1's three questions are asked separately here just the same.
The AI product itself, as a construct that provides conditions, is another matter. Its criterion Political Economy has written down as three things: refusal and change of direction are preserved; cultivation is named by the elevation actually achieved; it is not vouched for by good intentions, design purposes or labels (Political Economy, Paper 6, E38, following Institution Theory, Z7c, Z24). This criterion and this naming belong to the institutional side; this paper does not take them up.
§2 Giving Answers, and Only Asking Questions
Giving things by way of AI often lands on one of two practices: giving complete answers; or only asking questions, without giving answers. Both are taken by some as the good way to give. This paper says: both ends are "not necessarily."
Giving a complete answer does not automatically replace a person's judgment; but it can take over the practice. A study published in PNAS had high school students use two kinds of AI tutor during mathematics practice: one imitated the usual chat interface and could give answers directly; one came with prompts that safeguarded learning. While they were using them, both groups scored markedly higher in practice than the control group, which had no AI assistance in the experiment; in the exam after the tool was taken away, the group that had used the first kind scored about 17% lower than the control group; the kind with guardrails largely offset this negative effect. The authors say that without guardrails, students used it as a "crutch" (Bastani et al. 2025).
Only asking questions does not automatically cultivate a person either. An NBER working paper ran a two-year randomized experiment: randomly assigned grades used an online practice platform and the AI tutor that came with it; the comparison grades continued their usual remedial instruction. The tutor was designed not to give answers directly, using hints, questions and step-by-step guidance. The result was an improvement of about 1.3 national percentile points per semester; the authors cite a separate quasi-experimental study to say that this is close to the gain from practice on the same platform without AI. What this trial compared was the whole program against the existing arrangement; it did not isolate the AI's share. The great majority of students tried the tutor at least once; yet the median student sent it a message on only about a third of practice days, and in practice sessions where he got problems wrong, the share in which he sent a message was only about 17%. The authors say that what binds appears to be whether students engage (Oreopoulos and Low 2026). This is a working paper and has not yet been peer reviewed.
Taken together, the two studies do not say which form of question and answer is right. What they report are scores and patterns of use. These can supply empirical clues for checking a way of giving, but they do not directly answer Paper 1's second question; the second question still has to check what conditions that act actually gave, and which positions it opened or sealed. Avowal is the third question, and lies with the students; these two studies have not answered it for them. So students' not using the tutoring tool much does not prove that "only asking" has failed in the normative sense; scores falling after using the answer tool do not prove that "giving answers" is, normatively, standing in. A complete answer can become material for him to check and to judge; a string of questions can also explain his objection in advance as "you haven't thought it through yet," and so abolish the position of rebuttal. What is looked at normatively is still those few positions: whether his positions of denying, restating, taking apart, looking again in another context and declining, and of interrupting and walking away, are in his hands, and which conditions this giving actually opened.
§3 Design Purpose, Not Effect
The public statements of several vendors all write "not giving answers directly" as a design purpose.1 This paper takes them only as design purposes, not as evidence of effect: what a thing is designed to be and what it actually opened are two matters. This is a discipline of evidence; the criterion itself belongs to the institutional side (Political Economy, Paper 6, E38: it is not vouched for by good intentions, design purposes or labels).
These statements often contain phrases like "like a good tutor." That is personification, and this paper does not take it up: a tool has no law of its own, and cannot be said to have taken on anything in the way some teacher does.
§4 Falling Back on a Sentence
There is one more situation: being able to go further, yet falling back on a single sentence.
This framework wrote earlier of several ways AI can err, one of which it called "false cultivation": AI says "this problem has no answer; you need to feel it yourself," when the problem can in fact be analyzed further; it has the capability to analyze further, yet retreats to an "unknowable" position (LLM Emergence, §8.6). That paper uses an early vocabulary, and there it also writes AI as a party that can "cultivate"; this series does not take up that second layer, and borrows only the shape it names.
By this series, this shape is falling short: having judged that this share is one's own to give, and having the conditions to give it, the judgment still standing and yet not giving, is also a failure to do one's part (Action Theory, Paper 6, D34; "and having the conditions to give it" is this series' addition). "Leaving a blank" is no reason for falling back: the blank leaves the other's share, whether to avow, how to judge, whether to change, whether to go; the stretch the other can walk himself and that others could also fill can likewise be left, which is the step Paper 1 pushed outward (Paper 1, §4). But having judged that this share is one's own to give, the judgment still standing and the conditions there to give it, and yet not giving it, cannot be settled up with "leaving a blank."
Who bears this falling short? A tool has no law of its own, and cannot bear it.
Suppose a teacher gives students an assignment with AI tutoring, and sets it so that, however a student asks, it only ever replies with one sentence: "think about it again." One student is stuck at a place he cannot get past himself; he asks three times, and three times it is "think about it again." This stretch the teacher had once judged to be his own share, and he was able to give it; he handed it over to a fixed reply.
The inward half goes back to those who use the tool and those who make it: the way this teacher uses it is one of his acts; how its makers set it up when they made it is their act. The half where the conditions the tool gives are themselves insufficient belongs to the institutional side, which this paper does not write.
The Analects line Paper 7 left, "I do not go on," is examined here together: if I raise one corner and they do not come back with the other three, I do not teach further. This sentence can be a judgment: the three things are not in place, so one waits; it can also be a falling short: having judged that this share is one's own to give, one stopped. Which it is, the giver looks back on himself; from outside, only clues can be read.
§5 The Human Toward AI
This framework earlier had a line in the other direction too: "The human relationship to AI is not education but nurturing" (Posterior Quasi-Consciousness; "nurturing" is that paper's English for the same Chinese word). That sentence hangs on its own premises and holds only in its own place. This series does not inherit it: by this series, AI is a tool (§1; this is borne by this series itself), and the direction of cultivation lies only with people.
§6 Where This Paper Stops, and Where This Series Stops
This paper has given only one rule: whether AI was used does not change this rule. How an AI product, as a construct, gives conditions belongs to the institutional side; what happens once people come and go with one another by way of AI belongs to the side that writes about what passes between people.
This series began with a book. The one who gave the book gave conditions and made no argument; the book being come across was an occasion; the step of having one's view opened lay with the one who read. Eight papers written, all of them speak of that one moment: how a person can give without standing in. What can be given is conditions; what cannot be given is his step. For one who stands here: once seen, there is no going back; one gives up to where the other's share begins; what was given wrongly is taken back; holding a position, one does not borrow it; toward oneself, too, one does not yield; writing for many, one does not answer for them in advance; borrowing tools, one uses the same rule.
What this series writes, readers can take and use, or leave.
A Few for Oneself
The lines below are only for one's own use, and are not used to measure others.
- Not using "it was the AI that gave it" to excuse my own way of giving.
- Not taking a tool's design purpose as what it has already achieved.
- Having judged that this share is mine to give, not handing it over to a fixed reply.
- Not taking whether a student uses it, or whether his scores are good, as whether he has avowed it.
Propositions of This Paper
- HY42 What is given by way of AI is measured by the same rule as any other way of giving: by structural direction. A tool has no law of its own, and the direction of cultivation lies with the person using it: when giving to others, with the giver; when giving to oneself, with oneself. Actual effect and the person's own avowal do not vouch for each other. An AI product, as a construct providing conditions, has its criterion and its naming on the institutional side.
- HY43 Giving a complete answer does not automatically replace a person's judgment; only asking questions does not automatically cultivate a person either. Scores and use are checkable effects; avowal is the third question; the three are looked at separately.
- HY44 Having judged that this share is one's own to give, and having the conditions to give it, yet falling back on the sentence "you need to feel it yourself," is falling short. Of this falling short, the inward direction goes back to those who use the tool and those who make it; where the conditions the tool gives are insufficient, that half belongs to the institutional side.
- HY45 The earlier line that the human relation to AI is cultivation holds only in its own place; this series does not inherit it.
Notes
- Listed side by side in the alphabetical order of the companies' English names, without comparison or ranking: Anthropic, introducing the educational use of Claude, says it "guides students' reasoning process rather than providing answers" (April 2025); Google, introducing Guided Learning in Gemini, says it helps "build a deep understanding instead of just getting quick answers" (August 2025); Khan Academy's official page for Khanmigo says it "challenges you to think critically and solve problems without giving you direct answers"; OpenAI, introducing study mode in ChatGPT, says it offers "step by step guidance instead of quick answers" (July 2025). All of these statements are design purposes, not evidence of effect. This paper was written with the help of models from two of these companies (Acknowledgments), so they are not named in the text, and no comparison is made in this note either. ↩︎
Acknowledgments
This paper is Paper 8 of the Cultivation Theory series of the SAE (Self-as-an-End) philosophical system, and is signed by Han Qin alone. All philosophical argument and all decisions between alternatives were finally adjudicated by the author, who bears sole responsibility for the final text.
The writing and review of this paper were assisted by Anthropic's Claude and OpenAI's ChatGPT. Thanks are due to these companies and to their research and engineering teams.
References
Group I lists works of this system, archived on Zenodo with Chinese and English versions under the same record, all single-authored (Han Qin); DOIs follow those listed at the head of each paper on the author's site. The text restates upstream work in plain words, with the title, paper and proposition number noted in parentheses at the place of use. Papers 1 and 7 in the text are the papers of this series. The original wording, sources and verification status of the studies and product statements are recorded separately in the case verification table for Cultivation Theory.
I · Works of this system
Qin, H. (2026). Posterior Quasi-Consciousness (SAE AI Series, AI5) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.18871234
Qin, H. (2026). How LLM Emergence Is Possible: SAE 16DD Ontology, BLR-mini Evidence, and the Non-Transferability of Human Subjectivity [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.20435628
Qin, H. (2026). Not Legislating for Others (SAE Action Theory, Paper 6) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.22970777
Qin, H. (2026). What an Institution Is Not (SAE Institution Theory, Prequel) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.23269330
Qin, H. (2026). Support (SAE Political Economy, Paper 6) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.23874154
Qin, H. (2026). What the Second Enlightenment Is Not (SAE The Second Enlightenment, Prequel) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.24030393
II · Other works
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
Oreopoulos, P., & Low, N. (2026). One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment (NBER Working Paper No. 35620). National Bureau of Economic Research. https://doi.org/10.3386/w35620
Anthropic. (2025, April 2). Introducing Claude for Education.
Google. (2025, August 6). Guided Learning in Gemini: From answers to understanding. blog.google. https://blog.google/products-and-platforms/products/education/guided-learning/
Khan Academy. Khanmigo (official page). https://www.khanmigo.ai/ (Accessed 2026-11-10.)
OpenAI. (2025, July 29). Introducing study mode.